







































ChengHsien Tsai et al. /Future Technology                                                                     February 2026| Volume 05 | Issue 01 | Pages 
263-277 

263 

 

 

 

Article 

Research on real-time data display and production 

management in a digitalized management factory 

with an artificial intelligence-assisted flexible 

manufacturing execution system 
ChengHsien Tsai, Oyyappan Duraipandi*, Dhakir Abbas Ali 

Faculty of Business and Accountancy, Lincoln University College, Malaysia. Wisma Lincoln, 12-18, Jalan SS 6/12, Ss 6, 

47301 Petaling Jaya, Selangor, Malaysia 

A R T I C L E   I N F O 
 

Article history: 
Received 19 August 2025  
Received in revised form 
25 October 2025 
Accepted 02 December 2025 
 
Keywords:  
Manufacturing Execution Systems (MES), 
Artificial Intelligence (AI), Digital twin,  
Flexible manufacturing,  
Cyber-Physical Systems (CPS), 
Microservices architecture, 
Distributed stream processing 
 
*Corresponding author 
Email address: 
oyyappan@lincoln.edu.my 
 
DOI: 10.55670/fpll.futech.5.1.23 

A B S T R A C T 
 

Traditional Manufacturing Execution Systems (MES) face critical limitations in 
addressing Industry 4.0 demands for real-time processing, flexible scheduling, 
and adaptive decision-making, with less than 1% of manufacturing data 
effectively utilized. This research develops an Artificial Intelligence (AI)-
assisted flexible MES framework integrating real-time data visualization, digital 
twin technology, and distributed intelligence to enable proactive manufacturing 
management. The system employs Design Science Research (DSR) methodology 
and implements a microservices architecture using Apache Kafka for message 
streaming, Flink for real-time processing, and TensorFlow for AI inference, 
deployed across five production lines with 2,350 sensors and 45 Programmable 
Logic Controllers (PLCs). Results demonstrate exceptional performance with 
system throughput reaching 12,500 messages per second, the design target by 
25%, average data collection latency below 10 milliseconds, and 99.9% 
availability over 72-hour continuous operation. Production efficiency improved 
significantly with 25% increased output, 65.7% reduction in defect rates (from 
35,000 to 12,000 Parts Per Million), and 87.5% decrease in changeover time 
(from 120 to 15 minutes). Overall Equipment Effectiveness (OEE) increased 
from 60% to 82%, approaching world-class benchmarks (>85%). This research 
validates distributed intelligence architectures for achieving simultaneous 
improvements in manufacturing flexibility and efficiency, challenging 
traditional theoretical trade-offs while providing a practical implementation 
roadmap for digital transformation in manufacturing enterprises. 

1. Introduction 

In the introduction, explain why you did it (motivation). 
The transition to Industry 4.0 has reshaped manufacturing, 
imposing stringent demands for operational agility, real-time 
decision making, and seamless system integration [1]. The 
Manufacturing Execution System (MES) serves as a pivotal 
intermediary between Enterprise Resource Planning (ERP) 
and shop-floor operations, coordinating increasingly complex 
production processes [2]. Yet conventional MES architectures 
remain constrained when confronting modern requirements, 
particularly real-time data handling, flexible scheduling, and 
adaptive responses to volatile market conditions [1]. Despite 
generating vast data streams, traditional MES exploits less 
than 1% for decision making [3,4], highlighting the need for 
AI-integrated systems. This study proposes an AI-assisted, 
flexible MES augmented with advanced real-time 

visualization. The framework employs a digital twin for 
cyber-physical synchronization, applies machine-learning 
models for predictive analytics and optimization, and 
implements multi-layer dashboards to enhance operational 
transparency. The architecture contributes both theoretical 
frameworks and practical implementation strategies for 
intelligent manufacturing. Section 2 reviews existing 
literature to identify specific gaps this study 
addresses.Research Questions: This research addresses four 
specific questions: 
RQ1: How can AI capabilities be integrated with flexible MES 
to achieve sub-100ms latency and 10,000+ messages/second 
throughput at enterprise scale? 
RQ2: What are the measurable impacts of AI-assisted flexible 
MES on manufacturing performance (productivity, quality, 
flexibility, equipment efficiency)? 

 

 Open Access Journal 

ISSN 2832-0379 

February 2026| Volume 05 | Issue 01 | Pages 263-277 

https://doi.org/10.55670/fpll.futech.5.1.23 

 

 

 

 

 

 

 

 

 

 

 

 

 

Journal homepage: https://fupubco.com/futech 

 

Future Technology 

mailto:oyyappan@lincoln.edu.my
https://doi.org/10.55670/fpll.futech.5.1.23
https://fupubco.com/futech


ChengHsien Tsai et al. /Future Technology                                                                     February 2026| Volume 05 | Issue 01 | Pages 263-277 

264 

 

RQ3: What technical challenges emerge during industrial 
deployment, and what solutions enable 99.9%+ reliability? 
RQ4: Can distributed intelligence architecture overcome the 
traditional flexibility-efficiency trade-off in manufacturing 
systems? 

2. Literature review 

Over recent decades, MES has moved from transaction-
oriented middleware to a platform for cyber-physical 
production integration. Early deployments mainly bridged 
Enterprise Resource Planning (ERP) and the shop floor, 
emphasizing scheduling, resource allocation, and data 
capture [5]. Yet conventional designs struggled with real-time 
data streams, flexible manufacturing, and intelligent decision 
support [6]. Systematic reviews note that although MES has 
been commercial since the 1990s, scholarly work has only 
recently engaged with intelligent architectures aligned with 
Industry 4.0 [7]. The shift from model-based to data-driven 
manufacturing has prompted a reconceptualization of MES, 
with emerging frameworks favoring distributed intelligence, 
service-oriented architectures, and autonomous decision 
making [8]. While digital twin implementations have shown 
measurable improvements in specific applications [9,10], 
plant-wide integration remains challenging due to 
heterogeneous data formats and complex synchronization. 

Artificial intelligence has progressed from an auxiliary 
tool to a distinct production factor, with recent empirical 
analyses linking AI to productivity gains alongside traditional 
inputs [11]. In flexible manufacturing, machine learning—and 
especially deep learning—methods demonstrate strong 
performance in predictive maintenance, quality prediction, 
and adaptive scheduling [12]. Explainable AI (XAI) has gained 
traction as organizations seek trust in high-stakes decisions; 
interpretable models such as Generalized Additive Models 
(GAMs) provide transparency for process optimization and 
energy management despite advances, challenges persist, 
including large training data requirements, real-time 
inference complexity on resource-constrained hardware, 
robustness across variable operating conditions, and 
interoperability issues with legacy systems [13]. 

Real-time data visualization has evolved from simple 
dashboard displays to sophisticated multi-dimensional 
analytics platforms capable of processing high-velocity 
manufacturing data streams. Modern visualization 
frameworks leverage advanced technologies, including 
augmented reality (AR), edge computing, and AI-powered 
pattern recognition, to transform complex multivariate data 
into actionable insights [14]. Studies indicate significant 
operational improvements from real-time visualization 
systems [15]. However, current approaches face challenges in 
handling data volume, variety, and velocity, with many 
systems struggling to maintain sub-second response times. 
The lack of standardized frameworks and integration 
difficulties hinder widespread adoption. A critical reading of 
prior work indicates persistent gaps that hinder truly 
intelligent and flexible manufacturing. Individual 
technologies show promise, yet integration remains 
fragmented; most studies treat isolated deployments rather 
than end-to-end architectures. The lack of a standardized 
framework that unifies AI, digital twins, and real-time 
visualization within a single MES platform appears to be a 
core barrier to autonomous, adaptive manufacturing [16]. 
Scalability is also underexplored: few reports demonstrate 
sustained sub-second latency at enterprise scale across 
thousands of connected devices. To address these gaps, this 
study proposes an integrated architecture that fuses AI-

assisted decision making, digital-twin synchronization, and 
multi-layer real-time visualization within a flexible MES. The 
results suggest that superior performance can be achieved 
while preserving system scalability and adaptability. Section 
3 presents the theoretical framework and system architecture 
addressing these gaps through novel integration mechanisms. 
Research Novelty and Contributions: This research differs 
from prior work in four ways: 
Holistic Integration: Six AI models (LSTM, SVM+RF, CNN, 
Isolation Forest, GA, PSO) unified in one architecture, 
achieving 42ms latency—previous systems sacrifice 
modularity for performance or vice versa. 
Real-time Digital Twin: Bi-directional cyber-physical 
synchronization with 42ms latency (vs. minutes-to-hours in 
existing systems) through edge preprocessing and 
incremental updates. 
Manufacturing-aware Visualization: 12 FPS per-user with 
<100ms latency, exceeding literature reports (1-5 FPS, 500-
1000ms). 
Transcending Trade-offs: Simultaneous flexibility (+87.5% 
changeover speed) and efficiency (+25% output) 
improvements, challenging traditional theory that assumes 
inverse relationships. 

