Wheel-rail state recognition method based on feature fusion variable dimension convolutional neural network

Abstract

The application belongs to the technical field of high-speed railway wheel-rail state monitoring, and particularly discloses a wheel-rail state recognition method based on a feature fusion variable-dimension convolutional neural network. A shaft box vertical acceleration is selected as original vibration for state recognition, a vehicle-track coupled dynamic system is modeled, shaft box vertical vibration acceleration time domain responses generated under different speeds and different wheel-rail excitations are simulated, and the frequency spectrum of the time domain signal is obtained after Fourier transform, both of which are simultaneously sent to a feature fusion variable-dimension convolutional neural network for state recognition. The wheel-rail state of the train is quickly and accurately recognized, and the problem that wheel-rail state recognition is difficult under complex working conditions by using time-frequency analysis, principal component analysis methods and existing vehicle-mounted wheel-rail state monitoring systems is solved.

Description

Technical Field

[0001] This invention belongs to the field of high-speed rail wheel-rail condition monitoring technology, specifically involving the design of a wheel-rail condition recognition method based on feature fusion variable-dimensional convolutional neural network. Background Technology

[0002] In recent years, with the increase in railway operating mileage and train speed, wheel-rail interaction has intensified. High-speed trains operate under complex conditions, and wheel-rail defects occur frequently. Polygonal wheel wear and wavy rail wear, as common wheel-rail defects, significantly impact train safety, making wheel-rail condition monitoring essential.

[0003] Existing wheel-rail health monitoring systems are mainly divided into two categories: trackside monitoring systems and on-board monitoring systems. Trackside monitoring systems primarily rely on monitoring equipment installed beside the track to monitor the wheel and rail condition. Compared to trackside monitoring systems, on-board monitoring systems offer continuous monitoring. For subway vehicles, existing wheel-rail health monitoring systems mainly rely on methods such as image analysis to process the acquired vibration signals and determine the failure status of the wheel tread and rails. However, due to the complexity of the wheel and rail condition and the vertical vibration response components of the axle boxes in high-speed rail, relying solely on the above methods to identify the wheel and rail condition has certain limitations. Deep learning methods can effectively solve these problems.

[0004] While existing neural network models can accurately determine fault states, they typically only consider the difference between wheel or rail defects and healthy conditions, and generally require principal component analysis of the signal. This results in large amounts of data and long data processing times, making it difficult to meet the basic requirements for on-board wheel-rail health monitoring. Time-frequency analysis is usually more accurate in diagnosing single prominent components, but this method is greatly affected by time-frequency resolution and time-frequency plot size. Summary of the Invention

[0005] The purpose of this invention is to address the numerous limitations of existing wheel-rail health monitoring systems in the process of wheel-rail condition identification, and to propose a wheel-rail condition identification method based on feature fusion variable-dimensional convolutional neural networks.

[0006] The technical solution of this invention is: a wheel-track state recognition method based on feature fusion variable-dimensional convolutional neural network, comprising the following steps: S1. Establish a vehicle-track coupled dynamic system model and obtain the time-domain response of the vertical vibration acceleration of the axle box through simulation; S2. Analyze and preprocess the obtained time-domain response of the vertical vibration acceleration of the axle box to construct a dataset of wheel-rail defective states; S3. Establish a diagnostic model using a variable-dimensional convolutional neural network that integrates time and frequency features; S4. Input the wheel-rail defect data set into the neural network and train the neural network; S5. Use the trained variable-dimensional convolutional neural network diagnostic model to identify the wheel-rail status and output the wheel-rail status identification results.

[0007] The beneficial effects of this invention are: 1. It enables rapid and accurate identification of train wheel-rail conditions, solving the problems of time-frequency analysis, principal component analysis methods, and existing on-board wheel-rail condition monitoring systems in discerning wheel-rail conditions under complex working conditions.

[0008] 2. The constructed time-frequency feature fusion variable-dimensional convolutional neural network diagnostic model has low complexity, simple structure, few parameters, is less affected by human factors, and is not limited by expert knowledge. Its accuracy in identifying wheel track defects is higher than that of traditional machine learning methods and deep learning structures, and it has a certain generalization ability.

[0009] Preferably, the vehicle-track coupled dynamics model in step S1 includes a vehicle subsystem and a track subsystem; the vehicle subsystem includes one car body and two bogies, with the two bogies located below both ends of the car body; each bogie includes a frame, two wheelsets, and four axle boxes; the frame is connected to the two wheelsets below via a primary suspension and to the car body above via air springs; the primary suspension includes a primary steel spring and a primary vertical damper; the four axle boxes are fixed to both sides of the two wheelsets; the track subsystem includes rails and track slabs.

