Railway turnout is one of the critical equipment of Switch & Crossing (S&C) Systems in railway, related to the train’s safety and operation efficiency. With the advancement of intelligent sensors, data-driven fault detection technology for railway turnout has become an important research topic. However, little research in the literature has investigated the capability of data-driven fault detection technology for metro railway turnout. This paper presents a convolutional autoencoder-based fault detection method for the metro railway turnout considering human field inspection scenarios. First, the one-dimensional original time-series signal is converted into a two-dimensional image by data pre-processing and 2D representation. Next, a binary classification model based on the convolutional autoencoder is developed to implement fault detection. The profile and structure information can be captured by processing data as images. The performance of our method is evaluated and tested on real-world operational current data in the metro stations. Experimental results show that the proposed method achieves better performance, especially in terms of error rate and specificity, and is robust in practical engineering applications.
As an essential mode of transportation, urban rail transit has been rapidly developed in many countries. For instance, China’s urban rail transit (URT) has developed into the most prolonged and widespread urban rail transit network worldwide. By the end of 2021, the total length of China’s URT network has reached more than 9,192 kilometers [
Railway turnout is the critical infrastructure component in Switch&Crossing (S&C) Systems in high-speed rail, general-speed rail, and URT, which control the switch of tracks in operation [
Recent intelligent sensor advancements have contributed to data-driven PHM research on railway turnout. The related research contents include gap measurement [
The automatic features with deep learning approach have become an emerging research topic in PHM research [
As described, most existing approaches focus on the turnout of high-speed rail and normal-speed railway. Although some studies have applied the data-driven fault detection method to monitoring URT turnout [ This paper proposes a method for detecting metro turnout faults that applies to solving a practical engineering issue. The proposed data processing method fully considers the real detection scene of metro turnout by preserving the curve data’s profile and structure information. Data from real metro stations validated the proposed method.
This paper is organized as follows: railway turnout and field data description are given in
S&C Systems mainly include rails, actuators, and turnout (switch machine), as shown in
The current curve of ZDJ9 includes A, B, and C three-phase currents with the 380 V three-phase AC power supply. According to the literature [
The A-phase current curve consists of four stages: unlocking, switching, locking, and switching on. The specific description of each stage is as follows:
Stage 1 (unlocking) (T0--T1):
After the switch machine starts, it must overcome the strong resistance to complete the unlocking. Therefore, the motor needs to provide strong power support. The current increase rapidly, showing a prominent pulse peak on the curve. Afterward, power and current return to standard levels.
Stage 2 (conversion) (T1--T2):
The conversion process requires less resistance and relatively more minor power than the unlocking stage. The switch machine provides power to pull the switch to realize the conversion. This process is time-consuming, and the power is maintained at a relatively stable value. The current curve also keeps a smooth straight line with slight fluctuation.
Stage 3 (locking) (T2--T3):
After the conversion process is completed, the position of the switch needs to be fixed, and the tip rail is not allowed to move by an external force. This stage is locking. The locking process time is short, and the curve has no prominent feature.
Stage 4 (slow release) (T3--T4):
The switch state transition is completed when the lock ends and the current is disconnected with the switch action. The circuit is turned on, resulting in a rapid drop in the switch machine’s operating current.
ZDJ9 turnout takes 7–9 s to accomplish once state transition, with the sample rate of 25 Hz. Specifically, the field dataset consisted of 500 normal and 500 fault samples. Each current curve includes 256 sample points.
This paper develops a convolutional autoencoder-based (CAE) method for modeling and identifying turnout using the A-phase current curve. The method’s inputs are A-phase current signals of turnout. The outputs are whether the switch is faulty. As shown in Curve data pre-processing and image generation. The primary purpose of this module is to pre-process curve data to match the input requirements of the convolutional autoencoder-based model. The process involves the cleaning, normalization, and image generation of the current curve data. The MATLAB software is performed to generate images in a specified size. With the processed data, we can complete the next module’s convolution. Classification modeling. A CAE model is designed to identify the current curve. The input of this module is the current curve images, and the module’s output is the trained model for fault detection.