3. Theoretical framework and system architecture 

3.1 Conceptual framework development 
This study grounds an AI-assisted flexible MES in Cyber-

Physical Systems (CPS) theory and socio-technical principles. 
CPS denotes tight coupling of computation and physical 
processes, where embedded computing and networks 
monitor and control plants via feedback. The proposed 
framework extends classical CPS by embedding distributed 
intelligence and autonomous decision-making across 
hierarchical levels. AI is positioned as a cognitive layer that 
bridges the semantic gap between raw sensor streams and 
actionable insights, enabling what recent work refers to as 
“cognitive manufacturing.” As outlined in Figure 1, the 
framework comprises four functional dimensions 
coordinated by a central AI–MES orchestration hub.  

 

ERP

Cloud
Platform

MOM

SCM

Quality Prediction

Analysis

In
tellig

en
t In

v
en

to
ry

M
a
n

a
g
em

en
t

Human-Machine

Collaborative Sched.

A
n

o
m

a
ly

 D
et

ec
ti

o
n

&
 A

le
rt

Real-time

 Data

Quality Data

Equipment

 Status

Inventory

 Info

Energy Data

Scheduling Cmd.

Supply Info

Anomaly

 Alerts AI-MES
Core

AI Visual

Data

Flexible

 
 

Figure 1. Conceptual Framework of AI-Assisted Flexible MES. Eight 
operational modules: (1) Quality Prediction, (2) Predictive 
Maintenance, (3) Production Scheduling, (4) Energy Optimization, (5) 
Inventory Management, (6) Equipment Monitoring, (7) Supply Chain 
Collaboration, (8) Human-Machine Collaborative Scheduling 



ChengHsien Tsai et al. /Future Technology                                                                     February 2026| Volume 05 | Issue 01 | Pages 263-277 

265 

 

The theoretical basis follows hierarchical 
decomposition: complex operations are partitioned into 
manageable modules while system coherence is preserved 
through standardized data flows and interfaces. Each module 
acts as an autonomous agent that performs local optimization 
and contributes to global objectives via collaborative 
protocols. This multi-agent design accords with advances in 
distributed manufacturing intelligence that decentralize 
authority beyond monolithic control. 

Eight operational modules instantiate the theory into 
practice: (1) Quality Prediction, (2) Predictive Maintenance, 
(3) Production Scheduling, (4) Energy Optimization, (5) 
Inventory Management, (6) Equipment Monitoring, (7) 
Supply Chain Collaboration, and (8) Human-Machine 
Collaborative Scheduling as shown in Figure 1. Quality 
prediction employs probabilistic models to anticipate defects, 
whereas maintenance prediction uses temporal pattern 
recognition to detect degradation. Both rely on the 
assumption that manufacturing processes are deterministic 
dynamics corrupted by stochastic noise, formalized as 

( ) ( ( ), ) ( )Y t f X t t = +
           (1) 

In this architecture, the variables represent specific system 
components: Y(t) denotes output vectors (quality, health, 
performance metrics); X(t) represents input streams from 
2,350 sensors and 45 PLCs (100ms-10s sampling); f(.) 
embodies AI mapping functions (LSTM, SVM+RF, CNN, 
GA/PSO); 𝜃 denotes learnable parameters updated through 
online learning; and 𝜀(𝑡) captures system uncertainties 
(sensor noise, model errors), enabling machine learning 
while quantifying uncertainty. This formulation enables 
machine-learning methods to learn f(.) from data while 
quantifying uncertainty within probabilistic frameworks." 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

3.2 System architecture design 
The system architecture translates theoretical concepts 

into a practical implementation blueprint through a five-layer 
hierarchical structure that ensures scalability, modularity, 
and real-time performance. As illustrated in Figure 2, the 
architecture adopts a service-oriented approach where each 
layer provides well-defined services to adjacent layers 
through standardized Application Programming Interfaces 
(APIs). The presentation layer supports multi-modal human-
machine interaction through web dashboards, mobile 
applications, and large-format displays, implementing 
responsive design principles to adapt visualization 
complexity to device capabilities and user contexts.  

The service layer represents the architectural innovation 
that enables flexible integration of AI capabilities with 
traditional manufacturing operations. By separating AI 
services from business services, the architecture supports 
independent scaling and evolution of intelligent capabilities 
without disrupting core manufacturing processes. The AI 
service group implements prediction, optimization, diagnosis, 
classification, control, and learning functions through 
containerized microservices that can be dynamically 
orchestrated based on computational demands. Each AI 
service encapsulates specific algorithms while exposing 
uniform interfaces for service consumption. Table 1 
summarizes the deployed AI algorithms, their manufacturing 
applications, and selection rationale. 

GA and PSO handle discrete decision variables and 
combinatorial solution spaces that gradient-based methods 
cannot address. The data flow architecture implements a 
lambda pattern combining batch and stream processing to 
balance latency and throughput requirements.  

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 

Figure 2. System architecture diagram: (a) Five-layer architecture: Presentation layer, Application layer, AI+ business service layer, Data layer 

(Redis/PostgreSQL/ Hadoop), Device layer (PLCs n=45, Sensors n=2350), (b) Data flow: Kafka topics (100ms/1s/5-10s sampling) → Flink 

pipelines (<50ms latency) → Multi-temperature storage (Hot: Redis <1ms, Warm: PostgreSQL, Cold: Hadoop), (c) AI modules on GPU cluster 

(NVIDIA Tesla V100 ×4) with TensorFlow/PyTorch frameworks 



ChengHsien Tsai et al. /Future Technology                                                                     February 2026| Volume 05 | Issue 01 | Pages 263-277 

266 

 

 

 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Real-time streams from high-frequency sensors (100ms 

sampling) flow through Apache Kafka (>10,000 
messages/second), while batch processes aggregate 
historical data. The multi-temperature storage strategy 
maintains hot data in Redis (<1ms access), warm data in 
PostgreSQL for structured queries, and cold data in Hadoop 
for archival analytics. The data pipeline implements a lambda 
architecture combining real-time stream processing (Apache 
Kafka/Flink) and batch analytics, with a multi-temperature 
storage strategy optimizing for different data access patterns 
and latency requirements. The pipeline infrastructure is 
deployed on Kubernetes clusters, providing horizontal 
scalability, automated failover, and 99.9% availability over 
continuous operation. Kubernetes was selected for its 
superior resource utilization and ecosystem maturity. 

3.3 Real-time data processing framework 
The real-time data layer forms the computational 

backbone that converts raw sensor streams into actionable 
intelligence under strict latency constraints. A multi-stage 
pipeline progressively improves data quality while 
preserving temporal coherence across distributed nodes. 
Acquisition begins at the edge: smart sensors and 
Programmable Logic Controllers (PLCs) emit continuous 
streams at 10–100 Hz. Protocol translation services 
normalize industrial protocols—Open Platform 
Communications Unified Architecture (OPC UA), Modbus, and 
Message Queuing Telemetry Transport (MQTT)—into 
standardized formats for downstream processing. Edge 
nodes conduct initial validation and filtering to reduce 
bandwidth and latency. Lightweight anomaly detection using 
the Isolation Forest algorithm (as detailed in Section 4.1) flags 
suspicious readings prior to uplink. The Isolation Forest was 
selected for edge deployment due to its computational 
efficiency (O(n log n) complexity, <5ms inference), 
unsupervised learning capability, and 94.2% detection 
accuracy in production trials. The preprocessing pipeline 
applies six sequential transforms, including null removal, 
outlier detection, missing-value imputation, normalization, 
feature extraction, and temporal aggregation at multiple time 
scales (5-second, 1-minute, and 5-minute windows), 
optimized for different monitoring requirements. This 
structured approach yields ≥98.5% completeness, 99.2% 
accuracy, and 99.8% timeliness. 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

The framework implements a novel three-channel 
processing architecture that prioritizes data streams based 
on criticality and latency requirements. The fast channel 
processes critical alarms within 10ms latency through direct 
memory access and priority queuing, bypassing standard 
processing pipelines for immediate response. The standard 
channel handles production data through the complete 
analysis chain with sub-100ms latency, applying both rule-
based and machine learning inference. The batch channel 
processes historical data for complex analytics and model 
training, leveraging distributed computing frameworks to 
handle petabyte-scale datasets. 

3.4 Visualization module design 
The visualization module design addresses the cognitive 

challenges of presenting complex, multi-dimensional 
manufacturing data to diverse stakeholder groups ranging 
from shop-floor operators to executive management. 
Drawing upon principles from visual analytics and human-
computer interaction, the module implements a hierarchical 
information architecture that progressively reveals detail 
based on user interaction patterns and decision-making 
contexts. The design philosophy emphasizes glanceability for 
real-time monitoring, explorability for root-cause analysis, 
and actionability for decision support, implementing what 
recent research terms "manufacturing-aware visualization 
grammar". 

