Urban rail transit steel rail ground insulation defect diagnosis method
A deep learning model using CNN-BiLSTM-Attention fusion effectively addresses inefficiencies in rail-to-ground insulation defect detection, enhancing stray current management and safety in urban rail transit systems.
Patent Information
- Application Number
- CN202411234638.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-04
- Publication Date
- 2024-12-27
AI Technical Summary
Existing methods for detecting rail-to-ground insulation defects in urban rail transit systems are inefficient, requiring significant manpower and resources, and can only be conducted during specific times, leading to suboptimal stray current leakage management and safety risks.
A deep learning model utilizing a CNN-BiLSTM-Attention fusion network is employed to diagnose rail-to-ground insulation defects by analyzing rail potential data, incorporating a CNN for spatial feature extraction and BiLSTM for temporal analysis, with an attention mechanism to enhance feature correlation, enabling rapid and accurate defect identification.
The method allows for rapid and accurate diagnosis of insulation defects, reducing stray current leakage, lowering operational costs, and ensuring personnel and equipment safety by improving insulation between rails and ground.
Smart Images
Figure CN119202889A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of stray current corrosion protection in urban rail transit, and in particular relates to a method for diagnosing insulation defects of urban rail transit rails to the ground. Background Art
[0002] As the construction of urban rail transit continues to accelerate, as of February 2024, a total of 55 cities in my country have opened 308 urban rail transit lines, most of which use DC power supply, and the power supply voltage mainly includes DC750V, DC1500V, etc. Most of the DC traction power supply systems for urban rail transit are floating systems. Among them, the rails are used as both running rails for trains and return rails for current. The train draws current from the contact network or the third rail through the pantograph, and the current flows from the rails to the negative pole of the traction power supply station. Because the actual urban rail line is restricted by factors such as the harsh operating environment on site, the equivalent resistance of the rail fasteners does not meet the original design requirements, resulting in the inability to achieve complete insulation between the rails and the ground. This will cause part of the current to leak to the ground through the insulation defects of the rails to the ground when the current flows back to the negative pole of the contact network through the rails. This part of the leaked current is called stray current.
[0003] The stray current leakage of urban rail can have the following effects: 1) Electrochemical corrosion of pipelines: Usually, pipelines have spontaneous natural corrosion, and the corrosion rate is slow. The stray current flows from the cathode area to the anode area in the underground pipeline, and an oxidation reaction occurs in the anode area, causing metal dissolution, thinning of the pipeline wall, and even perforation. When the current value is large, the electrochemical corrosion rate is fast and the perforation time is short. 2) Transformer DC bias: The current flows in the soil near the substation and flows into the transformer through the neutral-grounded transformer, causing the transformer to have DC bias. After the DC bias occurs, the harmonics generated in the transformer excitation current will cause the transformer to be abnormal, with temperature rise, vibration, noise and other problems. While endangering the safety of personnel and equipment, large stray current interference causes additional economic losses.
[0004] There are many factors that affect the size and distribution of stray current leakage in the DC traction power supply system, such as important parameters such as traction return distance, rail body resistance, rail-to-ground transition resistance, etc. Among them, the rail-to-ground transition resistance has a direct impact on the size of stray current leakage. After the subway is built and enters the operation stage, the accumulation of line pollution and fastener corrosion over time will lead to a decrease in the rail-to-ground transition resistance, thereby exacerbating the stray current. When other factors such as the traction return distance and the size of the train current are fixed, as the rail transition resistance continues to increase, the stray current continues to decrease. Therefore, the rail transition resistance is the main factor that directly affects the stray current, and controlling the size of the rail transition resistance is the key to suppressing the stray current leakage problem. Therefore, the monitoring of the rail-to-ground transition resistance is particularly important.
