Paper Title (use style: paper title) Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 729 https://internationalpubls.com Mathematical Framework for Enhancing Machine Failure Prediction in Aviation and Beyond: Leveraging Deep Convolutional Neural Networks and Visual Analytic Jalajakshi V1,Dr. Myna A.N2 1Research scholar, Dept. of CSE, Navkis College of Engineering, Hassan. Email: jalajakshi.prakash@gmail.com 2Vice Principal, Research Supervisor, Dept. of CSE, Navkis College of Engineering, Hassan Email: : mynaan@gmail.com Google Scholar ID: PwtzkUMAAAAJ Article History: Received: 31-05-2024 Revised: 30-06-2024 Accepted: 22-07-2024 Abstract: In industries such as aviation, accurate prediction of machine failures is essential to prevent costly and potentially hazardous breakdowns. Traditional methods rely on detailed degradation models of components, which are often not available, posing challenges to precise predictions. To overcome this limitation, the study proposes the use of deep convolutional neural networks (DCNN), sophisticated algorithms capable of learning from extensive datasets without the need for intricate degradation models. DCNNs employ a time window approach to discern significant patterns in data, eliminating the requirement for specialized knowledge in failure prediction and increasing their effectiveness across diverse scenarios. Evaluation using aircraft engine data illustrates that DCNNs surpass traditional methods in reliably predicting failures, representing a significant advancement in predictive maintenance. The study utilizes scatterplots and histograms to visually analyze the distribution of fault cycles among engine units. One scatter plot subplot displays each engine unit as a point, revealing patterns or trends in fault cycles. The other subplot features a histogram illustrating the frequency of fault cycles across all units, including statistical measures such as minimum, average, and maximum cycles. These visualizations provide a thorough understanding of fault cycle distribution, aiding in deeper analysis and interpretation of engine performance data. The performance evaluation compares several models: Linear Regression, Random Forest, Improved Linear Regression, Improved Random Forest, and the Cox Proportional Hazards model. Evaluating RMSE (Root Mean Squared Error), MAE (Mean Absolute Error), and R² values for both training and test datasets reveals that the Improved Random Forest and Improved Linear Regression models demonstrate lower RMSE and MAE values and higher R² values. This indicates superior predictive accuracy and data fit. The Cox Proportional Hazards model offers unique insights into risk factors but may vary in performance depending on the dataset, emphasizing the importance of model refinement and tuning for precise predictions. This comprehensive approach underscores the potential of DCNNs and enhanced traditional models to advance predictive maintenance across industries. Keywords: RUL, DCNN, NASA, MAPSS, PHM, CBM, MAE. 1. INTRODUCTION A. Remaining Useful Life RUL estimation is similar to predicting the future performance of machines. Imagine knowing exactly how long your car engine or a specific piece of equipment will operate efficiently before needing Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 730 https://internationalpubls.com repair or replacement. This insight helps businesses improve maintenance planning by identifying the optimal times to address potential issues, thus preventing expensive and disruptive unexpected breakdowns. It's like being aware of when your phone battery will run out, allowing you to recharge it before it completely dies. By analyzing historical data on the wear and tear of machines over time, RUL estimation enables cost and time savings while ensuring smooth operations. Remaining Useful Life (RUL) is a pivotal concept in predictive maintenance and reliability engineering, representing the estimated duration that a machine, component, or system can operate before needing repair or replacement. Precise RUL predictions empower organizations to schedule maintenance proactively, preventing unforeseen failures, reducing downtime, and optimizing maintenance schedules. This proactive approach is especially beneficial in industries where equipment reliability is paramount, including aviation, manufacturing, and energy sectors. Estimating RUL involves examining historical and real-time data collected from diverse sensors and operational parameters. Advanced AI techniques such as deep convolutional neural networks (DCNNs) play a crucial role in this analysis. These models are trained on extensive datasets like NASA's Predictive Maintenance (RUL) Database, which contains detailed records of machinery operations and failures. By detecting patterns and anomalies within the data, DCNNs can deliver highly precise RUL predictions. This predictive capability not only enhances maintenance planning but also enhances safety and efficiency by allowing potential issues to be addressed proactively before they escalate. As technology continues to advance, integrating sophisticated predictive models with comprehensive datasets will further refine the accuracy and relevance of RUL predictions across various industries. B. Importance of RUL C. Motivation 2. LITERATURE SURVEY A. Survey Accurately assessing the remaining useful life (RUL) of lithium batteries is