Railway turnout system RUL prediction based on feature fusion and genetic programming
Introduction
Due to the ever-increasing demand on equipment reliability and qualities, modern mechanical/electrical systems in railways, e.g., power systems [1], electric multiple unit (EMU) [2], and railway turnout system (RTS) [3], [4], [5], [6], [7] are overwhelming more and more complex [8], [9]. To monitor their operation status, there are a lot of sensors installed in these systems since a slight hitch can cause a devastating catastrophe with potentially profound impacts or huge losses.
RTS is an integrated system consisting of complex mechanical and electrical devices, which allows trains to transfer from one track to another. In general, the harsh environment often causes RTS failures, which may lead to catastrophic consequences to equipment and environment such as property loss, human injury and death, and severe damage. According to the statistics of a railway corporation in China, RTS failure accounts for over 40% of the railway signaling equipment failures [10]. Therefore how to predict the turnout’s failure and keep the system in a safe mode in advance becomes an important and urgent issue to be addressed. In order to solve these problems, many scholars at home and abroad are focusing on providing comprehensive frameworks or approaches for RTS remaining useful life prediction for prognostic and health management (PHM). These frameworks or approaches usually include three steps, i.e., feature evaluation and selection, feature fusion & health indicator (HI) construction, and remaining useful life (RUL) estimation.
Feature evaluation and selection is a key step in RUL estimation [11]. There are usually three metrics for feature evaluation and selection, i.e., trendability, monotonicity, and robustness. For example, in [12], Liao points out that an ideal feature which best represents the fault evolution process should have a continuous upward or downward trend, i.e., monotonicity. According to this standard, an HI is built, and an RUL prediction work is employed on a bearing. In [13] another metric, trendability, is proposed for feature selection and simplify a specific prognostic model. Using these two criteria, Saidi presented a methodology based on support vector regression to realized fault prediction of wind turbines. For more meticulous quantitative evaluation of features and HI, a new criterion (i.e., robustness) is added to the metric mentioned before. The new metric measures the resistance to noise. Through comprehensive utilization of all above metrics, Duong successfully selected features and formed the proper HI, and verified it on two public benchmark bearing datasets for RUL prediction [14]. Similar to rolling bearings, the incipient failures of railway turnout systems are usually caused by wearing or degradation of mechanical components, e.g., stock rails and slide chair plates. Hence, the widely used monotonicity, trendability, and robustness in the field of rolling bearings also become important evaluation metrics for the health index of turnout systems. In addition, some existing works [15], [16] also made use of these three metrics to construct optimal prognostic metrics of railway turnout systems and achieved promising performances. Moreover, although there are different feature extraction and selection methods that were proposed for HI construction, few of them have a generic feature selection technique covering feature extraction, smoothing, feature rank, and optimal feature subset choice to build a good HI. Thus, developing more feature evaluation and selection methods is necessary for HI construction for improving the accuracy of RUL estimation.
The main goal of feature fusion is to construct a generic machine HI to enhance the information content related to the system degradation [17]. Since a single feature may only contain partial degradation information [18], so it is beneficial to fuse different features to contain more complementary information about the health state of the degrading machine. Baraldi et al. proposed a novel fusion method based on Auto-Associative Kernel Regression (AAKR) in [19], in which the features collected from health operation conditions are constructed as a weighted sum, and the failure detection is qualified by the similarity between current signal and the reconstructions. Zhang et al. [20] proposed an intersection feature fusion method for overcoming shortcomings of the simple features, and the simulation results show its superiority. Based on simulated dataset Song proposed another multi-sensor feature fusion method for the prediction of degradation trajectory of aircraft engines [21]. Another type of feature fusion is based on AI algorithms. In [22], Zhao et al. divided the last hidden layer of an enhanced auto encoder (EAE) into several groups as a deep feature, and a weighted fusion of them is conducted to represent the degradation process of bearings optimally. Chen [23] applied sparse auto encoder (SAE) for rotating machinery fault diagnosis. However, these AI and deep learning methods often pose challenges for model parameters tuning and computational complexity.
