Acta Polytechnica CTU Proceedings https://doi.org/10.14311/APP.2025.52.0063 Acta Polytechnica CTU Proceedings 52:63–68, 2025 © 2025 The Author(s). Licensed under a CC-BY 4.0 licence Published by the Czech Technical University in Prague ABRUPT CHANGE DETECTION IN RAILWAY NOISE DATA Jan Kruntoráda,∗, Tetiana Reznychenkob, Petr Červenkaa a Czech Technical University in Prague, Faculty of Transportation Sciences, Department of Transport Engineering, Konviktská 20, 110 00 Prague, Czech Republic b Czech Technical University in Prague, Faculty of Transportation Sciences, Department of Applied Mathematics, Konviktská 20, 110 00 Prague, Czech Republic ∗ corresponding author: kruntjan@fd.cvut.cz Abstract. Current methods for diagnosing the quality of the railway superstructure are mainly based on optical sensors, which are relatively expensive compared to acoustic sensors. As part of the HLUKOS research project, a pair of microphones is installed near the wheel-rail contact point on a diagnostic vehicle of the Railway Administration (Czech railway infrastructure manager). The research task is to detect when the sound level changes significantly. A likelihood ratio method has been used in this paper to detect abrupt changes, which is a current scientific topic. Experiments with different input thresholds are performed on a sample of 250 m of track data. Initial experimental results show that this method is meaningfully able to detect locations of abrupt changes with input threshold values h = 4.58 and number of steps from N = 5 to N = 40. Keywords: Abrupt change, change detection, likelihood ratio, railway diagnostics, railway noise, wheel-rail contact. 1. Introduction The last decades have been marked by the devel- opment of technological processes and sophisticated sensors, followed by the need for sophisticated infor- mation processing systems. Addressing these issues is important for a number of reasons, particularly safety, environmental and economic. The large amount of data from a wide range of sensors means that algo- rithms must be found or developed to analyse the information. The core of the interest is the detection of one or more abrupt changes in some characteristic of the system under consideration. An abrupt change is considered to be any change in system parameters that occurs either immediately or at least very quickly relative to the measurement sampling period used. Abrupt changes certainly do not refer to changes of large magnitude; on the con- trary, in most applications the main problem is the detection of small changes. Moreover, in some specific cases, early warning of small (and not necessarily fast) changes is essential to avoid economic or even fatal consequences, such as small malfunctions in the sen- sors of aircraft navigation systems or small variations in the geometric position of the track [1]. The quality of track geometry, respect of clearance profile and other parameters necessary for safe railway operation are continuously monitored and evaluated by infrastructure managers. One of the objectives of the HLUKOS research project is to determine whether abrupt changes in rolling noise recordings at the wheel- rail contact point can be detected, and to use these changes to warn of possible local degradation (defects) of the railway superstructure. 2. Theoretical background Abrupt change detection in noisy signals is a critical area of research that spans various fields, including signal processing, climate science, and power systems. The methods employed for detecting these changes can be broadly categorized into wavelet-based techniques, statistical approaches, and machine learning methods. Each of these methodologies offers unique advantages and is suitable for different types of signals and noise characteristics. 2.1. Current methods used Wavelet transforms have emerged as a powerful tool for detecting abrupt changes in signals, particularly in the presence of noise. The wavelet transform allows for the decomposition of a signal into its constituent components at various scales, making it easier to iden- tify discontinuities or abrupt changes. For instance, Ukil and Živanović [2] demonstrated the effectiveness of using an adjusted Haar wavelet for segmenting power system disturbance signals into event-specific sections, highlighting the method’s robustness in noisy environments. In addition to wavelet methods, statistical tech- niques such as the Kolmogorov-Smirnov statistic have been utilized for abrupt change detection. Qi et al. [3] proposed a fast framework based on binary search trees and the Kolmogorov statistic to identify changes in the statistical properties of signal series, which are indicative of qualitative transitions in the underlying data generation mechanism. This approach is more ef- ficient due to the shortest computation time, the high- est hit