crucial in battery management systems for efficient management, optimization of performance, and improving user experience. Current deep learning methods for estimating battery RUL primarily rely on individual neural networks, often neglecting important spatial-temporal characteristics. In this research, we present a time series prediction framework that utilizes a gated recurrent unit (GRU) combined with a spatial-temporal attention mechanism. Our approach begins with screening data highly correlated with battery capacity using correlation coefficient analysis. The screened data is subsequently processed by a series of GRU network layers to illustrate the relationship between battery data and capacity. After this, the extracted features undergo a deeper analysis, where a spatial-temporal attention mechanism assigns varying weights to these features across temporal and spatial dimensions, thereby capturing complex relationships more effectively. We tested the proposed method on the NASA dataset, and the experimental results confirm that our network achieves accurate time series predictions with a mean absolute error (MAE) of less than 1.5%.[1] Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 731 https://internationalpubls.com Estimating the Remaining Useful Life (RUL) is crucial in prognostics and health management (PHM) as it effectively enhances reliability, increases availability, and reduces costs. However, assessing the uncertainty level in deep learning-based RUL prediction models is challenging, and stochastic process-based models struggle with handling complex and large datasets. Therefore, this paper proposes a Wiener process model for RUL prediction that integrates a Transformer neural network. First, the historical data is filtered and smoothed, and the degradation trend in the processed historical data is extracted using the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) technique. Based on the obtained degradation pattern, the training and test datasets have been developed using moving window techniques to train an optimized Transformer neural network. The trained Transformer neural network is then used to adaptively identify the Wiener drift coefficient function. Subsequently, the analytical probability density function of the RUL is derived based on the first passage time (FPT). The proposed method is demonstrated using publicly available lithium-ion battery capacity degradation data from NASA.[2] Remaining useful life (RUL) estimation is crucial for predictive maintenance. Due to the limited availability of labeled data caused by infrequent failures, semisupervised RUL estimation methods have been developed to leverage unlabeled data collected during maintenance intervals to enhance performance. However, these methods often fall short because they only partially utilize unlabeled data during training. This article introduces a novel semisupervised method for RUL estimation, where both the main and auxiliary tasks focus on estimating relative RULs. It employs a geometric progression to draw more unlabeled samples from shorter RULs and utilizes different strategies for pairing unlabeled samples at various training stages. The proposed method is evaluated using the NASA C-MAPSS dataset and the Backblaze HDD dataset. Results demonstrate that this approach significantly improves performance with limited labeled data for training. [3] As a crucial energy source, the 18650 lithium-ion battery is widely used in electric vehicles. Predicting the remaining useful life (RUL) of these batteries is essential for the reliable operation of electric vehicles. Traditional methods often struggle with adaptively estimating model parameters and detecting capacity regeneration, areas that require further investigation. To address these challenges, this article introduces a novel approach combining expectation maximization, unscented particle filter, and Wilcoxon rank sum test (EM-UPF-W) to adaptively estimate noise variables in the degradation model and accurately detect battery capacity regeneration. Specifically, for scenarios involving small unlabeled samples, this article develops a dynamic degradation model for a single battery using the UPF framework, which adaptively estimates noise variables with the help of the EM algorithm. Additionally, the Wilcoxon rank sum test is applied to determine the capacity recovery point, thereby reducing prediction error. The methodology is demonstrated using a 18650 lithium-ion battery dataset provided by NASA. Experimental results indicate that the proposed EM-UPF-W approach outperforms several existing data-driven techniques. [4] With the widespread adoption of deep learning in condition monitoring system prognostics, its lack of interpretability has always been a concern. This article introduces an interpretable method for remaining useful life (RUL) prediction. The proposed approach comprises an augmenter network based on ordinary differential equations and an estimator network employing feature-temporal attention. The augmenter network is designed to mitigate additive noise in the original data and to Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 732 https://internationalpubls.com infer unobservable health-related variables using embedded formulas. Following this, an uncertainty- aware estimator is employed to predict RUL through a quantile regression module, simultaneously