The real on-site turnout detection in URT is to read the images of electrical characteristic curves. The 2D image represents the raw current signals based on this domain knowledge. MATLAB’s ‘plot’ and ‘saveas’ functions are used to generate images from raw current curves. The size of the generated images in this study was set to 32 × 32. The whole process is shown in
To summarize, unlike the previous study [
The convolutional autoencoder is a type of autoencoder that is effective for unsupervised learning. It introduces convolutional operations into the encoding and decoding steps. Through the combination of convolutional operations and autoencoders, it provides powerful feature extraction and unsupervised feature clustering. In encoding, convolution and pooling are used to map high-dimensional image data to low-dimensional feature space. In the decoding part, the feature space is reconstructed and transformed into the original data by deconvolution and unpooling. Intermediate hidden layers can effectively represent the original data and provide features for classifiers [
In practical engineering application scenarios, the sample size is often relatively small. In this situation, we designed the network structure as shown in
The configuration of each layer of the proposed structure is shown in
Let us consider the operation dataset is
Following data processing and image transformation, the input matrix The convolution operation for matrix The encoding part of the convolutional autoencoder is as follows:
Through the convolution operation, the original data is mapped into the feature space
And the pooling operation is denoted as follows:
Then the decoder operation is as follows:
Then the unsupervised loss function is as follows:
The
Then the selected hidden layer features are fed into the fully connected layer for classification. The SoftMax function is used as the activation function of the classifier’s output layer, and the SoftMax function is represented as follows:
The specific formula of the classifier’s output layer is:
The corresponding loss function of the classifier is
The proposed method is validated by the real field current turnout data in the metro system, as described in
The deep learning framework keras accomplish the developed method with python 3.6. All models are implemented on the workstation with NVIDIA RTX 2080 GPU and Intel i7-8700 CPU. There are four evaluation metrics selected for comparison. The metrics are error rate, F1-score, sensitivity, and specificity. The definitions and formulas of these evaluation metrics are shown as follows:
Error rate: the probability of recognition error.
F1-score: a measure of prediction accuracy.
Sensitivity: the probability of a positive sample being predicted to be positive.
Specificity: the probability that a negative sample is predicted to be negative.
where the values TP, FP, TN, and FN correspond to True Positive, False Positive, True Negative, and False Negative, respectively.
Specifically, true positives indicate the number of positive samples predicted correctly. False negatives indicate the number of positive samples predicted as negatives. False positives indicate the number of negative samples predicted as positives, and true negatives indicate the number of negative samples predicted correctly.
To evaluate the performance of the presented method, we compared it with the fault detection models in the existing literature. The comparison models include the convolutional neural networks (CNN) model [
Model | Hyper-parameters | Training range |
---|---|---|
CNN | { |
{[0.0001, 1], [5,25], [5,25]} |
Stacked AE | { |
{[0.0001, 1], [5,25], [5,25]} |
Stacked SAE | { |
{[0.0001, 1], [5,25], [5,25]} |
Proposed CAE | { |
{[0.0001, 1], [5,25], [5,25]} |
Experiment 1: Comparison of different models with changes in the number of epochs.
With other parameters fixed, models’ performances in terms of evaluation metrics in different epoch numbers are shown in
Evaluation indicators | Number of epochs | |||||
---|---|---|---|---|---|---|
5 | 10 | 15 | 20 | 25 | ||
CNN | Error rate (%) | 14.0 | 14.5 | 13.5 | 14.0 | 14.5 |
F1-score (%) | 84.6 | 84.0 | 85.2 | 84.6 | 84.0 | |
Stacked AE | Error rate (%) | 16.5 | 16.0 | 15.0 | 17.0 | 14.5 |
F1-score (%) | 82.3 | 83.0 | 84.2 | 81.7 | 84.8 | |
Stacked SAE | Error rate (%) | 15.0 | 15.5 | 15.0 | 17.5 | 15.5 |
F1-score (%) | 83.3 | 82.7 | 83.3 | 0.8 | 82.7 | |
Proposed CAE | Error rate (%) | 13.5 | 14.0 | 12.5 | 11.5 | 13 |
F1-score (%) | 85.2 | 84.6 | 86.5 | 87.7 | 85.9 |
Evaluation indicators | Number of epochs | |||||
---|---|---|---|---|---|---|
5 | 10 | 15 | 20 | 25 | ||
CNN | Sensitivity (%) | 93.9 | 93.8 | 94.0 | 93.9 | 93.8 |
Specificity (%) | 80.5 | 79.8 | 81.2 | 80.5 | 79.8 | |
Stacked AE | Sensitivity (%) | 88.5 | 88.6 | 88.9 | 88.4 | 89.0 |
Specificity (%) | 79.6 | 80.4 | 81.8 | 78.9 | 82.6 | |
Stacked SAE | Sensitivity (%) | 93.8 | 93.7 | 93.8 | 93.3 | 93.7 |
Specificity (%) | 79.2 | 78.5 | 79.2 | 76.0 | 78.5 | |
Proposed CAE | Sensitivity (%) | 94.0 | 93.9 | 94.1 | 94.3 | 94.0 |
Specificity (%) | 81.2 | 80.5 | 82.6 | 84.1 | 82.0 |
Experiment 2: Comparison of different models with the change in batch size.