rate and accuracy than compared conventional methods Kolmogorov-Smirnov statistic, t-statistic and 63 https://doi.org/10.14311/APP.2025.52.0063 https://creativecommons.org/licenses/by/4.0/ https://www.cvut.cz/en J. Kruntorád, T. Reznychenko, P. Červenka Acta Polytechnica CTU Proceedings singular-spectrum analyses which was tested on real EEG recordings. Another significant method for detecting abrupt changes is the use of Bayesian approaches, particu- larly in online applications such as automatic speech recognition (ASR). Chowdhury et al. [4] emphasized the importance of developing noise estimation algo- rithms that continuously track and estimate noise levels, enabling the detection of abrupt changes in the noise spectrum. This adaptability is crucial in real-time systems where the acoustic environment can vary significantly. In the realm of climate science, various detection methods have been developed to identify regime shifts, which can be classified into categories based on the type of change – mean value, variance, frequency, prob- ability density changes, and multivariable analysis [5]. These methods provide a comprehensive framework for analyzing abrupt changes in climate data, empha- sizing the need for tailored approaches depending on the nature of the data. The integration of machine learning techniques into abrupt change detection has also shown promise. For instance, sparse identification of nonlinear dynamics has been utilized for rapid model recovery in chaotic systems, allowing for efficient detection of abrupt changes with reduced data requirements [6]. This approach exemplifies the trend towards leveraging ad- vanced computational techniques to enhance detection capabilities in complex systems. In summary, the methods for detecting abrupt changes in noisy signals are diverse and multifaceted, encompassing wavelet transforms, statistical tech- niques, Bayesian approaches, and machine learning. Each method has its strengths and is suited to par- ticular types of signals and noise characteristics. In this paper we focus on another technique for detect- ing abrupt changes, namely the use of likelihood ra- tio. It is based on the principle of change detection on difference synthetic aperture radar images [7] or pipeline damage detection using torsional guided wave mode [8]. 2.2. Problem formulation Lets have a dataset which a sequence of signal level observations X = [x1, x2, . . . , xn], where n is a number of observations. The task is to find a sufficient change in the signal level. This is achieved by identifying the intervals at which significant changes in signal strength occur. This allows for the identification of potential issues and an improvement in the accuracy of data analysis. 2.3. Likelihood-ratio change detection Likelihood-ratio change detection is a statistical method that is used to detect a point in time when the distribution of signal yt changes [9, 10]. The goal of the detection algorithm is to determine S(yt) when the distribution shifts from f(yt, θ0) to f(yt, θ1) where it is assumed that θ0 and θ1 are known. At each time step t, the log-likelihood-ratio is com- puted as: S(yt) = log f(yt, θ1) f(yt, θ0) (1) This ratio in Eq. (1) compares the probability of observing the current value of the signal yt under the assumption that a change has already taken place (with parameter θ1) to the probability of observing the same value under the assumption that no change has yet occurred with θ0. The following algorithm solves the detection prob- lem. It detects the change in the level of the signal under test. It is possible to apply this approach to the Fourier transform, thus capturing changes in fre- quency. Initialization. Set the values of the signal yt, θ0, and θ1 – means, σ – standard deviation of the data set. Signal observation. The signal yt is sampled in K, where K is the sequence number of the large period of sampling. In addition, each large period is divided into N small periods, as indicated in Figure 1. Figure 1. Illustration of the signal divided into K intervals [9]. Compute and update statistics. Calculate the likelihood ratio for yt on N -th small interval, and update statistics over small intervals using the alpha coefficient in Eq. (2). The closer α is to 1, the more weight is given to the value of St−1: St = α · St−1 + (θ1 − θ2) σ2 [ yt − θ1 + θ2 2 ] , (2) where St is the current value of the statistic, yt is the current observed value, and α is the coefficient of forgetting where 0 < α < 1. After obtaining the result on the small interval N , the value of SN 1 (K) big interval is computed, as shown in Eq. (3), where SN 1 (K) is the value of the statistic on the K-th interval. SN 1 (K) = (θ1−θ0) σ2 ∑N t=N(K−1)+1 K [ yt − θ1+θ0 2 ] . (3) Threshold test. Compare SN 1 (K) with a predeter- mined threshold h. The decision rule is: d = { 0, if SN 1 (K) < h, 1, if SN i (K) > h, (4) 64 vol. 52/2025 Abrupt change detection in railway noise data where SN 1 (K) = SNK N(K−1)+1 [10]. If SN 1 (K) > h, it is considered that there was a change in the distribution from θ0 to θ1. 