identifying key features and temporal dependencies via attention mechanisms. Extensive evaluations are conducted using the N-CMAPSS aero engine dataset. . Our method surpasses baseline approaches, achieving a 14% improvement in the NASA scoring metric and a 7% reduction in root-mean-square error. Furthermore, we delve into the interpretability of our framework. The proposed method can infer physical constraints, detect anomalous subsystems, and pinpoint critical flight stages even with limited prior knowledge. [5] Lithium-ion batteries are ubiquitous today, powering everything from small electronic devices to large electric vehicles. Consequently, there's a growing focus on assessing the remaining useful life (RUL) of lithium-ion batteries to predict failures and ensure safe operation. However, most battery RUL research neglects testing models for execution on low-power edge nodes, the ideal target devices. This paper proposes an Edge-Based RUL Estimation (EBRULE) approach leveraging deep learning. EBRULE enables battery RUL prediction directly on an edge device. Four deep neural network architectures were trained on two datasets, NASA and CALCE, using a GPU. Subsequently, inference was performed on a Raspberry Pi single-board computer. Our experiments demonstrate that the models consistently delivered predictions with an accuracy exceeding 89% within seconds. Furthermore, our approach achieved a RMSE of 0.02-6.81% and a MAE of 0.8-5.67%. Notably, the DeTransformer model significantly outperformed the baseline DeTransformer method, reducing the MAE and RMSE by up to 71.95% and 77.31% respectively. Similarly, the RNN model achieved impressive reductions in MAE and RMSE by 78.89% and 90.77% compared to the baseline RNN approach.[6] Maintaining an accurate estimate of the satellite's attitude is essential for pointing performance to ensure the quality of products such as spatial images. Nevertheless, if sensors and actuators are affected by the severe space environment, the attitude control systems may be disrupted. The South Atlantic Anomaly (SAA) is a well-known disturbance in space whose effects are consistently monitored by space agencies such as NASA. This work provides a straightforward yet robust estimation strategy that enhances the attitude determination system's (ADS) performance when sensor-filter communication loss occurs due to the SAA. Our method utilizes the statistical properties of SAA to adapt the attitude estimator's gain following measurement restoration. This technique helps reduce the attitude innovation commonly regarded as an indicator of ADS health. We used data collected by attitude sensors on an actual satellite in low earth orbit to validate our method. [7] While deep neural network (DNN) architectures have been widely explored for assessing the State of Health (SOH) of Lithium-ion batteries (LIBs), the influence of hyperparameters on their performance remains under-investigated. This lack of understanding makes it challenging to evaluate the suitability of different DNN architectures and to effectively build upon or modify previous research. This study delves into the impact of hyperparameters on the performance of feedforward neural networks for LIB SOH estimation. SpecificallyThis study proposes two feed-forward neural networks: an accuracy model specifically optimized for highly accurate SOH prediction of individual cells, and a generalized model optimized for predicting SOH across various cells. To illustrate the effects of hyperparameters on these models, we leverage the National Aeronautics and Space Administration Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 733 https://internationalpubls.com (NASA) prognostic battery dataset. We meticulously clean this dataset to remove anomalies and analyze its features. Our investigations reveal that the accuracy model achieves a root mean square error of 0.33%, demonstrating its effectiveness in predicting the health state of individual cells. Furthermore, we compare our findings with existing research, establishing a benchmark for future studies. Our open resources, including code, cleaned data, feature analysis, and experimental reports, are publicly available on GitHub (https://github.com/jhrrlee/batterystates). [8] B. Understanding from survey Enhancing machine failure prediction in avionics and beyond leverages a powerful combination: deep convolutional neural networks (DCNNs) and visual analytics. This approach integrates advanced AI techniques with sophisticated data visualization tools. DCNNs excel at handling large, complex datasets, making them adept at uncovering patterns and anomalies in machinery data that traditional methods might overlook. By training on historical failure data and operational parameters, DCNNs can generate highly accurate predictions of potential failures. This capability is critical in aviation, where equipment malfunctions can be catastrophic. Integrating DCNNs enables real-time monitoring and predictive maintenance, leading to significant reductions in downtime and maintenance costs, while simultaneously bolstering safety and reliability. Visual analytics serve as a powerful complement to DCNNs. These techniques provide intuitive and interactive data visualizations, aiding engineers and decision-makers in deciphering complex data and model outputs. By visually highlighting patterns, trends, and anomalies identified by DCNNs, stakeholders can gain rapid insights and make well-informed decisions. This synergy between deep learning and visual analytics strengthens the overall predictive maintenance framework, promoting a proactive approach to machine health management. Beyond aviation, these technologies hold similar promise for industries like manufacturing and transportation. By enabling the anticipation of failures before they occur and facilitating timely interventions, they can contribute to improved efficiency, reduced costs, and enhanced safety. 