[0005] In the actual rail-to-ground transition resistance detection, the detection length is generally not more than 4 km due to the influence of the upstream and downstream flow lines and the flow lines between the same line. There are problems such as low efficiency of single detection, increased number of detections when the line is long, and a lot of manpower and material resources are consumed. In addition, the transition resistance detection of the operating line can only be carried out in the early morning. In view of the above situation, it is very necessary to study a method for diagnosing rail-to-ground insulation defects. The most direct characterization factor of the rail-to-ground insulation condition should be the distribution of the rail potential of the entire line. When the rail insulation of a certain section of the line is defective, the dynamic distribution of the rail potential of the entire line will change significantly compared with the normal insulation. Therefore, accurate segmental diagnosis of rail-to-ground insulation defects is of great significance to reducing stray current leakage of DC traction power supply system during operation, reducing operating costs, and maintaining the safety of personnel and equipment. When it is diagnosed that the rail insulation to the structure or to the ground is reduced as a whole, the rail insulation to the ground can be improved by patrolling the track, flushing the track, and cleaning the tunnel. For local sections with insulation defects between rails and ground, measures such as improving the track environment of local sections, inspecting and replacing rail fasteners, finding connection problems between trackside equipment and rails, and improving the electrical penetration performance of the return network can be taken to increase the transition resistance between rails and ground, improve the insulation between rails and ground, and thus reduce the leakage of stray current. Summary of the invention
[0006] In view of the above problems, the present invention provides a method for diagnosing insulation defects of urban rail transit rails to the ground.
[0007] A method for diagnosing insulation defects of urban rail transit rails to the ground of the present invention comprises the following steps:
[0008] Step 1: Obtain training sample data.
[0009] A rail-to-ground insulation defect diagnosis training sample database is established, including a line information table, a sample time information table, a sample data information table, and a sample label table.
[0010] Step 2: Build a deep learning model network for rail-to-ground insulation defect diagnosis based on CNN-BiLSTM-Attention fusion.
[0011] The deep learning model network takes the rail potential of the entire line as input and the rail resistance of the entire line to the ground as output; a deep learning hybrid model with a two-dimensional convolutional neural network CNN and a bidirectional long short-term memory neural network BiLSTM as the core is built, and the convolutional neural network CNN is used as the spatial feature extraction module, and the BiLSTM layer is used as the frequency domain feature extraction module; a fast Fourier transform is added to the front end of the network to process the time domain data of the rail potential, and a random enhancement module for the rail potential input features is added during the neural network training.
[0012] Firstly, the training sample data of station rail potential, polarization potential of stray current collection network and surrounding surface potential gradient are normalized. CNN is used to extract the scattered current regularity correlation features between spatial and multi-source information of training samples, and at the same time, the data dimension is compressed to reduce the number of parameters of network training to speed up the learning speed of the diagnostic model. On this basis, the cascaded BiLSTM network is used to extract the time features of the data in parallel in positive and reverse order to establish the corresponding relationship between the features and the categories of rail-to-ground insulation defects, and an attention mechanism is added to make the model pay attention to the synchronization of feature data, the correlation of multi-source information, and the change law of data features when rail-to-ground insulation defects occur.
[0013] Step 3: After cleaning the historical data within a monitoring cycle, the dynamic station rail potentials of multiple trains on the entire line are input into the rail insulation defect diagnosis deep learning model, and the diagnosis results are obtained through the rail insulation defect diagnosis deep learning model.
[0014] Furthermore, in step 1, the line information table contains pre-established line information, including a line serial number, a line name, a description of the corresponding line, and a creation time of the line.
[0015] In the sample time information table, select the date DATE to get the date accurate to year, month and day; then select the dynamic rail potential data of multiple trains in a certain time period of the day from the start time to the end time, and use the Long data type to represent the time of the day as 0-86400 seconds.
[0016] In the sample data information table, the sample time SAMPLETIME is used to indicate the simulation time of the learning sample in the second, and the Char character type is used to store the rail potential PAILPOT of all monitoring points along the entire line and the OVPD state OVPDSTATE of all monitoring points along the entire line.