The dashboard adopts a tile-based layout in which each 
tile is a self-contained visualization module with independent 
refresh cycles and interaction handlers. This modularity 
enables role-aware, priority-driven composition. Real-time 
binding uses WebSockets to push updates at 12 frames per 
second, exceeding the 10-FPS threshold for perceived real-
time response. To render thousands of concurrent streams 
while preserving clarity, the visualization pipeline applies 
intelligent data reduction—temporal aggregation, spatial 
clustering, and semantic filtering—thereby controlling 
computational complexity without sacrificing 
interpretability. Section 4 details the research methodology 
and implementation strategy to translate these theoretical 
designs into functioning industrial systems. 

 

 

 

Table 1. AI algorithm portfolio and selection rationale 

Algorithm Application Domain Key Performance Metrics Selection Rationale 
 LSTM Time-series prediction: 

equipment degradation 
forecasting, demand 

prediction 

156ms inference latency, 
94.2% accuracy for 72-
hour failure prediction 

Superior temporal dependency capture; handles 
variable-length sequences; effective for non-

stationary manufacturing processes 

SVM + RF Quality classification: defect 
categorization across six 

types 

420ms per frame 
processing time for multi-

class classification 

Robust with limited training samples; effective in 
high-dimensional feature spaces; ensemble 

mitigates individual weaknesses 

CNN Image-based defect detection 
from optical inspection 

systems 

Real-time processing of 
visual inspection data 

Automatic hierarchical feature extraction from raw 
images; translation-invariant pattern recognition; 

eliminates manual feature engineering 

GA Production scheduling: job 
sequencing, resource 

allocation 

Handles discrete decision 
variables and constraint 

satisfaction 

Global optimization avoiding local minima; handles 
combinatorial problems with discrete choices; 
accommodates multi-constraint environments 

intractable for gradient methods 

PSO Dynamic rescheduling, real-
time resource reallocation 

Fast convergence for 
online adaptation (<2 

seconds response time) 

Computational efficiency for real-time response; 
lower overhead than GA for continuous parameters; 

balances exploration-exploitation for dynamic 
environments 

 



ChengHsien Tsai et al. /Future Technology                                                                     February 2026| Volume 05 | Issue 01 | Pages 263-277 

267 

 

4. Research methods 

This study adopts a Design Science Research (DSR) 
methodology to build and assess the AI-assisted flexible MES 
framework [17]. DSR offers a systematic route to create 
artifacts that solve practical problems while extending theory 
[18]. The process comprises six activities—problem 
identification, objective definition, design and development, 
demonstration, evaluation, and communication—and 
proceeds iteratively, with each cycle incorporating feedback 
to refine architecture and implementation strategies [19]. To 
structure technical realization, the System Development Life 
Cycle (SDLC) complements DSR [20]. An agile–waterfall 
hybrid is employed: waterfall rigor governs critical 
infrastructure, while agile sprints drive AI module 
development and user-interface design. This hybrid enables 
rapid algorithm prototyping without compromising stability 
and reliability. Action-research principles foster close 
collaboration with manufacturing practitioners throughout 
development; recurring stakeholder workshops and feedback 
sessions ensure that the system addresses real-world 
challenges and operational constraints. 

4.1 System implementation strategy  
The system implementation follows a phased 

deployment strategy designed to minimize operational 
disruption while maximizing learning opportunities. The 
technology stack selection prioritizes open-source 
frameworks and industry-standard protocols to ensure 
interoperability and scalability, as shown in Table 2 [21]. The 
implementation architecture leverages containerization 
through Docker and Kubernetes to enable microservices 
deployment and horizontal scaling [22]. The development 
phases consist of four major stages: infrastructure setup, core 
MES functionality implementation, AI integration, and 
visualization layer development.  

Table 2. System development technology stack and tools 

Level/Category Technology 
Component 

Implementation 

Frontend Layer Web Framework React.js 

 Visualization Library D3.js + ECharts 

 Mobile React Native 

 Large Display Grafana 

Application Service 
Layer 

Backend Framework Spring Boot 

 API Gateway Kong 

 Message Queue Apache Kafka 

 Cache Redis 

AI Service Layer Deep Learning 
Framework 

TensorFlow + 
PyTorch 

 Model Service TensorFlow Serving 

 MLOps Platform MLflow 

 GPU Computing NVIDIA CUDA 

Data Processing 
Layer 

Stream Processing 
Engine 

Apache Flink 

 Batch Processing 
Framework 

Apache Spark 

 Time-series Database InfluxDB 

 Data Lake Apache Hadoop 

Device Access 
Layer 

Sensor Deployment Distributed Sensor 
Network 

 OPC UA Server KEPServerEX 

 MQTT Broker Eclipse Mosquitto 

 Edge Computing Azure IoT Edge 

Development Tools Containerization Docker + Kubernetes 

 CI/CD Jenkins + GitLab 

 Monitoring & 
Operations 

Prometheus + ELK 
Stack 

 

Each phase incorporates continuous integration and 
continuous deployment (CI/CD) pipelines to automate testing 
and deployment processes. The infrastructure setup phase 
establishes the foundational components, including message 
queuing systems, time-series databases, and edge computing 
nodes. Apache Kafka serves as the primary message broker, 
configured with three topic partitions to handle over 10,000 
messages per second [23]. Integration protocols follow 
Industry 4.0 standards, implementing OPC UA for equipment 
connectivity and MQTT for lightweight IoT device 
communication [24]. The system adopts a Service-Oriented 
Architecture (SOA) approach where each functional module 
exposes RESTful APIs for inter-service communication. 
GraphQL endpoints provide flexible data querying 
capabilities for front-end applications, while WebSocket 
connections enable real-time data streaming to visualization 
dashboards. The complete implementation workflow is 
illustrated in Figure 3. 

4.2 Data collection and processing methods  
The data collection framework implements a multi-

tiered architecture, as depicted in Figure 4, that captures 
heterogeneous manufacturing data from 2,350 sensor points 
distributed across 20 major monitoring locations. High-
frequency sensors operating at 100Hz sampling rates 
monitor critical parameters including temperature, pressure, 
vibration, and electrical current. The edge computing layer 
performs initial data validation and compression using the 
LZ4 algorithm, achieving a 70% compression ratio while 
maintaining sub-10ms processing latency. The preprocessing 
pipeline applies six sequential transformations to ensure data 
quality. Outlier detection utilizes statistical process control 
limits (±3σ) to identify anomalous readings [25]. Missing 
value interpolation employs cubic spline functions to 
maintain temporal continuity, achieving a 95.5% fill rate 
across all data streams while maintaining interpolation error 
below 2% of signal variance [26]. Feature extraction 
techniques combine time-domain analysis (mean, variance, 
peak values) with frequency-domain analysis via the Fast 
Fourier Transform (FFT), reducing dimensionality from 100 
to 20 features while preserving 98% of the variance. Real-
time analysis implements a three-channel processing 
architecture optimized for different latency requirements. 
The fast channel processes critical alarms within 10ms 
through direct memory access and priority queuing. The 
standard channel handles production data with sub-100ms 
latency through the complete analysis pipeline, applying both 
rule-based logic and machine learning inference. The batch 
channel leverages distributed computing frameworks for 
complex analytics on historical data, supporting petabyte-
scale processing [27]. 

4.3 Validation methodology  
The validation methodology employs a comprehensive 
performance evaluation framework detailed in Table 3 that 
assesses five key dimensions: real-time performance, system 
throughput, production efficiency, data quality, and system 
reliability [28]. Performance metrics are collected 
continuously through embedded monitoring agents and 
aggregated using Prometheus for real-time analysis [29]. 
Experimental validation follows a three-phase approach: 
laboratory testing, pilot deployment, and full-scale 
implementation. Laboratory testing utilizes synthetic data 
generators to simulate production scenarios and stress-test 
system components.  

 



ChengHsien Tsai et al. /Future Technology                                                                     February 2026| Volume 05 | Issue 01 | Pages 263-277 

268 

 

 
Figure 3. System development and implementation process flow 

 

 

 

 

 

 

 

 

 

Figure 4. Data collection and processing flow. Sampling rates: 100ms (high-frequency sensors for vibration/current), 1s (standard monitoring for 

temperature/pressure), 5s (auxiliary metrics). Edge processing: LZ4 compression (70% ratio), latency <10ms. Data sources: 20 monitoring 

locations, 2,350 sensor endpoints 



ChengHsien Tsai et al. /Future Technology                                                                     February 2026| Volume 05 | Issue 01 | Pages 263-277 

269 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

The pilot deployment phase implements the system on a 
single production line for 30 days, collecting baseline 
performance data and identifying optimization opportunities. 
Full-scale implementation incorporates lessons learned from 
pilot testing and extends deployment across five production 
lines with different product configurations. Statistical 
validation employs paired t-tests to compare pre- and post-
implementation performance metrics across five key 
dimensions: Overall Equipment Effectiveness (OEE), defect 
rate (PPM), changeover time, first-pass yield, and daily 
output. Significance levels were set at α = 0.05 [30]. Results 
demonstrate statistically significant improvements (p< 
0.001) across all metrics: OEE (60% to 82%, t = 8.42), defect 
rate (35,000 to 12,000 PPM, t = 6.73), changeover time (120 
to 15 minutes, t = 12.35), first-pass yield (96.5% to 98.8%, t = 
5.91), and daily output (1,200 to 1,500 units, t = 7.28). Effect 
sizes (Cohen's d: 1.8-3.2) indicate large practical significance, 
with statistical power >0.95 confirming robustness. System 
reliability assessment follows IEC 61508 standards for 
functional safety, targeting Safety Integrity Level (SIL) 2 for 
critical control functions [31]. Section 5 presents concrete 
implementation details and case study results from deploying 
the system in an operational manufacturing facility.  