With other parameters fixed, models’ performances in evaluation metrics in different batch sizes are shown in
Evaluation indicators | Batch size | |||||
---|---|---|---|---|---|---|
5 | 10 | 15 | 20 | 25 | ||
CNN | Error rate (%) | 14.5 | 14.0 | 13.5 | 14.0 | 15.0 |
F1-score (%) | 84.0 | 84.6 | 85.2 | 84.6 | 84.2 | |
Stacked AE | Error rate (%) | 16.5 | 14.5 | 16.0 | 15.0 | 16.5 |
F1-score (%) | 82.4 | 84.8 | 83.0 | 84.2 | 82.4 | |
Stacked SAE | Error rate (%) | 15.5 | 14.5 | 14.0 | 13.5 | 16.5 |
F1-score (%) | 82.7 | 84.0 | 84.6 | 85.2 | 81.4 | |
Proposed CAE | Error rate (%) | 14.0 | 14.5 | 13.5 | 13.0 | 13.5 |
F1-score (%) | 84.6 | 84.0 | 85.2 | 85.9 | 85.2 |
Evaluation indicators | Batch size | |||||
---|---|---|---|---|---|---|
5 | 10 | 15 | 20 | 25 | ||
CNN | Sensitivity (%) | 93.8 | 93.9 | 94.0 | 93.9 | 88.9 |
Specificity (%) | 79.8 | 80.5 | 81.2 | 80.5 | 81.8 | |
Stacked AE | Sensitivity (%) | 88.5 | 89.0 | 88.6 | 88.9 | 88.5 |
Specificity (%) | 79.6 | 82.6 | 80.4 | 81.8 | 79.6 | |
Stacked SAE | Sensitivity (%) | 93.7 | 93.8 | 93.9 | 94.0 | 93.5 |
Specificity (%) | 78.5 | 79.8 | 80.5 | 81.2 | 77.2 | |
Proposed CAE | Sensitivity (%) | 93.9 | 93.8 | 94.0 | 94.0 | 94.0 |
Specificity (%) | 80.5 | 79.8 | 81.2 | 81.9 | 81.2 |
Experiment 3: Models’ performance under the best parameters.
Under the best parameters, models’ performances in evaluation metrics are shown in
To summarize, the proposed model offers two distinct advantages over the other method. Firstly, the data processing process considers the curve data’s profile and structure information. Meanwhile, the convolution autoencoder model combines the advantages of convolution operation and autoencoder. Consequently, the proposed method performs well compared to other related studies in experiments 1, 2, and 3, demonstrating our approach’s robustness.
Evaluation indicators | Proposed CAE | CNN | Stacked AE | Stacked SAE |
---|---|---|---|---|
Error rate (%) | 8.0 | 8.5 | 11.5 | 9.0 |
F1-score (%) | 91.8 | 91.2 | 87.7 | 90.6 |
Sensitivity (%) | 94.7 | 94.6 | 94.3 | 94.6 |
Specificity (%) | 89.6 | 88.8 | 84.1 | 88.0 |
The samples that failed to be identified in experiment 3 were clustered by k-means clustering and analyzed. Three types of fault samples could not be entirely correctly identified. The specific examples are shown in
As a result, the proposed method is difficult to completely identify the fault curve if the profile or structure is relatively small changes compared to the standard curve.
This study proposes a convolutional autoencoder-based fault detection method for metro railway turnout. The presented method included 1) Curve data pre-processing and image generation and 2) Classification modeling based on the convolutional autoencoder. Furthermore, the main contribution lies in developing a new data-driven fault detection method for metro railway turnout without expert experience-based feature engineering. The proposed method combines the advantages of convolutional operations and autoencoders. Specifically, the developed method is evaluated and validated with real-world operation data. While our approach is highly dependent on the quality of the data, it is difficult to identify some kinds of samples. Furthermore, the proposed method still needs to improve accuracy and interpretability, and incipient fault detection is not considered.
Future work will focus on the model’s accuracy, interpretability, and incipient fault detection. We will improve the model’s accuracy and interpretability by circuit physics modeling and the neural network’s loss functions embedded with the partial differential equations. For the incipient fault detection, a dynamic model of the turnout circuit will be constructed based on circuit parameters, describing the degradation process. Afterward, threshold technology will be used to recognize the incipient fault.
The authors would like to thank the all-round rail transit control system integrator (CASCO) for providing research data and domain knowledge support.