During the last few years, a lot of RUL estimation models have been designed based on physical models [24], ANN [25], and deep learning [26]. However, ANN and deep learning-based methods are considered as black-box, which lack actual physical meanings due to the opacity of model structures and have difficulty in choosing optimal parameters.
Some researchers had successfully applied Genetic programming (GP) to other time-domain prognosis and data analysis [27], [28]. Unfortunately, few existing works focus on explicit model relationships for RTS life models, which hindered from truly understanding the life model or degradation process. For scenarios requiring high reliability, this uncertainty and opacity may be fatal.
In order to solve the aforementioned problems, this paper proposes a novel RUL estimation method for RTS based on feature selection & fusion and Genetic Programming. The main contributions of this work include:
- (1)
A novel feature extraction and selection method for prediction of RUL of turnout is proposed, which includes the time-domain features extraction, local smoothing, and feature importance evaluation considering three inherit metrics (i.e., monotonicity, trendability, and robustness) and consistency. The proposed feature selection method takes advantages of the locally weighted regression to avoid the overfitting in feature smoothing. Moreover, combining the feature importance evaluation based on the inherit three metrics and the choice of the optimal number of features based on the consistency metric, the proposed approach can obtain a more effective and computationally efficient feature subset.
- (2)
An AAKR-based feature fusion approach is utilized. Compared to the traditional fusion methods, such as ANN-based methods, this feature fusion method is more computationally efficient.
- (3)
A GP-based RUL estimation is conducted. Instead of obtaining the prognostic model in “black-box-like” methods, which lack actual physical meanings, the proposed method may provide the explicit mapping relationships between RUL and HIs and provide a reliable and effective RUL estimation result.
The rest of this paper is organized as follows. In Section 2 railway turnout system and its monitoring measures are briefly introduced. The methodology for constructing HI, denoise technology, and the RUL prediction algorithm are detailedly illustrated in Section 3. Section 4 shows the experimental results. Our conclusions are made in Section 5.
Section snippets
Railway turnout system and conditional monitor
Railway turnout system, which is a basic signaling facility of railways, is used for the transformation of trains between different tracks. Take a single RTS as an example, as shown in Fig. 1, it includes blades, stock rails, and turnout machines, etc. Also, RTS is equipped with a variety of sensors to obtain monitoring data, such as current (indicates the health states of the RTS circuit), voltage (indicates conversion voltage), proximity signal (for gap detection) and force (indicates the
Methodology
The general framework of this prognostic method is shown in Fig. 2. In the proposed framework, (i) the RTS’s monitoring signals are firstly fed into the feature extraction & smoothing module. (ii) Then eight signal statistical features can be obtained, e.g., root mean square and standard deviation. (iii) Next, the feature selection module selects the best subset of these statistic features by a variant correlation-based feature selection (VCFS) process. (iv) After that, the selected features
Turnout simulating dataset experiment
RTS degradation caused by conversion resistance can be observed by monitoring signals such as force and power. In the following parts, we apply force signals from the simulated datasets to verify the proposed method, and power signals from on-site condition monitoring datasets to test the proposed method.
Conclusions
In this paper, we develop an RUL prediction model for RTS based on the feature fusion method and genetic programming, which can effectively reflect the degradation process of RTS. To extract features from the conversion force and power signal of RTS, different time-domain based feature extraction methods are applied. Moreover, we utilize run-to-failure trajectories from both simulated and real condition monitoring datasets to conduct the experiments. In addition, a comparative experiment
Fundings
This work was supported by the Project of Beijing Education Commission (No. I17H100010) and Subsidized Project by China Railway Electrification Bureau (No. I19L00180).
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgement
This research was supported by Beijing Laboratory of Urban Rail Transit and National Engineering Laboratory for Urban Rail Transit Communication and Operation Control.