5. System implementation and case study 

5.1 Implementation environment  
The implementation was conducted at a discrete 

manufacturing facility specializing in electronic enclosure 
production, operating five production lines with an annual 
capacity of 6 million units. The facility encompasses 75,000 
square feet of production space equipped with injection 
molding machines, Computer Numerical Control (CNC) 
machining centers, automated assembly lines, and quality 
inspection stations.  

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

The hardware infrastructure comprised 45 PLCs 
distributed across production equipment, 2,350 IoT sensors 
monitoring critical parameters including temperature, 
pressure, vibration, and electrical current at 20 major 
monitoring points. Edge computing nodes based on NVIDIA 
Jetson AGX Xavier platforms were deployed at each 
production line, selected for their superior AI inference 
performance (32 TOPS), power efficiency (30W), and 
industrial-grade reliability suitable for harsh manufacturing 
environments, enabling sub-10ms processing latency. 

The software environment integrated existing ERP (SAP 
S/4HANA) and Manufacturing Operations Management 
(MOM) systems through standardized APIs and message 
queuing protocols. The technology stack leveraged 
containerized microservices deployed on Kubernetes 
clusters, ensuring horizontal scalability and fault tolerance. 
Real-time data streaming was handled by Apache Kafka, 
configured with three topic partitions to handle message 
throughput exceeding 12,500 messages per second. The 
implementation followed a phased approach aligned with 
agile-waterfall hybrid methodology, enabling iterative 
development while maintaining system stability. 

5.2 AI-MES integration details  
The AI integration architecture implemented six 

specialized machine learning models deployed as 
containerized microservices within the service layer. 
Predictive maintenance algorithms utilized LSTM networks 
trained on 18 months of historical equipment data, achieving 
156ms inference latency for real-time anomaly detection. The 
LSTM model processed sequences of 100 time steps with 20 
features extracted through FFT, maintaining prediction 
accuracy of 94.2% for equipment failure events within a 72-
hour horizon. Quality prediction employed ensemble 
methods combining SVM classifiers and Random Forest 
algorithms, processing image data from optical inspection 

Table 3. System performance evaluation index system 

Evaluation 
Dimension 

Key Metrics Calculation Formula Target Value Industry Benchmark Weight 

Real-time 
Performance 

Data Collection 
Latency 

Time from sensor trigger to data storage <10ms (high-
speed)/<100ms 

(regular) 

50-100ms 15% 

 Stream Processing 
Latency 

Time from data queue to processing 
completion 

<50ms 100-500ms 15% 

 Visualization 
Refresh Rate 

Time from data update to interface 
display 

<100ms 100-1000ms 10% 

 Alarm Response 
Time 

Time from anomaly occurrence to alarm 
trigger 

<5s 10-30s 10% 

System Throughput Concurrent 
Connections 

Number of simultaneous device 
connections 

>5000 1000-3000 8% 

 Message 
Processing Rate 

Messages processed per second >10000 msg/s 1000-5000 msg/s 12% 

 Data Write Rate Data points written per second >100000 points/s 10000-50000 points/s 8% 

Production 
Efficiency 

Equipment OEE Availability × Performance × Quality 82% Industry average 60%, 
World-class 85% 

10% 

 Changeover Time Time required for product switching 15 minutes 90 minutes 
(traditional) 

8% 

 Capacity 
Utilization 

Actual output/Theoretical capacity >85% 70-80% 4% 

Data Quality Data 
Completeness 

% of required data collected >98.5% 95-98% 3% 

 Data Accuracy % of accurate data >99.2% 97-99% 3% 

 Data Timeliness % meeting time requirements >99.8% 95-98% 2% 

System Reliability System 
Availability 

MTBF/(MTBF+MTTR) >99.9% 99.5-99.9% 3% 

 Failure Recovery 
Time 

Time to restore normal operation <30 minutes 1-4 hours 2% 

 Backup Success 
Rate 

% successful backups 100% 99-100% 1% 

 



ChengHsien Tsai et al. /Future Technology                                                                     February 2026| Volume 05 | Issue 01 | Pages 263-277 

270 

 

systems at 420ms per frame for defect classification across six 
categories (scratches, dents, discoloration, dimensional 
defects, contamination, and surface finish anomalies). 
Integration with existing manufacturing systems required 
protocol adapters supporting OPC UA, Modbus TCP, and 
MQTT. The AI orchestrator implemented reinforcement 
learning-based resource allocation across four NVIDIA Tesla 
V100 GPUs. Model versioning utilized MLflow, maintaining 
three versions (production, staging, experimental) with 
automated A/B testing. Real-time processing capabilities 
were achieved through a three-tier caching strategy: Redis for 
hot data with sub-millisecond access latency, PostgreSQL for 
structured queries with indexed access patterns, and Apache 
Hadoop for historical data analysis. The stream processing 
pipeline implemented Apache Flink for complex event 
processing, maintaining sub-50ms latency for the 99.5th 
percentile of transactions while processing concurrent data 
streams from multiple production lines. 

5.3 Real-time data visualization implementation  
The visualization implementation adopted a component-
based architecture using React.js (v18.2.0) for dynamic user 
interfaces and D3.js combined with Apache ECharts for 
complex data visualizations. The dashboard framework 
implemented WebSocket connections, maintaining 12 frames 
per second (FPS) update rates per concurrent user session, 
exceeding the 10 FPS threshold required for perceived real-
time responsiveness.  

 
Figure 5. Real-time data visualization dashboard interface 

As shown in Figure 5, the implementation comprised 
four integrated dashboard views: production overview 
displaying single-line real-time data with key performance 
indicators, quality monitoring featuring SPC control charts 
and defect analysis, equipment status with interactive facility 
layout visualization, and predictive warning systems with 
temporal forecasting displays. The visualization pipeline 
implemented intelligent data reduction techniques to manage 
rendering complexity while maintaining visual clarity. 
Temporal aggregation algorithms compressed high-
frequency sensor data into 5-second, 1-minute, and 5-minute 
windows based on user zoom levels. Spatial clustering 
techniques grouped related equipment data points, reducing 
visual clutter while preserving critical information density. 
The implementation incorporated progressive disclosure 
patterns, revealing additional detail layers through user 
interaction rather than overwhelming initial views. 

User interaction features included drill-down 
capabilities enabling navigation from facility-level overviews 
to individual equipment details, configurable alert thresholds 
with visual highlighting of out-of-range conditions, and role-
based dashboard customization supporting operator, 
supervisor, and executive personas. Mobile responsiveness 
was achieved through adaptive layouts optimized for tablets 
and smartphones, maintaining functionality across 4G 
network conditions with 180ms average response times.  

 

 



ChengHsien Tsai et al. /Future Technology                                                                     February 2026| Volume 05 | Issue 01 | Pages 263-277 

271 

 

The visualization framework is integrated with existing 
Business Intelligence (BI) tools through standardized data 
export formats, enabling advanced analytics in Tableau and 
Power BI environments. 

5.4 Case study: Manufacturing facility application  
The case study deployed the AI-assisted flexible MES on 

five lines producing electronic enclosures with a high variety: 
>150 Stock Keeping Units (SKUs) and batch sizes of 50–5,000 
units. The environment posed notable challenges—frequent 
changeovers (~15 per day), mixed-model assembly, and 
stringent quality targets of <1,000 Parts Per Million (PPM). A 
staged approach was adopted, starting with a pilot on Line 1 
selected for a representative mix and moderate complexity. 
The 30-day pilot established baselines and validated 
performance under production conditions. Key Performance 
Indicators (KPIs) tracked included Overall Equipment 
Effectiveness (OEE), changeover time, first-pass yield, and 
energy per unit. After achieving 82% OEE versus a 60% 
baseline, the rollout expanded to the remaining lines over 12 
weeks. Each deployment incorporated lessons learned, 
reducing per-line implementation time from 15 days on Line 
1 to 7 days on Line 5. 