[0017] The sample label table is the rail insulation defect status distribution R2EC between stations along the entire line starting from the rail potential sample data SAMPLEID.
[0018] Furthermore, the rail potential training data for the entire line takes into account the train operation diagram, line parameters and train parameters in different operating periods, and restores the load process of multiple trains driven by the operation diagram; simulates the working conditions of rail-to-ground insulation defects between stations, and obtains training samples through parallel calculation and simulation of working conditions.
[0019] Furthermore, the training process of the deep learning model network for rail-to-ground insulation defect diagnosis in step 2 is specifically as follows:
[0020] First, the rail potential data is extracted from the deep learning sample database. After the model receives the rail potential data, it is divided into a training set and a test set; the weight and bias matrix of the model are initialized, and then the rail insulation defect diagnosis deep learning model is trained according to the training set part of the rail potential data, and the accuracy and loss value of the network in each round are saved; the offline training of the model includes forward calculation and back propagation, and the back propagation optimization algorithm is used to optimize the network parameters based on the loss value; the offline training will end when the deep learning model training round reaches the preset maximum number of iterations, and the optimal deep learning model for rail insulation defect diagnosis is saved, which means that the model training is completed; finally, the pre-divided test set is used as the network model input to evaluate the accuracy of the deep learning network.
[0021] The beneficial technical effects of the present invention are:
[0022] The rail-to-ground insulation defect diagnosis model adopted by the present invention cascades CNN, BiLSTM and self-attention mechanism. The sparse interaction and parameter sharing properties of the two-dimensional convolutional network CNN are helpful to explore the spatial distribution characteristics of the rail potential. The simultaneous learning strategy of the rail potential of the front and rear monitoring points of the bidirectional long short-term memory network BiLSTM is conducive to exploring the frequency domain distribution characteristics of the rail potential, realizing rapid and accurate diagnosis of rail insulation defects, which is of great significance to reducing the stray current leakage of the DC traction power supply system during the operation period, reducing the operating cost, and maintaining the safety of personnel and equipment. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 This is the network architecture of the rail-to-ground insulation defect diagnosis model of the present invention.
[0024] Figure 2 The invention discloses a method for acquiring training samples of a rail-to-ground insulation defect diagnosis model.
[0025] Figure 3 This is the training process of the rail-to-ground insulation defect diagnosis model of the present invention. DETAILED DESCRIPTION
[0026] The present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.
[0027] A method for diagnosing insulation defects of urban rail transit rails to the ground of the present invention is specifically as follows:
[0028] Step 1: Obtain training sample data.
[0029] A rail-to-ground insulation defect diagnosis training sample database is established, including a line information table, a sample time information table, a sample data information table, and a sample label table.
[0030] The line information table contains pre-established line information, including line serial number, line name, description of the corresponding line, and creation time of the line. By selecting the corresponding urban rail line in the line information table, the rail potential learning sample data in the corresponding line is searched. The line information table of this embodiment is shown in Table 1.
[0031] Table 1 Line information table
[0032]
[0033] In the sample time information table, the date is accurately expressed as year, month, and day by selecting the date DATE; then the dynamic rail potential data of multiple trains in a certain time period of the day are selected from the start time STARTTIME to the end time ENDTIME. At the same time, the Long data type is used to represent the time of a day as 0-86400 seconds. The sample time information table of this embodiment is shown in Table 2.
[0034] Table 2 Sample time information table
[0035]
[0036] In the sample data information table, the sample time SAMPLETIME is used to represent the simulation time of the learning sample in the second, and the Char character type is used to store the rail potential PAILPOT of the monitoring points on the entire line and the OVPD state OVPDSTATE of the monitoring points on the entire line, so that the field data can be extracted and parsed conveniently and quickly later. The sample data information table of this embodiment is shown in Table 3.