3. Railway noise data The purpose of the experiment part is to find intervals where noise significantly and rapidly changes using the likelihood ratio method. Within the framework of the HLUKOS research project, data collection is carried out by means of microphones placed symmetrically on the wheelset of the diagnostic vehicle MVŽSv2 (owned and operated by Czech rail infrastructure manager: Technology and Diagnostics Centre, Railway Admin- istration, State organisation), in such a reference po- sition (see Figure 2), which is closest to the measured wheel and, according to further technical analysis and listening to the sound recording, shows the lowest level of interference by air flow and other ambient phenomena. The key is the choice and correct configuration of the sound analyser, whose function is to process the data recorded by the measuring microphones in real time and convert them into specific acoustic pa- rameters. The problem lies in the combination of the high speed of the diagnostic vehicle (currently max. 160 km h−1, in the near future up to 200 km h−1) and the requirement for very fine sampling of the track parameters. For the solution of this project, a specially designed sensor-acoustic system developed by the project team – EKOLA group, spol. s r.o., and CTU in Prague, Faculty of Transportation Sci- ences – was used. This system enables the collection of acoustic data with an accuracy comparable to conven- tional sound measurement technology at a sampling frequency of up to 48 kHz, which corresponds to a sam- pling duration of 0.02 ms [11]. The study was conducted in a station throat sec- tion with three single rail switches running against the switch blades at a low speed, within the range of 35.6 to 36.8 km h−1. The data set used of 1 001 obser- vations contained two key variables: ID and railway noise values. The ID variable is dimensionless and in- dicates the order of observations, with measurements taken every 0.25 m of track. The ID range is from 0 to 1 000, which means that the measured track sample is 250 m long. The next variable is the noise in dB. These values reflect the amplitude of the Z-network (unweighted) equivalent sound pressure level recorded at corresponding ID points. The range of noise values is from 87.5 to 118.4 dB. 4. Results and discussion A model was defined with the parameters θ0 = 90 and θ1 = 107. These values represent the mean val- ues obtained from the measured signal, which was segmented into two distinct models. The signal was divided into K samples to determine the moment of signal change. Figure 2. Position of the microphone for measuring the contact noise between the rail wheel and the rail on the MVŽSv2 diagnostic vehicle [12]. Table 1 presents the number of samples on which SN 1 (K) changes from θ0 to θ1 for different values of N , where N is 5, 10, 20, 40. The threshold h was established at 4.58. N 5 10 20 40 Total 15 9 5 3 Table 1. Demonstration of the relationship between segment length N and detection sensitivity h. Figures 3–6 presents a comprehensive chart detail- ing the detection of changes in railway noise. The top graph displays the raw signal data, the middle graph illustrates the values of the St statistic on the K-th sample, and the bottom graph visualizes the decision-making process for detecting changes in the signal, based on the analysis of SN 1 (K) and h. The coefficient α was applied during calculations over small periods, preserving α = 0.95 when N = 5 and N = 10, and α = 1.0 when N is 20 and 40. For each combination of parameters N and h, the results have shown the number of detected changes were obtained: • at N = 5 and h = 4.58, changes were identified due to the high sensitivity of the threshold, • at N = 10 and h = 4.58, the sensitivity decreased and the number of false alarms decreased, • at N = 20 and h = 4.58, the lowest number of false alarms and the most accurate detection of changes were recorded, • at N= 40 and h = 4.58, three samples exhibited a change in the signal, and no other notable peaks were identified. This suggests that this specific combination of segment length and threshold might limit the detection of additional signal changes. 5. Conclusions Listening to the sound recording, three clear sound perceptions can be heard in the analysed section of track (at ID = 81, 251 and 834). The abrupt change at ID = 81 was only detected at N = 5; at IDs = 251 and 834 at ID = 834 at all N values examined. 