Operational scenarios demonstrated system flexibility 
through rapid response to dynamic conditions. During a 
critical customer order requiring an 87.5% reduction in 
standard changeover time, the AI scheduling optimizer 
reconfigured production sequences, grouped similar 
products, and pre-positioned materials, achieving 15-minute 
changeovers compared to the previous 120-minute standard. 
Quality emergencies were addressed through real-time SPC 
monitoring, with the system detecting process drift 25 
minutes before traditional control limits would trigger, 
preventing the production of 1,250 potentially defective 
units. The predictive maintenance system successfully 
identified bearing degradation in the injection molding 
machine INJ-003 eighteen hours before failure, enabling 
scheduled maintenance during planned downtime. 

5.5 System performance evaluation  
System performance evaluation employed 

comprehensive metrics validating achievement of design 
targets across all critical dimensions. Real-time processing 
metrics confirmed sub-10ms data collection latency for high-
speed sensors and 42ms average stream processing delay, 
enabling true real-time decision support. The evaluation 
methodology incorporated continuous monitoring via 
embedded agents, stress testing under maximum load 
conditions, and statistical validation using paired t-tests with 
significance levels at α = 0.05. The system demonstrated 
robust scalability, supporting 5,832 concurrent device 
connections while maintaining 99.9% availability over 72 
hours of continuous operation, exceeding the initial design 
specifications by 16.6% in connection capacity. Section 6 
presents comprehensive results across five evaluation 
dimensions and discusses the findings in relation to research 
questions and industry benchmarks. 

6. Results and discussion 

Results addressing research questions: This section 
presents comprehensive evaluation results organized to 
address the four research questions posed in Section 1. 
RQ1 (Real-time Performance): Achieved - System throughput 
12,500 messages per second, data collection latency 8.5ms, 
stream processing 42ms, 5,832 concurrent connections, 
99.9% availability (Section 6.1). 

RQ2 (Performance Impacts): Productivity +25%, defects -
65.7%, changeover time -87.5%, OEE +22 points, all with p 
less than 0.001 (Section 6.2). 
RQ3 (Deployment Challenges): Eight challenges documented 
with solutions - data completeness 85% to 98.5%, AI override 
rate 40% to 12%, 127 vulnerabilities remediated (Section 
6.4). 
RQ4 (Trade-off Resolution): Simultaneous flexibility and 
efficiency improvements confirmed (r = 0.12, p = 0.43), 
challenging traditional inverse relationship theory (Sections 
6.2-6.3). 

6.1 System performance results  
The comprehensive evaluation demonstrated 

exceptional performance across all critical metrics. As shown 
in Table 4, the system achieved or exceeded all target 
specifications. Real-time data collection performance 
exceeded targets with sub-10ms latency, enabling effective 
cyber-physical synchronization [32]. Stream processing 
exceeded design targets with Kafka throughput at 12,500 
messages/second and Flink latency at 42ms [33]. AI inference 
achieved sub-200ms latency across all models, meeting real-
time requirements [34]. Visualization responsiveness 
exceeded industry standards at 12 FPS, with the system 
supporting 5,832 concurrent connections  [35]. 

6.2 Production efficiency improvement  
The implementation yielded substantial improvements 

across all production efficiency dimensions, demonstrating 
the transformative potential of AI-integrated flexible 
manufacturing systems. As shown in Table 5 and Figure 6, the 
system delivered measurable enhancements in productivity, 
quality, flexibility, and resource utilization. Production 
efficiency metrics revealed a 25% increase in daily average 
output from 1,200 to 1,500 units per day, significantly 
exceeding the industry average improvement of 15%. As 
illustrated in Figure 6(a), all five production lines 
demonstrated consistent 25% improvements, with Line 3 
achieving the highest absolute output of 1,625 units per day. 
Capacity utilization improved by 17 percentage points to 
reach 85%, surpassing the 80% benchmark of excellent 
companies. 

Quality indicators demonstrated exceptional gains with 
first-pass yield increasing by 2.3 percentage points to 98.8%, 
approaching world-class levels of 99%+. As shown in Figure 
6(b), quality variability reduced dramatically with the 
standard deviation decreasing from ±1.2% to ±0.3% after 
implementation. The defect rate measured in Parts Per 
Million (PPM) decreased from 35,000 to 12,000, representing 
a 65.7% reduction. Equipment efficiency improvements were 
particularly striking, with OEE increasing by 22 percentage 
points from 60.0% to 82.0%, approaching world-class 
benchmarks of 85%. This improvement magnitude exceeds 
typical AI-driven manufacturing enhancements (10-20 
percentage points reported in literature) due to several case-
specific factors that created exceptional improvement 
potential: 
(1) Baseline Performance Gap: The pre-implementation OEE 
of 60% was substantially below industry norms (75-80% for 
discrete manufacturing), indicating significant latent 
improvement opportunities. The facility had operated with 
reactive maintenance and manual scheduling for over a 
decade, resulting in accumulated inefficiencies ripe for 
optimization. 

 



ChengHsien Tsai et al. /Future Technology                                                                     February 2026| Volume 05 | Issue 01 | Pages 263-277 

272 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Table 4. System performance key indicators test results 

Test Item Test Conditions Test Method Target Value Measured Value Compliance 
Status 

Data Collection 
Performance 

     

Sensor Sampling Frequency 20 main sensor groups 24h monitoring 10-100Hz Main 100Hz, 
Auxiliary 10-50Hz 

✓ Compliant 

Sensor Coverage Plant-wide deployment Coverage test >95% 2350 collection 
points, 98% 

coverage 

✓ Exceeds 

Data Collection Latency High-load scenario Timestamp test <10ms 8.5 ms ± 1.2 ms ✓ Better than 
target 

Protocol Conversion Delay OPC UA/Modbus/MQTT E2E test <20ms 15.3ms ✓ Compliant 
Stream Processing 
Performance 

     

Kafka Throughput 3 Topics, 10 partitions Stress test 10000 msg/s 12500 msg/s ✓ Exceeds by 
25% 

Flink Processing Latency 3 parallel pipelines RT monitoring <50ms 42ms average ✓ Compliant 
Data Aggregation Delay 5s/1min/5min windows Perf analysis <100ms 78ms ✓ Compliant 
AI Inference Performance      
LSTM Prediction Delay Batch size 32 GPU test <200ms 156ms ✓ Compliant 
Anomaly Detection Response Isolation Forest Real-time data stream <100ms 85ms ✓ Compliant 
Image Recognition 
Processing 

CNN model 1080p images <500ms 420ms ✓ Compliant 

Visualization Response      
Dashboard Refresh Rate 20 concurrent users Frontend test 10 FPS 12 FPS ✓ Compliant 
Large Screen Rendering 
Delay 

4K resolution Chrome DevTools <100ms 95ms ✓ Compliant 

Mobile Response Time 4G network Real device testing <200ms 180ms ✓ Compliant 
System Capacity      
Concurrent Device 
Connections 

Simulated 5000 devices Load balancing test >5000 5832 ✓ Exceeds by 
16.6% 

Data Storage Rate Time-series data write InfluxDB stress test 100k points/s 125k points/s ✓ Exceeds by 
25% 

Query Response Time 1-month historical data SQL query test <3s 2.4s ✓ Compliant 
System Stability      
72-hour Stress Test Full load operation Continuous monitoring No crashes 0 crashes ✓ Compliant 
Memory Leak Detection Long-term operation JVM monitoring <5% growth 2.3% growth ✓ Compliant 
CPU Usage Normal load System monitoring <70% 62% average ✓ Compliant 

 

 

Figure 6. Comparison of production efficiency before and after implementation: (a) Production rate improvement: Y-axis in units/day, 

baseline 1200 → 1500 (+25%), (b) Quality improvement: Y-axis in PPM (Parts Per Million), defects 35,000 → 12,000 (-65.7%). (c) OEE 

Components: Percentage scale, Availability 75%→92%, Performance 85%→91%, Quality 94%→97%. (d) Flexibility Metrics: Y-axis in 

minutes, changeover time 120→15 min, order response 1440→360 min, exception handling 30→5 min 

 



ChengHsien Tsai et al. /Future Technology                                                                     February 2026| Volume 05 | Issue 01 | Pages 263-277 

273 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

(2) Availability Improvements (75% to 92%, +17pp): 
Predictive maintenance dramatically reduced unplanned 
downtime. The LSTM-based failure prediction system (94.2% 
accuracy, 72-hour warning window) enabled scheduled 
maintenance during planned downtime, reducing unplanned 
stops by 85%. The 87.5% changeover time reduction (120 to 
15 minutes) further increased available production time. The 
22-point OEE improvement aligns with academic literature 
reporting 15-25 percentage point gains in comprehensive 
digital transformation initiatives. 
 