[0037] Table 3 Sample data information table
[0038]
[0039]
[0040] The sample label table is the rail insulation defect state distribution R2EC between stations on the entire line starting from the rail potential sample data SAMPLEID. The sample label table of this embodiment is shown in Table 4.
[0041] Table 4 Sample label table
[0042]
[0043] The rail potential training sample simulation is carried out through the Network Power resource sharing and scheduling simulation platform (NPRSS) independently developed by our company. The training samples are quickly obtained using multi-machine cluster parallel computing technology and stored in the database. The rail potential simulation process is as follows: Figure 2 shown.
[0044] Step 2: Build a deep learning model network for rail-to-ground insulation defect diagnosis based on CNN-BiLSTM-Attention fusion.
[0045] The network structure of the deep learning model is as follows Figure 1 As shown in the figure, the deep learning model network takes the rail potential of the entire line as input and the rail resistance of the entire line as output; a deep learning hybrid model with two-dimensional convolutional neural network CNN and bidirectional long short-term memory neural network BiLSTM as the core is built, and the convolutional neural network CNN is used as the spatial feature extraction module, and the BiLSTM layer is used as the frequency domain feature extraction module; Fast Fourier Transform (FFT) is added to the front end of the network to process the time domain data of the rail potential, and a rail potential input feature random enhancement module is added during neural network training.
[0046] Furthermore, the rail potential training data for the entire line takes into account the train operation diagram, line parameters and train parameters in different operating periods, and restores the load process of multiple trains driven by the operation diagram; simulates the working conditions of rail-to-ground insulation defects between stations, and obtains training samples through parallel calculation and simulation of working conditions.
[0047] exist Figure 1 In it, M is the time dimension, N is the space dimension, S is the convolution step, σ is the sigmoid activation function, and δ is the softmax activation function. The output size of the convolution layer is calculated as follows:
[0048] m=(M+2×P-K1) / S+1
[0049] n=(N+2×P-K2) / S+1
[0050] Where P is the padding step, K1 and K2 are the dimensions of the convolution kernel.
[0051] Firstly, the training sample data of station rail potential, polarization potential of stray current collection network and surrounding surface potential gradient are normalized. CNN is used to extract the scattered current regularity correlation features between spatial and multi-source information of training samples, and at the same time, the data dimension is compressed to reduce the number of parameters of network training to speed up the learning speed of the diagnostic model. On this basis, the cascaded BiLSTM network is used to extract the time features of the data in parallel in positive and reverse order to establish the corresponding relationship between the features and the categories of rail-to-ground insulation defects, and an attention mechanism is added to make the model pay attention to the synchronization of feature data, the correlation of multi-source information, and the change law of data features when rail-to-ground insulation defects occur.
[0052] The training process of the deep learning model network for rail-to-ground insulation defect diagnosis is as follows: Figure 3 As shown, specifically:
[0053] First, the rail potential data is extracted from the deep learning sample database. After the model receives the rail potential data, it is divided into a training set and a test set; the weight and bias matrix of the model are initialized, and then the rail insulation defect diagnosis deep learning model is trained according to the training set part of the rail potential data, and the accuracy and loss value of the network in each round are saved; the offline training of the model includes forward calculation and back propagation, and the back propagation optimization algorithm is used to optimize the network parameters based on the loss value; the offline training will end when the deep learning model training round reaches the preset maximum number of iterations, and the optimal deep learning model for rail insulation defect diagnosis is saved, which means that the model training is completed; finally, the pre-divided test set is used as the network model input to evaluate the accuracy of the deep learning network.
[0054] Step 3: After cleaning the historical data within a monitoring cycle, the dynamic station rail potentials of multiple trains on the entire line are input into the rail insulation defect diagnosis deep learning model, and the diagnosis results are obtained through the rail insulation defect diagnosis deep learning model.