[0010] The beneficial effects of the above preferred solution are: By establishing an accurate vehicle-track coupled dynamic model, the time-domain response of axle box vibration acceleration is obtained through simulation. Based on this, a variable-dimensional convolutional neural network diagnostic model that integrates time and frequency features is constructed, which has a higher accuracy rate in identifying wheel-rail defective conditions than traditional machine learning methods and deep learning structures.

[0011] Preferably, step S2 specifically includes the following steps: S21. Take 0.5 to 1.0 mm portions respectively. s The time-domain response of the vertical vibration acceleration of the axle box is taken as the time-domain signal of each sample; S22. Perform a Fourier transform on the time-domain signal to obtain its corresponding spectrum; S23. Based on the obtained spectrum, select the first 1200. Hz As a frequency domain signal; S24. Perform min-max normalization on the time-domain and frequency-domain signals to obtain the processed time-domain and frequency-domain signals, and construct a wheel-rail defect state dataset.

[0012] The beneficial effects of the above preferred solution are: By analyzing and preprocessing the time-domain response of axial and vertical vibration acceleration, a dataset of wheel-rail defective conditions is formed. A variable-dimensional convolutional neural network diagnostic model is constructed that can accurately capture the complex relationship between the vertical vibration acceleration of the axle box and the wheel-rail condition, thereby achieving accurate identification of wheel-rail condition.

[0013] Preferably, the minimum-maximum normalization formula in step S24 is as follows:

[0014] in, This represents the data after min-max normalization; Indicates the actual value; Represents the maximum value in the sequence; This represents the minimum value in the sequence.

[0015] The beneficial effects of the above preferred solution are: The min-max normalization process accelerates the convergence of the model, reduces the impact of sequence extrema on the model, and eliminates the limitation of different dimensions between samples.

[0016] Preferably, the time-frequency feature fusion variable-dimensional convolutional neural network diagnostic model described in step S3 includes 1 DCNN and 2 DCNN , and 1 DCNN and 2 DCNN Connected via a fully connected layer; the 1 DCNN It includes 3 convolutional layers, 3 pooling layers, and 1 batch normalization layer; the 2 DCNN It consists of only 4 convolutional layers and 4 pooling layers.

[0017] The beneficial effects of the above preferred solution are: The variable-dimensional convolutional neural network diagnostic model that integrates time and frequency features enables the monitoring and accurate identification of wheel-rail status without the need for complex signal processing and principal component analysis.

[0018] Preferably, step S4 specifically includes the following steps: S41. Divide the wheel-rail defective state dataset into a training set and a test set; S42. Input the training set into the variable-dimensional convolutional neural network diagnostic model and train it for wheel-track state recognition; S43. Input the test set into the variable-dimensional convolutional neural network diagnostic model after each training session, test the variable-dimensional convolutional neural network diagnostic model after each training session, and determine whether the training number has been reached. If yes, the training is complete; otherwise, return to step S42.

[0019] The beneficial effects of the above preferred solution are: By training a neural network, the complex relationship between the vertical vibration acceleration of the axle box and the wheel-rail condition can be accurately captured, thereby achieving precise identification of the wheel-rail condition.

[0020] Preferably, step S5 specifically includes the following steps: Step S5 specifically includes the following steps: S51. Input the wheel-rail time-domain signal and frequency-domain signal into 1 DCNN In the middle, and output the feature points; S52. Connect the feature points end to end to form a feature vector, and use a fully connected neural network for dimensionality reduction; S53. Increase the dimensionality of the reduced feature vectors and rearrange them to form a time-frequency feature map; S54. Input the time-frequency feature map into 2 DCNN Perform wheel-rail status identification and output the wheel-rail status identification results.

[0021] The beneficial effects of the above preferred solution are: Based on the rapid dimensionality reduction of vibration time series and corresponding spectral series by one-dimensional convolutional neural network, time and frequency features are fused and rearranged in higher dimensions to form a time-frequency feature map. A two-dimensional convolutional neural network is then used to extract features and determine wheel-rail status from the small-sized time-frequency feature map. This solves the problem of difficulty in determining wheel-rail status under complex working conditions in time-frequency analysis, principal component analysis methods, and existing vehicle-mounted wheel-rail status monitoring systems.