 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
As shown in Figure 6(c) (Overall Equipment Effectiveness, 
OEE), this improvement resulted from coordinated 
enhancements across availability (75% to 92%), performance 
(85% to 91%), and quality (94% to 97%). Flexible 
manufacturing capabilities showed the most dramatic 
improvements. Figure 6(d) (Flexible manufacturing response 
time) illustrates the waterfall effect of time reductions across 
five key scenarios, with product changeover time decreasing 
by 87.5% from 120 to 15 minutes, approaching Single-Minute 
Exchange of Die (SMED) targets. 

 

 

Table 5. Production efficiency improvement key indicator comparison 

Improvement 
Dimension 

Specific Indicator Before 
Implementation 

After 
Implementa

tion 

Improvement 
Range 

Industry Benchmark 

Production Efficiency      
Daily Average Output 

(units/day) 
5-line average 1200 1500 +25.0% Industry average 

+15% 

Capacity Utilization Actual/Theoretical 
capacity 

68% 85% +17 percentage 
points 

Excellent companies 
80% 

Production Cycle Time Average time per unit 45.2 seconds 36.2 seconds -19.9% Industry leading 35 
seconds 

Quality Indicators      
First Pass Yield Monthly average 96.5% 98.8% +2.3 percentage 

points 
World-class 99%+ 

Defect Rate (PPM) PPM value 35000 12000 -65.7% Six Sigma <3400 
Total Defect Rate Percentage 3.5% 1.2% -2.3 percentage 

points 
Industry excellent 

<2% 

Rework Rate Rework volume/Total 
output 

2.8% 0.9% -67.9% Industry excellent 
<1% 

Equipment Efficiency      
Equipment OEE Overall efficiency 60.0% 82.0% +22 percentage 

points 
World-class 85% 

- Availability Operating 
time/Planned time 

75% 92% +17 percentage 
points 

Target >90% 

- Performance Actual/Standard 
speed 

85% 91% +6 percentage 
points 

Target >95% 

- Quality Good units/Total 
output 

94% 97% +3 percentage 
points 

Target >99% 

Mean Time Between 
Failures (MTBF) 

Hours 168 420 +150% Industry excellent 
>400 

Mean Time To Repair 
(MTTR) 

Minutes 45 12 -73.3% Target <15 minutes 

Flexible Manufacturing      
Product Changeover Time Average changeover 

time 
120 minutes 15 minutes -87.5% SMED target <10 

minutes 
Order Response Time Order to delivery 24 hours 6 hours -75.0% Industry leading 4 

hours 
Exception Handling Time Discovery to 

resolution 
30 minutes 5 minutes -83.3% Real-time response <5 

minutes 

Planning Adjustment 
Time 

Rescheduling time 90 minutes 15 minutes -83.3% Agile manufacturing 
<20 minutes 

New Product Introduction 
Cycle 

Design to production 15 days 3 days -80.0% Rapid prototyping 2-5 
days 

Small Batch Production 
Capability 

Minimum batch size 500 units 50 units -90.0% One-piece flow 
production 

Energy Efficiency      
Unit Energy Consumption kWh/unit 2.85 2.14 -24.9% Green manufacturing 

<2.0 
Energy Utilization Rate Effective 

consumption/Total 
consumption 

72% 88% +16 percentage 
points 

Energy saving target 
>85% 

Inventory Management      
Work-in-Process 

Inventory 
Turnover days 5.2 days 2.1 days -59.6% JIT target <2 days 

Raw Material Inventory Turnover times/year 12 24 +100% Lean target >20 
Finished Goods Inventory Inventory value 

reduction 
Baseline -45% -45% Industry excellent -

40% 

 



ChengHsien Tsai et al. /Future Technology                                                                     February 2026| Volume 05 | Issue 01 | Pages 263-277 

274 

 

6.3 Comparative analysis  
A comparative analysis with traditional MES 

implementations and contemporary intelligent 
manufacturing systems reveals the distinctive advantages of 
the AI-assisted, flexible architecture. Traditional MES 
typically achieves 10-15% productivity improvements and 5-
10% quality enhancements, while the proposed system 
delivered 25% productivity gains and 65.7% defect reduction 
[36]. This performance differential stems from the integration 
of real-time AI inference capabilities with adaptive scheduling 
algorithms, enabling proactive rather than reactive 
manufacturing management. When benchmarked against 
recent intelligent manufacturing implementations, the 
system demonstrates competitive advantages in several key 
areas. Recent studies of cloud-based MES report average 
response times of 200-500ms for critical operations, while 
the implemented system maintains sub-100ms latency for 
99.5th percentile transactions [37]. The ability to process 
12,500 messages per second significantly exceeds typical 
industry implementations handling 5,000-8,000 messages 
per second, enabling more granular process monitoring and 
control. 

Cost-benefit analysis reveals superior Return on 
Investment (ROI) compared to traditional automation 
approaches. While initial implementation costs were 35% 
higher than conventional MES due to AI infrastructure 
requirements, the payback period was reduced to 18 months 
compared to the industry average of 36 months. Total Cost of 
Ownership (TCO) analysis over five years indicates 40% 
lower operational costs due to reduced downtime, improved 
quality, and decreased maintenance expenses [38]. The 
modular microservices design enables selective upgrades and 
targeted technology adoption without system-wide 
disruption, supporting evolutionary rather than disruptive 
transformation. Interoperability with legacy systems while 
introducing advanced capabilities mitigates a major barrier to 
Industry 4.0 adoption, particularly for small and medium-
sized enterprises operating under capital constraints. 

6.4 Challenges and solutions 
Implementation surfaced challenges across technical, 

organizational, and operational domains, requiring adaptive 
remedies. Technically, data quality and integration 
complexity dominated. Initial sensor streams exhibited 15% 
missing values and 8% anomalies, motivating a robust 
preprocessing pipeline with advanced interpolation and 
outlier detection. Cascaded validation at edge nodes reduced 
central processing by 60% and raised completeness to 98.5%. 
System integration was hindered by heterogeneous protocols 
and legacy constraints. Equipment from multiple vendors 
relied on proprietary interfaces, necessitating 12 custom 
adapters. A universal translation layer standardizing on OPC 
UA enabled seamless connectivity while preserving vendor-
specific optimizations. 

Organizational resistance to AI-guided actions required 
structured change management. Early operator skepticism 
produced a 40% override rate. Deploying explainable AI 
views that exposed decision rationales lowered overrides to 
12% within three months. Continuous training—hands-on 
workshops and success-story sharing—further improved 
acceptance. Computational resource pressure emerged 
during peaks: concurrent inference pushed GPU utilization to 
95%. Dynamic allocation based on priority queuing and 
model complexity maintained latency within bounds. An 
edge-cloud hybrid distributed loads and reduced central GPU 
requirements by 45%. 

Cybersecurity challenges required specialized 
mitigation: (i) AI model integrity threats mitigated through 
cryptographic signing and blockchain-based provenance 
tracking (3 tampering attempts blocked); (ii) data exfiltration 
risks addressed via AES-256 encryption, TLS 1.3, and network 
micro-segmentation; (iii) real-time control attacks prevented 
using anomalous command detection (7 suspicious 
sequences identified). Following NIST Cybersecurity 
Framework and IEC 62443 standards, the system 
implemented continuous authentication, least-privilege 
access control (237 operator accounts, 45 PLC service 
accounts), network segmentation via software-defined 
networking, and behavioral analytics detecting 12 anomalous 
access patterns. 

6.5 Theoretical Implications  
The research contributes significant theoretical 

advancements to manufacturing systems theory by 
demonstrating the viability of distributed intelligence 
architectures for achieving flexible automation. The 
successful integration of AI cognitive capabilities with 
traditional MES functions validates the conceptual 
framework of cognitive manufacturing systems, extending 
CPS theory beyond simple automation to encompass adaptive 
learning and autonomous optimization [39]. The findings 
challenge existing assumptions regarding the trade-off 
between flexibility and efficiency in manufacturing systems. 
Traditional theory posits inverse relationships between these 
objectives, yet the implemented system achieved 
simultaneous improvements in both dimensions through AI-
mediated dynamic optimization. This suggests a need to 
reconceptualize manufacturing system design principles, 
incorporating intelligence as a fundamental rather than 
auxiliary component [40]. The research establishes new 
theoretical constructs for understanding human-AI 
collaboration in manufacturing contexts. The observed 
evolution from initial resistance to productive partnership 
suggests staged acceptance models requiring further 
theoretical development. These findings contribute to 
emerging theories of augmented intelligence in industrial 
applications. 