Claims
1. A method for diagnosing rail-to-ground insulation defects in urban rail transit, characterized in that: The following steps are involved: Step 1: Obtain training sample data; Establish a rail-to-ground insulation defect diagnosis training sample database, including a line information table, a sample time information table, a sample data information table, and a sample label table; Step 2: Build a deep learning model network for rail-to-ground insulation defect diagnosis based on CNN-BiLSTM-Attention fusion; The deep learning model network uses the rail potential of the entire line as input and the rail resistance of the entire line to the ground as output; a deep learning hybrid model with a two-dimensional convolutional neural network (CNN) and a bidirectional long short-term memory (BiLSTM) neural network as the core is built, using the convolutional neural network (CNN) as the spatial feature extraction module and the BiLSTM layer as the frequency domain feature extraction module; Fast Fourier transform is added to the front end of the network to process the time domain data of rail potential, and a random enhancement module of rail potential input features is added during neural network training; Firstly, the training sample data of the station rail potential, the polarization potential of the stray current collection network and the surrounding surface potential gradient are normalized; CNN is used to extract the scattered current regularity correlation features between the space and multi-source information of the training samples, and at the same time, the data dimension is compressed to reduce the number of parameters of network training to speed up the learning speed of the diagnostic model; on this basis, the cascaded BiLSTM network is used to extract the time features of the data in parallel in positive and reverse order to establish the corresponding relationship between the features and the categories of rail-to-ground insulation defects, and an attention mechanism is added to make the model pay attention to the synchronization of feature data, the correlation of multi-source information, and the law of data feature changes when rail-to-ground insulation defects occur; Step 3: After cleaning the historical data within a monitoring cycle, the dynamic station rail potentials of multiple trains on the entire line are input into the rail insulation defect diagnosis deep learning model, and the diagnosis results are obtained through the rail insulation defect diagnosis deep learning model.
2. The method for diagnosing rail-to-ground insulation defects in urban rail transit according to claim 1, characterized in that: In step 1, the line information table contains pre-established line information, including a line serial number, a line name, a description of the corresponding line, and a creation time of the line; In the sample time information table, select the date DATE to get the date accurate to year, month and day; then select the dynamic rail potential data of multiple trains in a certain time period of the day from the start time to the end time, and use the Long data type to represent the time of the day as 0-86400 seconds; In the sample data information table, the sample time SAMPLETIME is used to indicate the simulation time of the learning sample in the second, and the Char character type is used to store the rail potential PAILPOT of all monitoring points and the OVPD state OVPDSTATE of all monitoring points; The sample label table is the rail insulation defect status distribution R2EC between stations along the entire line starting from the rail potential sample data SAMPLEID.
3. The method for diagnosing rail-to-ground insulation defects in urban rail transit according to claim 1, characterized in that: The rail potential training data for the entire line takes into account the train operation diagrams, line parameters and train parameters in different operating periods, and is driven by the operation diagram to restore the load process of multiple trains; the rail-to-ground insulation defect working conditions between stations are simulated, and training samples are obtained through parallel calculation and simulation of the working conditions.
4. The method for diagnosing rail-to-ground insulation defects in urban rail transit according to claim 1, characterized in that: The training process of the deep learning model network for rail-to-ground insulation defect diagnosis in step 2 is specifically as follows: First, the rail potential data is extracted from the deep learning sample database. After the model receives the rail potential data, it is divided into a training set and a test set. The weight and bias matrix of the model are initialized. Then, the rail insulation defect diagnosis deep learning model is trained based on the training set part of the rail potential data, and the accuracy and loss value of the network in each round are saved. The offline training of the model includes forward calculation and back propagation, and the back propagation optimization algorithm is used to optimize the network parameters based on the loss value. The offline training will end when the deep learning model training round reaches the preset maximum number of iterations, and the optimal deep learning model for rail insulation defect diagnosis is saved, which means that the model training is completed. Finally, the pre-divided test set is used as the network model input to evaluate the accuracy of the deep learning network.