65 J. Kruntorád, T. Reznychenko, P. Červenka Acta Polytechnica CTU Proceedings Figure 3. Abrupt change the visualization of railway noise using likelihood ratio detection: Graph of signal detection with the sample length N = 5. Figure 4. Abrupt change the visualization of railway noise using likelihood ratio detection: Graph of signal detection with the sample length N = 10. 66 vol. 52/2025 Abrupt change detection in railway noise data Figure 5. Abrupt change the visualization of railway noise using likelihood ratio detection: Graph of signal detection with the sample length N = 20. Figure 6. Abrupt change the visualization of railway noise using likelihood ratio detection: Graph of signal detection with the sample length N = 40. 67 J. Kruntorád, T. Reznychenko, P. Červenka Acta Polytechnica CTU Proceedings The experiment demonstrated that the parameter N is essential for the accurate detection of signal changes. Shorter segments (N = 5) enhance sensitiv- ity but increase the likelihood of false alarms. The threshold h must be selected to achieve an optimal balance between sensitivity and accuracy. With a cho- sen threshold of h = 4.58, a value of N in the range 20–40 seems to be optimal. This research paper advances the knowledge base and capabilities for identifying railway noise, paving the way for more effective monitoring and maintenance strategies. Ultimately, this research contributes to the enhancement of railway safety and reliability. Acknowledgements This paper has been financed with the state support of the Technology Agency of the Czech Republic and the Ministry of Transport of the Czech Republic within the project CK03000099 Dynamic opto-acoustic method for railway superstructure emission noise assessment within the TRANSPORT 2020+ Programme. References [1] J. Kruntorád, L. Týfa. Vision Zero application on Czech railways. Acta Polytechnica CTU Proceedings 41:15–19, 2023. https://doi.org/10.14311/APP.2023.41.0015 [2] A. Ukil, R. Živanović. Adjusted Haar wavelet for application in the power systems disturbance analysis. Digital Signal Processing 18(2):103–115, 2008. https://doi.org/10.1016/j.dsp.2007.04.001 [3] J.-P. Qi, J. Qi, Q. Zhang. A fast framework for abrupt change detection based on binary search trees and Kolmogorov statistic. Computational Intelligence and Neuroscience 2016(1):8343187, 2016. https://doi.org/10.1155/2016/8343187 [4] M. F. R. Chowdhury, S.-A. Selouani, D. O’Shaughnessy. Bayesian on-line spectral change point detection: a soft computing approach for on-line ASR. International Journal of Speech Technology 15(1):5–23, 2012. https://doi.org/10.1007/s10772-011-9116-2 [5] Q. Liu, S. Wan, B. Gu. A review of the detection methods for climate regime shifts. Discrete Dynamics in Nature and Society 2016(1):3536183, 2016. https://doi.org/10.1155/2016/3536183 [6] M. Quade, M. Abel, J. Nathan Kutz, S. L. Brunton. Sparse identification of nonlinear dynamics for rapid model recovery. Chaos: An Interdisciplinary Journal of Nonlinear Science 28(6):063116, 2018. https://doi.org/10.1063/1.5027470 [7] H. Zhuang, Z. Tan, K. Deng, G. Yao. Adaptive generalized likelihood ratio test for change detection in SAR images. IEEE Geoscience and Remote Sensing Letters 17(3):416–420, 2020. https://doi.org/10.1109/LGRS.2019.2922198 [8] S. Mariani, P. Cawley. Change detection using the generalized likelihood ratio method to improve the sensitivity of guided wave structural health monitoring systems. Structural Health Monitoring 20(6):3201–3226, 2021. https://doi.org/10.1177/1475921720981831 [9] M. Basseville, I. V. Nikiforov. Detection of Abrupt Changes: Theory and Application. Prentice-Hall, New Jersey, 1993. ISBN 978-0-13-126780-0. [10] K. S. Riedel. Detection of abrupt changes: Theory and application. Technometrics 36(3):326–327, 1994. https://doi.org/10.1080/00401706.1994.10485821 [11] J. Kruntorád, L. Týfa, O. Simon. Analysis of noise emissions at wheel-rail contact. In O. Prentkovskis, I. Yatskiv (Jackiva), P. Skačkauskas, et al. (eds.), TRANSBALTICA XV: Transportation Science and Technology, pp. 502–511. Springer Nature Switzerland, Cham, 2025. ISBN 978-3-031-85390-6. https://doi.org/10.1007/978-3-031-85390-6_47 [12] L. Ládyš, O. Simon, M. Ládyš, et al. CK03000099 Dynamická opto-akustická metoda hodnocení emisní hlučnosti železničního svršku. Interim report 2023, EKOLA group, Prague, 2024. 68 https://doi.org/10.14311/APP.2023.41.0015 https://doi.org/10.1016/j.dsp.2007.04.001 https://doi.org/10.1155/2016/8343187 https://doi.org/10.1007/s10772-011-9116-2 https://doi.org/10.1155/2016/3536183 https://doi.org/10.1063/1.5027470 https://doi.org/10.1109/LGRS.2019.2922198 https://doi.org/10.1177/1475921720981831 https://doi.org/10.1080/00401706.1994.10485821 https://doi.org/10.1007/978-3-031-85390-6_47 Acta Polytechnica CTU Proceedings 52:63–68, 2025 1 Introduction 2 Theoretical background 2.1 Current methods used 2.2 Problem formulation 2.3 Likelihood-ratio change detection 3 Railway noise data 4 Results and discussion 5 Conclusions Acknowledgements References