6.6 Practical implications for industry  
The demonstrated success provides actionable insights 

for manufacturing practitioners considering intelligent 
system implementations. Organizations should prioritize 
data infrastructure development before AI deployment, as 
data quality directly impacts system effectiveness. The 
phased implementation approach, beginning with pilot 
deployments on representative production lines, reduces risk 
while building organizational capabilities and confidence 
[41]. Investment strategies should balance immediate 
automation needs with long-term flexibility requirements. 
The modular architecture approach enables incremental 
capability addition without wholesale system replacement, 
protecting capital investments while maintaining 
technological currency. Manufacturing leaders should 
allocate 20-30% of digitalization budgets to workforce 
development, as human factors significantly influence 
implementation success [42]. Strategic partnerships with 
technology providers accelerate implementation while 
reducing technical risks. However, organizations must 
maintain internal competencies in system architecture and 
data management to avoid vendor lock-in and ensure 
sustainable competitive advantages. The development of 
cross-functional teams combining operational expertise with 
data science capabilities proves essential for maximizing AI-



ChengHsien Tsai et al. /Future Technology                                                                     February 2026| Volume 05 | Issue 01 | Pages 263-277 

275 

 

driven manufacturing benefits. Small and medium 
manufacturers can leverage cloud-based deployment models 
to access advanced capabilities without prohibitive 
infrastructure investments. The demonstrated scalability 
from single-line pilots to multi-line deployments provides a 
roadmap for gradual digital transformation aligned with 
business growth and market opportunities. 

6.7 Research limitations  
The research exhibits several limitations requiring 

acknowledgment for the appropriate interpretation of 
findings. The implementation occurred within a single 
manufacturing facility producing electronic enclosures, 
potentially limiting generalizability to other manufacturing 
contexts. Process-intensive industries with continuous 
production may experience different implementation 
challenges and benefit profiles. The evaluation period of 30 
days for pilot testing and 12 weeks for full implementation 
may not capture long-term performance variations or 
degradation patterns. Seasonal demand fluctuations, 
equipment aging effects, and evolving worker expertise could 
influence sustained performance metrics. Extended 
longitudinal studies would provide more comprehensive 
performance assessments [43]. Technical limitations include 
dependence on high-quality sensor data and reliable network 
connectivity. Manufacturing environments with harsh 
conditions or limited infrastructure may face additional 
implementation barriers not addressed in this research. The 
computational requirements for real-time AI inference may 
prove prohibitive for resource-constrained organizations, 
suggesting a need for further optimization research [44]. 
Section 7 synthesizes these findings into conclusions, 
articulates principal contributions, and identifies future 
research directions. 

7. Conclusion  
This research successfully developed and implemented 

an AI-assisted flexible manufacturing execution system that 
addresses critical limitations of traditional MES architectures 
in the Industry 4.0 era. The proposed framework, integrating 
real-time data visualization, digital twin technology, and 
distributed AI intelligence, achieved all design objectives 
while demonstrating superior performance metrics across 
multiple dimensions. The implementation significantly 
exceeded industry benchmarks across all performance 
metrics. The research contributes theoretical advancements 
by establishing cognitive manufacturing systems as a viable 
extension of cyber-physical systems theory, demonstrating 
that distributed intelligence architectures can achieve 
simultaneous improvements in both flexibility and efficiency, 
challenging traditional trade-off assumptions. For 
practitioners, the modular microservices architecture and 
phased implementation approach provide a practical 
roadmap for digital transformation, particularly beneficial for 
small and medium enterprises seeking evolutionary rather 
than revolutionary change. While the evaluation period and 
single-facility implementation present limitations regarding 
long-term performance assessment and cross-industry 
generalizability, the demonstrated benefits justify continued 
investigation. Future research should focus on developing 
industry-specific optimization algorithms and exploring 
federated learning approaches for multi-site deployments 
while maintaining data privacy and competitive advantages 
in increasingly connected manufacturing ecosystems. 

 

 

Ethical issue 
The authors are aware of and comply with best practices in 
publication ethics, specifically regarding authorship 
(avoidance of guest authorship), dual submission, 
manipulation of figures, competing interests, and compliance 
with research ethics policies. The authors adhere to 
publication requirements that the submitted work is original 
and has not been published elsewhere. 

Data availability statement 
The manuscript contains all the data. However, more data will 

be available upon request from the authors. 

Conflict of interest 

The authors declare no potential conflict of interest. 

References 

[1] A. Shojaeinasab et al., "Intelligent manufacturing 

execution systems: A systematic review," Journal of 

Manufacturing Systems, vol. 62, pp. 503-522, 2022, 

doi: 10.1016/j.jmsy.2022.01.004. 

[2] A. Tariq, S. A. Khan, W. H. But, A. Javaid, and T. 

Shehryar, "An IoT-enabled real-time dynamic 

scheduler for flexible job shop scheduling (FJSS) in an 

industry 4.0-based manufacturing execution system 

(MES 4.0)," IEEE Access, vol. 12, pp. 49653-49666, 

2024, doi: 10.1109/ACCESS.2024.3378838. 

[3] L. Allen, J. Atkinson, D. Jayasundara, J. Cordiner, and P. 

Z. Moghadam, "Data visualization for Industry 4.0: A 

stepping-stone toward a digital future, bridging the 

gap between academia and industry," Patterns, vol. 2, 

no. 5, 2021, doi: 10.1016/j.patter.2021.100266. 

[4] H. Wang, Z. Yang, Q. Zhang, Q. Sun, and E. Lim, "A 

digital twin platform integrating process parameter 

simulation solution for intelligent manufacturing," 

Electronics, vol. 13, no. 4, p. 802, 2024, doi: 

10.3390/electronics13040802. 

[5] S. Mantravadi and C. Møller, "An overview of next-

generation manufacturing execution systems: How 

important is MES for industry 4.0?," Procedia 

manufacturing, vol. 30, pp. 588-595, 2019, doi: 

10.1016/j.promfg.2019.02.083. 

[6] R. Beregi, G. Pedone, B. Háy, and J. Váncza, 

"Manufacturing execution system integration through 

the standardization of a common service model for 

cyber-physical production systems," Applied 

Sciences, vol. 11, no. 16, p. 7581, 2021. 

[7]  T. Dieguez, M. T. Malheiro, N. Leal, and J. Machado, 

"Systematic Literature Review on Manufacturing 

Execution Systems in the Era of Industry 4.0: A 

Bibliometric Analysis," in International Conference 

Innovation in Engineering, 2025: Springer, pp. 298-

310, doi: 10.1007/978-3-031-94484-0_24.  

[8] H. ElMaraghy, L. Monostori, G. Schuh, and W. 

ElMaraghy, "Evolution and future of manufacturing 

systems," Cirp Annals, vol. 70, no. 2, pp. 635-658, 

2021, doi: 10.1016/j.cirp.2021.05.008. 

[9] M. Liu, S. Fang, H. Dong, and C. Xu, "Review of digital 

twin about concepts, technologies, and industrial 

applications," Journal of manufacturing systems, vol. 

58, pp. 346-361, 2021, doi: 

10.1016/j.jmsy.2020.06.017. 



ChengHsien Tsai et al. /Future Technology                                                                     February 2026| Volume 05 | Issue 01 | Pages 263-277 

276 

 

[10] C. Su, X. Tang, Q. Jiang, Y. Han, T. Wang, and D. Jiang, 

"Digital twin system for manufacturing processes 

based on a multi-layer knowledge graph model," 

Scientific Reports, vol. 15, no. 1, p. 12835, 2025, doi: 

10.1038/s41598-024-85053-0. 

[11] D. G. Lindberg, "Harnessing AI for smart 

manufacturing: insights from Industry 4.0," Discover 

Artificial Intelligence, vol. 5, no. 1, p. 111, 2025, doi: 

10.1007/s44163-025-00363-0. 

[12] J. F. Arinez, Q. Chang, R. X. Gao, C. Xu, and J. Zhang, 

"Artificial intelligence in advanced manufacturing: 

current status and future outlook," Journal of 

Manufacturing Science and Engineering, vol. 142, no. 

11, p. 110804, 2020, doi: 10.1115/1.4047855. 

[13] Y. Zhang, G. Zhang, J. Wang, S. Sun, S. Si, and T. Yang, 

"Real-time information capturing and integration 

framework of the internet of manufacturing things," 

International Journal of Computer Integrated 

Manufacturing, vol. 28, no. 8, pp. 811-822, 2015, doi: 

10.1080/0951192X.2014.900874. 

[14] J. Wang, C. Xu, J. Zhang, and R. Zhong, "Big data 

analytics for intelligent manufacturing systems: A 

review," Journal of Manufacturing Systems, vol. 62, 

pp. 738-752, 2022, doi: 10.1016/j.jmsy.2021.03.005. 

[15] T. Zigart, G. Kormann-Hainzl, H. Lovasz-Bukvova, M. 

Hölzl, T. Moser, and S. Schlund, "From lab to industry: 

lessons learned from the evaluation of augmented 

and virtual reality use cases in the Austrian 

manufacturing industry," Production & 

Manufacturing Research, vol. 11, no. 1, p. 2286345, 

2023, doi: 10.1080/21693277.2023.2286345. 

[16] V. M. Tabim, N. F. Ayala, G. A. Marodin, G. B. Benitez, 

and A. G. Frank, "Implementing Manufacturing 

Execution Systems (MES) for Industry 4.0: 

Overcoming buyer-provider information 

asymmetries through knowledge sharing dynamics," 

Computers & Industrial Engineering, vol. 196, p. 