3. METHODOLOGY A. Nasa predictive Maintenance (RUL) Database NASA's C-MAPSS dataset is a valuable resource that fosters advancements in predictive maintenance and prognostics. This comprehensive dataset offers time-series data on machinery operation, maintenance interventions, and failure events. This rich data foundation empowers researchers to develop and rigorously test prognostic algorithms. The C-MAPSS dataset focuses on predicting the remaining useful life (RUL) of various components and systems. This focus empowers more effective maintenance scheduling and minimizes the risk of unforeseen failures. Researchers and engineers leverage this data to train and validate machine learning models, such as deep convolutional neural networks (DCNNs). These models can analyze historical and real-time data to predict when a machine or component is nearing failure. The RUL database boasts a rich variety of machinery and operational scenarios, mirroring the versatility of predictive maintenance across diverse industries. Take the aviation sector for instance, where the database houses data from aircraft engines. Here, pinpointing RUL with high accuracy is paramount for both safety and optimizing maintenance schedules. The dataset incorporates a diverse range of sensor readings, operational parameters, and failure modes, https://github.com/jhrrlee/batterystates Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 734 https://internationalpubls.com offering a rich information pool for developing advanced predictive models. By capitalizing on the RUL database, researchers can propel the continuous improvement of reliable predictive maintenance frameworks. This ultimately translates to enhanced operational efficiency, reduced maintenance costs, and increased safety in avionics and other critical industries. Proposed Flow chart Fig 1: Proposed Flowchart Following an extensive literature review, this work presents a systematic approach for implementing predictive maintenance solutions across industries, as outlined in the proposed model flowchart (Figure 1). The process commences with data acquisition. Here, relevant data on machine operation, maintenance logs, and historical failure events are gathered from various sources. This comprehensive data collection phase equips the predictive maintenance system with rich and diverse datasets, crucial for generating accurate predictions. After collection, the data undergoes rigorous preparation and preprocessing to guarantee its quality and suitability for modeling. This critical step involves tasks like data cleaning, feature engineering, and normalization, all aimed at enhancing the dataset's integrity and relevance for machine learning. Once the data is prepared, the next stage focuses on model evaluation. Here, various predictive Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 735 https://internationalpubls.com models, including deep learning algorithms like convolutional neural networks (CNNs), are trained, tested, and validated using appropriate performance metrics to determine the most effective approach. After selecting the optimal model based on its performance metrics, visualization techniques are employed to translate the results into actionable insights. Subsequently, the validated model is deployed in real-world industrial settings. Here, it empowers proactive identification of potential machine failures, allowing for optimized maintenance scheduling and minimized downtime. Ultimately, this leads to improved operational efficiency and cost savings. Results and Discussion This study leverages a scatter plot, visualized in the top portion of Figure 2, to depict the distribution of fault cycles across engine units. Each data point on the scatter plot represents an individual engine, with the horizontal axis (x-axis) indicating the engine unit number and the vertical axis (y-axis) representing the corresponding number of fault cycles. This visualization allows us to identify any patterns or trends in how fault cycles vary among the different engine units. Fig 2: Distribution of the number of fault cycles across different engine units The bottom portion of Figure 3 displays a histogram visualizing the distribution of fault cycles across all engine units. The horizontal axis (x-axis) represents the number of fault cycles, while the vertical axis (y-axis) shows the number of engine units within each range (bin). Additionally, vertical dashed lines highlight key statistical measures such as the minimum, average, and maximum number of fault cycles. These lines provide insights into the center point (central tendency) and the spread (variability) of the distribution. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 736 https://internationalpubls.com Fig 3: Distribution of the number of fault cycles across all engine units Figure 4 visualizes the distribution of RUL values within the training and testing sets. The training set exhibits a wider spread of RUL values, indicating a variety of degradation patterns and life spans across the components. This distribution likely leans towards lower RUL values, suggesting a higher concentration of components nearing failure. In contrast, the test set might display a more concentrated distribution of RUL values, potentially clustered around mid-range or lower values. This reflects the real-world scenarios the model aims to predict. This distinction ensures the model encounters diverse failure patterns during training, ultimately enhancing its robustness and accuracy in predicting RUL for the unseen test data. Fig 4: Distribution of RULs for train and test sets Fig 5 & Fig 6 shows Actual vs predicted RUL for test set with Linear Regression & Random Forest respectively. A key assumption in our initial approach was a linear decrease in RUL over time. However, upon closer inspection, we observed that sensor readings in many cases remain constant initially before experiencing a sharp rise or fall. This pattern aligns with the reality of engine degradation, which typically progresses over time. The point where the signal curve deviates from its initial constancy is the first indication of engine deterioration. From this point onward, assuming a linear decline in RUL becomes more justifiable. Prior to this inflection point, accurate RUL assessment is challenging due to the lack of information about initial wear and tear. Figures 7 and 8 illustrate the comparison of actual versus predicted RUL for the test set using the Improved Linear Regression and Improved Random Forest models, respectively. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 737 https://internationalpubls.com Fig 5: Linear Regression Fig 6: Random Forest Fig 7: Improved Linear Regression Fig 8: Improved Random Forest Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 738 https://internationalpubls.com The Cox Proportional Hazards model is a widely used statistical technique in survival analysis for investigating the relationship between the survival time of subjects and one or more predictor variables as shown in Fig 9. Unlike other models, it does not assume a specific baseline hazard function, making it semi-parametric. The Cox proportional hazards model offers remarkable adaptability, effectively handling diverse survival data. This model centers around the concept of a hazard function. This function describes the instantaneous risk of an event happening at a specific time, considering it hasn't happened yet. The model assumes this hazard is a combination of a baseline hazard and an exponential function of the covariates. This enables the Cox model to evaluate how various factors influence the survival rate, providing valuable insights into which variables significantly impact the time until the event of interest occurs, such as engine failure or disease onset. Fig 9: Cox Proportional Hazards models Table 1 showcases the performance metrics for the models, enabling a clear comparison of their effectiveness in RUL prediction. The basic Linear Regression model serves as a benchmark, but more intricate approaches often surpass its performance. The Random Forest model, leveraging an ensemble approach, typically achieves higher accuracy and robustness by effectively capturing non- linear patterns within the data. The Improved Linear Regression model, likely incorporating additional features or regularization methods, demonstrates a significant improvement over the standard version by reducing prediction errors. Likewise, the Improved Random Forest model, potentially optimized through hyper parameters tuning or feature engineering, exhibits superior performance compared to its base counterpart. Finally, the Cox Proportional Hazards model, suited for analyzing time-to-event data, offers valuable insights into the factors influencing RUL. However, its performance can vary depending on dataset characteristics and hazard proportionality assumptions. Overall, the table emphasizes that well-designed and well-tuned models consistently yield more accurate RUL predictions. Table 1: Performance Measure for models Linear Regression Random Forest Improved Linear Regression Improved Random Forest Cox Proportional Hazards models Train RMSE 44.69 36.8 21.52 17.16 26.22 Train MAE 34.13 25.81 17.61 12.64 22.48 Train R2 0.58 0.71 0.73 0.83 0.6 Test 32.2 32.33 22.07 18.53 26.61 Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 739 https://internationalpubls.com RMSE Test MAE 25.64 23.98 17.79 13.64 23.08 Test R2 0.4 0.39 0.72 0.8 0.59 This visualization offers a holistic view of how fault cycles are distributed within the dataset. This comprehensive perspective empowers researchers to delve deeper into the engine unit performance data and extract meaningful insights. To evaluate the performance of various models, this study compares Linear Regression, Random Forest, their improved variants, and the Cox Proportional Hazards model. By calculating the Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and R-squared (coefficient of determination) for both training and testing datasets, we can assess the accuracy and robustness of each model An analysis of the results reveals that Improved Random Forest and Improved Linear Regression models consistently achieve lower RMSE and MAE values. This translates to greater accuracy in predictions and reduced errors compared to their standard counterparts. The improved models also exhibit higher R-squared values, signifying a stronger fit to the data. While the Cox Proportional Hazards model offers valuable insights into risk factors, its performance can vary depending on the specific characteristics of the dataset. Overall, these findings highlight the significant impact of model feature enhancements and tuning on improving prediction accuracy. 