References (39)
- et al.
A bayesian network model for prediction of weather-related failures in railway turnout systems
Expert Syst. Appl.
(2017) - et al.
A review on data-driven fault severity assessment in rolling bearings
Mech. Syst. Signal Process.
(2018) - et al.
Differential evolution-based multi-objective optimization for the definition of a health indicator for fault diagnostics and prognostics
Mech. Syst. Signal Process.
(2018) - et al.
Wind turbine high-speed shaft bearings health prognosis through a spectral kurtosis-derived indices and svr
Appl. Acoust.
(2017) - et al.
Degradation-level assessment and online prognostics for sliding chair failure on point machines
IFAC-PapersOnLine
(2018) - et al.
A neural network approach for remaining useful life prediction utilizing both failure and suspension histories
Mech. Syst. Signal Process.
(2010) - et al.
Hygp-msam based model for slewing bearing residual useful life prediction
Measurement
(2019) - et al.
Wavelet filter-based weak signature detection method and its application on rolling element bearing prognostics
J. Sound Vibr.
(2006) - et al.
Rolling element bearing diagnostics in run-to-failure lifetime testing
Mech. Syst. Signal Process.
(2001) - et al.
Machinery health prognostics: a systematic review from data acquisition to rul prediction
Mech. Syst. Signal Process.
(2018)
Predictive diagnosis of high-power transformer faults by networking vibration measuring nodes with integrated signal processing
IEEE Trans. Instrum. Meas.
Bond graph modeling and fault injection of crh5 traction system
A simple state-based prognostic model for railway turnout systems
IEEE Trans. Industr. Electron.
Gap measurement of point machine using adaptive wavelet threshold and mathematical morphology
Sensors
Adaboost and least square based failure prediction of railway turnouts
Failure cause extraction of railway switches based on text mining
A review: Prognostics and health management
J. Electronic Measure. Instru.
Research on switch fault detection and health assessment method on svdd
J. Southwest Jiaotong Univ.
Discovering prognostic features using genetic programming in remaining useful life prediction
IEEE Trans. Industr. Electron.
Cited by (43)
Evolutionary modeling approach based on multiobjective genetic programming for strip quality prediction
2024, Swarm and Evolutionary ComputationFeature fusion model based health indicator construction and self-constraint state-space estimator for remaining useful life prediction of bearings in wind turbines
2023, Reliability Engineering and System SafetyLearning of physical health timestep using the LSTM network for remaining useful life estimation
2022, Reliability Engineering and System SafetyCitation Excerpt :Kumar et al. [42] developed and compared singular value decomposition, average value of the cumulative feature, and Mahalanobis distance to build a HI of the rolling bearing elements. Chen et al. [43] proposed Auto-Associative Kernel Regression to fuse selected features into a HI of the railway turnout systems. Shi and Chehade [44] proposed a Dual-LSTM framework to estimate a novel HI with change point detection for predicting the RUL of aircraft engines that have different initial degradation conditions.
Relation between prognostics predictor evaluation metrics and local interpretability SHAP values
2022, Artificial IntelligencePump remaining useful life prediction based on multi-source fusion and monotonicity-constrained particle filtering
2022, Mechanical Systems and Signal ProcessingCitation Excerpt :To further illustrate the performance advantages of the proposed MAAKR fusion approach, four multi-source fusion approaches based on the original AAKR are introduced to carry out the same multi-source fusion task for comparison. These four fusion approaches include Euclidean distance-based AAKR reported in [21], Mahalanobis distance-based AAKR claimed by Baraldi et al. [20], cosine distance-based AAKR, and AAKR based on Euclidean distance and cosine distance. The fusion results of the four comparative approaches are successively shown in Fig. 13a-d.
Investigations on generalized Hjorth's parameters for machine performance degradation assessment
2022, Mechanical Systems and Signal Processing