110483, 2024, doi: 10.1016/j.cie.2024.110483. 

[17] X. Wang, Y. Guo, and Y. Gao, "Unmanned autonomous 

intelligent system in 6G non-terrestrial network," 

Information, vol. 15, no. 1, p. 38, 2024, doi: 

10.3390/info15010038. 

[18] A. Hevner and S. Chatterjee, Design research in 

information systems: theory and practice. Springer 

Science & Business Media, 2010. 

[19] K. Peffers, T. Tuunanen, M. A. Rothenberger, and S. 

Chatterjee, "A design science research methodology 

for information systems research," Journal of 

management information systems, vol. 24, no. 3, pp. 

45-77, 2007, doi: 10.2753/MIS0742-1222240302. 

[20] A. Bompota, "A theoretical and empirical comparative 

analysis of waterfall, agile and hybrid project 

management methods in digital projects: The case of 

consolut," M.S. thesis, Dept. Applied Informatics, Univ. 

of Macedonia, Thessaloniki, Greece, 2023. [Online]. 

Available: 

https://dspace.lib.uom.gr/handle/2159/29724  

[21] K. L. Yépez Zambrano, "Tecnologías de 

comunicaciones: Prototipo para adquisición de datos 

de un sistema de semáforos para administración de 

tráfico vehicular en una ciudad inteligente," B.S. 

thesis, Facultad de Ingeniería Eléctrica y Electrónica, 

Escuela Politécnica Nacional, Quito, Ecuador, 2024.  

[22] K. Kaur, S. Garg, G. Kaddoum, S. H. Ahmed, and M. 

Atiquzzaman, "KEIDS: Kubernetes-based energy and 

interference driven scheduler for industrial IoT in 

edge-cloud ecosystem," IEEE Internet of Things 

Journal, vol. 7, no. 5, pp. 4228-4237, 2019, doi: 

10.1109/JIOT.2019.2939534. 

[23] T. P. Raptis and A. Passarella, "A survey on networked 

data streaming with apache kafka," IEEE access, vol. 

11, pp. 85333-85350, 2023, doi: 

10.1109/ACCESS.2023.3303810. 

[24] P. Strelec, T. Horak, S. Kovac, E. Nemlaha, and P. 

Tanuska, "IIoT device prototype design using state 

machine according to OPC UA," IEEE Access, vol. 10, 

pp. 134004-134017, 2022, doi: 

10.1109/ACCESS.2022.3233147. 

[25] M. Benková, D. Bednárová, and G. Bogdanovská, 

"Process capability evaluation using capability 

indices as a part of statistical process control," 

Mathematics, vol. 12, no. 11, p. 1679, 2024, doi: DOI: 

10.3390/math12111679. 

[26] P. Reali, R. Lolatto, S. Coelli, G. Tartaglia, and A. M. 

Bianchi, "Information retrieval from 

photoplethysmographic sensors: A comprehensive 

comparison of practical interpolation and breath-

extraction techniques at different sampling rates," 

Sensors, vol. 22, no. 4, p. 1428, 2022, doi: 

10.3390/s22041428. 

[27] X. Sun, Y. He, D. Wu, and J. Z. Huang, "Survey of 

distributed computing frameworks for supporting big 

data analysis," Big Data Mining and Analytics, vol. 6, 

no. 2, pp. 154-169, 2023, doi: 

10.26599/BDMA.2022.9020014. 

[28] D. Kibira, K. C. Morris, and S. Kumaraguru, "Methods 

and tools for performance assurance of smart 

manufacturing systems," Journal of Research of the 

National Institute of Standards and Technology, vol. 

121, p. 282, 2016, doi: 10.6028/jres.121.013. 

[29] Prometheus Authors, "Prometheus - Monitoring 

system & time series database," 2017. [Online]. 

Available: https://prometheus.io/ 

[30] X. Lv, L. Shi, Y. He, Z. He, and D. K. Lin, "Joint 

optimization of production, maintenance, and quality 

control considering the product quality variance of a 

degraded system," Frontiers of Engineering 

Management, vol. 11, no. 3, pp. 413-429, 2024, doi: 

10.1007/s42524-024-3103-1. 

[31] International Electrotechnical Commission, 

"Functional safety of 

electrical/electronic/programmable electronic 

safety-related systems," IEC 61508, 2010. [Online]. 

Available: https://webstore.iec.ch/publication/5515 

[32] H. Xu, W. Yu, D. Griffith, and N. Golmie, "A survey on 

industrial Internet of Things: A cyber-physical 

systems perspective," Ieee access, vol. 6, pp. 78238-

78259, 2018, doi: 10.1109/ACCESS.2018.2884906. 

[33] S. Deng, H. Zhao, W. Fang, J. Yin, S. Dustdar, and A. Y. 

Zomaya, "Edge intelligence: The confluence of edge 

computing and artificial intelligence," IEEE Internet 

of Things Journal, vol. 7, no. 8, pp. 7457-7469, 2020, 

https://webstore.iec.ch/publication/5515


ChengHsien Tsai et al. /Future Technology                                                                     February 2026| Volume 05 | Issue 01 | Pages 263-277 

277 

 

doi: 

10.1109/JIOT.2020.298488710.1109/JIOT.2020.298

4887. 

[34] E. Sisinni, A. Saifullah, S. Han, U. Jennehag, and M. 

Gidlund, "Industrial internet of things: Challenges, 

opportunities, and directions," IEEE transactions on 

industrial informatics, vol. 14, no. 11, pp. 4724-4734, 

2018, doi: 10.1109/TII.2018.2852491. 

[35] T. Hai et al., "Task scheduling in cloud environment: 

optimization, security prioritization and processor 

selection schemes," Journal of Cloud Computing, vol. 

12, no. 1, p. 15, 2023, doi: 10.1186/s13677-022-

00374-7. 

[36] R. X. Gao, J. Krüger, M. Merklein, H.-C. Möhring, and J. 

Váncza, "Artificial Intelligence in manufacturing: State 

of the art, perspectives, and future directions," CIRP 

Annals, vol. 73, no. 2, pp. 723-749, 2024, doi: 

10.1016/j.cirp.2024.04.101. 

[37] C. Li, P. Zheng, Y. Yin, B. Wang, and L. Wang, "Deep 

reinforcement learning in smart manufacturing: A 

review and prospects," CIRP Journal of Manufacturing 

Science and Technology, vol. 40, pp. 75-101, 2023, 

doi: 10.1016/j.cirpj.2022.11.003. 

[38] J. Lee, M. Azamfar, and J. Singh, "A blockchain enabled 

Cyber-Physical System architecture for Industry 4.0 

manufacturing systems," Manufacturing letters, vol. 

20, pp. 34-39, 2019, doi: 

10.1016/j.mfglet.2019.05.003. 

[39] Y. He, B. Wu, J. Mao, W. Jiang, J. Fu, and S. Hu, "An 

effective MID-based visual defect detection method 

for specular car body surface," Journal of 

Manufacturing Systems, vol. 72, pp. 154-162, 2024, 

doi: 10.1016/j.jmsy.2023.11.014. 

[40] A. Setyadi, S. Soekotjo, S. D. Lestari, S. Pawirosumarto, 

and A. Damaris, "Trends and opportunities in 

sustainable manufacturing: a systematic review of 

key dimensions from 2019 to 2024," Sustainability, 

vol. 17, no. 2, p. 789, 2025, doi: 10.3390/su17020789. 

[41] S.-C. Chen, H.-M. Chen, H.-K. Chen, and C.-L. Li, "Multi-

objective optimization in Industry 5.0: Human-centric 

AI integration for sustainable and intelligent 

manufacturing," Processes, vol. 12, no. 12, p. 2723, 

2024, doi: 10.3390/pr12122723. 

[42] J. Pochmara and A. Świetlicka, "Cybersecurity of 

industrial systems—A 2023 report," Electronics, vol. 

13, no. 7, p. 1191, 2024, doi: 

10.3390/electronics13071191. 

[43] T. Khan, U. Khan, A. Khan, C. Mollan, I. Morkvenaite-

Vilkonciene, and V. Pandey, "Data-driven digital twin 

framework for predictive maintenance of smart 

manufacturing systems," Machines, vol. 13, no. 6, p. 

481, 2025, doi: 10.3390/machines13060481.  

[44] Y. Lu, X. Xu, and L. Wang, "Smart manufacturing 

process and system automation–a critical review of 

the standards and envisioned scenarios," Journal of 

Manufacturing Systems, vol. 56, pp. 312-325, 2020, 

doi: 10.1016/j.jmsy.2020.06.010. 

 

  

This article is an open-access article distributed under the 

terms and conditions of the Creative Commons Attribution 

(CC BY) license 

(https://creativecommons.org/licenses/by/4.0/). 

https://creativecommons.org/licenses/by/4.0/