4. CONCLUSION This study introduces a novel deep learning approach for air motor unit prognostics, achieving high accuracy in predicting remaining useful life (RUL). The proposed method utilizes convolutional neural networks (CNNs) to analyze raw sensor data. To prepare the data for training, the method involves feature selection, preprocessing, and the implementation of a time window technique for sample creation. By comparing the proposed deep convolutional architecture against other neural network structures, this study highlights its effectiveness in prognostic tasks. The research investigates the influence of hidden layer count and time window size on performance, aiming to identify the optimal network configuration. The results demonstrate the method's superior prognostic accuracy when compared to existing approaches. Additionally, the paper explores data preprocessing techniques suitable for real-time applications and proposes future research directions to optimize the architecture and reduce training time. Future scope The future of machine disappointment forecasting using deep convolutional neural networks (DCNNs) holds significant potential for various industries. Researchers can delve deeper by refining the time window approach and exploring different parameters to better capture nuanced patterns and temporal relationships. Furthermore, by combining DCNNs with techniques like recurrent neural networks or reinforcement learning, hybrid approaches could unlock even greater accuracy in predictions. DCNN-based models' scalability and generalizability unlock a wider range of applications. Researchers can tailor architectures to specific domains and leverage transfer learning to expedite deployment. Additionally, incorporating visualization techniques can boost model Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 6s (2024) 740 https://internationalpubls.com transparency and user comprehension, facilitating more informed decision-making in crucial operational settings. In essence, progress in DCNN-powered machine failure prediction paves the way for technical improvements, broader adoption, and enhanced interpretability, ultimately resulting in more precise and actionable predictive maintenance solutions across industries. References [1] Z. Bao, H. Lin, S. Guan, X. Zheng, Z. Li and M. Gao, "Spatial-temporal Attention-Based Time Series Prediction Network for Lithium Battery Remaining Useful Life Estimation," 2024 4th International Conference on Neural Networks, Information and Communication (NNICE), Guangzhou, China, 2024, pp. 91-94, doi: 10.1109/NNICE61279.2024.10498716. [2] J. Ren, J. Zheng, J. Li, H. Dong and Z. Zhang, "A Hybrid Remaining Useful Life Prognosis Method Integrating Transformer Networks and Wiener process," 2023 CAA Symposium on Fault Detection, Supervision and Safety for Technical Processes (SAFEPROCESS), Yibin, China, 2023, pp. 1-5, doi: 10.1109/SAFEPROCESS58597.2023.10295588. [3] M. Natsumeda and T. Yairi, "Consistent Pretext and Auxiliary Tasks With Relative Remaining Useful Life Estimation," in IEEE Transactions on Industrial Informatics, vol. 20, no. 4, pp. 6879-6888, April 2024, doi: 10.1109/TII.2024.3353923. [4] J. Zhang, Y. Jiang, X. Li, H. Luo, S. Yin and O. Kaynak, "Remaining Useful Life Prediction of Lithium-Ion Battery With Adaptive Noise Estimation and Capacity Regeneration Detection," in IEEE/ASME Transactions on Mechatronics, vol. 28, no. 2, pp. 632-643, April 2023, doi: 10.1109/TMECH.2022.3202642. [5] L. Qin, S. Zhang, T. Sun and X. Zhao, "An Interpretable Neuro-Dynamic Scheme With Feature-Temporal Attention for Remaining Useful Life Estimation," in IEEE Transactions on Industrial Informatics, vol. 20, no. 4, pp. 5505-5516, April 2024, doi: 10.1109/TII.2023.3333933. [6] C. J. C. Adducul, J. R. I. Macasaet and N. M. C. Tiglao, "Edge-based Battery Remaining Useful Life Estimation Using Deep Learning," 2023 International Conference on Smart Applications, Communications and Networking (SmartNets), Istanbul, Turkiye, 2023, pp. 1-6, doi: 10.1109/SmartNets58706.2023.10215733. [7] H. Henna, E. Gasmi, H. Toubakh, M. R. Kafi, M. Djemai and M. Sayed-Mouchaweh, "Fault Tolerant Attitude Estimation for Satellite at Low Earth Orbit*," 2023 9th International Conference on Control, Decision and Information Technologies (CoDIT), Rome, Italy, 2023, pp. 275-280, doi: 10.1109/CoDIT58514.2023.10284329. [8] J. Lee, "Li-ion Battery Health State Estimation with Two Feed Forward Neural Networks," 2023 IEEE International Conference on Industrial Technology (ICIT), Orlando, FL, USA, 2023, pp. 1-6, doi: 10.1109/ICIT58465.2023.10143124. [9] Y. Zeng, Y. Zhang, X. Yan and Q. Miao, "EMA Health Indicator Extraction Based on Improved Multivariate State Estimation Technique With a Composite Operator," in IEEE Sensors Journal, vol. 23, no. 17, pp. 19894-19904, 1 Sept.1, 2023, doi: 10.1109/JSEN.2023.3298349.