[0022] Preferably, the wheel-rail condition described in step S54 includes three conditions: wheel-rail health condition, wheel polygonal shape, and rail corrugation.

[0023] Preferably, the formula for describing the wheel polygon is:

[0024]

[0025] in, This represents the passing frequency corresponding to each order of polygon; Indicates the wheel is Temporal roughness; Indicates the train's speed; Indicates the order of the polygon; Indicates the wheel radius; Indicates the first The magnitude of the polygon; Indicates the first Two-dimensional polygons; Indicates the first Phase of a polygon.

[0026] Preferably, the formula describing the rail corrugation is:

[0027] in, This represents the frequency of the wave-damping response at a single wavelength. Indicates the wavelength of the wave mill.

[0028] The beneficial effects of the above preferred solution are: Simultaneously considering the combination of poor wheel and rail condition with healthy wheel and rail condition, accurate identification of wheel and rail condition is achieved. Attached Figure Description

[0029] Figure 1 The diagram shows a flowchart of a wheel-track state recognition method based on a feature fusion variable-dimensional convolutional neural network. Detailed Implementation

[0030] Exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the drawings are merely exemplary and are intended to illustrate the principles and spirit of the invention, and are not intended to limit the scope of the invention.

[0031] like Figure 1 As shown, a wheel-track state recognition method based on feature fusion variable-dimensional convolutional neural networks includes the following steps: S1. Establish a vehicle-track coupled dynamic system model and obtain the time-domain response of the vertical vibration acceleration of the axle box through simulation; S2. Analyze and preprocess the obtained time-domain response of the vertical vibration acceleration of the axle box to construct a dataset of wheel-rail defective states; S3. Establish a diagnostic model using a variable-dimensional convolutional neural network that integrates time and frequency features; S4. Input the wheel-rail defect data set into the neural network and train the neural network; S5. Use the trained variable-dimensional convolutional neural network diagnostic model to identify the wheel-rail status and output the wheel-rail status identification results.

[0032] Considering the simulation conditions, the track is set as a straight segment. The simulated track conditions under both polygonal wheel and healthy wheel excitation are the same as the measured long-wave irregularities of the rail, identical to the vertical irregularities of the rail before the corrugated composite sine wave. The vehicle runs at a constant speed, with a simulated speed of 100–450 km / h. km / h The speed interval is 25 km / h The simulation time was 1.2 seconds. s The sampling frequency is 10000. Hz Simulations were performed to obtain the time-domain response of the vertical vibration acceleration of the axle box under different working conditions.

[0033] In this embodiment, the vehicle-track coupled dynamics model in step S1 includes a vehicle subsystem and a track subsystem. The vehicle subsystem includes one car body and two bogies, with the two bogies located below both ends of the car body. Each bogie includes a frame, two wheelsets, and four axle boxes. The frame is connected to the two wheelsets below via a primary suspension and to the car body above via air springs. The primary suspension includes primary steel springs and primary vertical dampers. The four axle boxes are fixed to both sides of the two wheelsets. The track subsystem includes rails and track slabs. The rails are considered as... Timoshenko Beams and track slabs are adopted CRTS Ⅲ The track slab structure. Accurate description of wheelset flexibility is crucial to the reliability of the axle box vibration response. This invention utilizes finite element software to analyze the modes of the flexible wheelset and its first 600... Hz Wheelset mode import Simpack A rigid-flexible coupled vehicle model was established to calculate the vertical vibration acceleration of the axle box, and the wheel-rail normal force calculation of this model was based on... Hertz Contact theory, tangential force calculation based on FASTSIM Fast algorithm. This invention also considers the influence of track slab flexibility on the vertical vibration acceleration of the axle box. CRTS Ⅲ The first 30 modes of the track slab, including the first 6 rigid body modes and 24 flexible modes, make the calculated vibration acceleration of the axle box closer to the actual response.

[0034] In this embodiment, step S2 specifically includes the following steps: S21. Take 0.5 to 1.0 mm portions respectively. s The axial vertical vibration acceleration time-domain response is used as the time-domain signal for each sample; S22. Perform a Fourier transform on the time-domain signal to obtain its corresponding spectrum; S23. Based on the obtained spectrum, select the first 1200. Hz As a frequency domain signal; S24. Perform min-max normalization on the time-domain and frequency-domain signals, mapping the time-domain signal and spectrum sequence of each sample to between 0 and 1, to obtain the processed time-domain and frequency-domain signals; S25. Construct a wheel-rail defect state dataset based on the processed time-domain and frequency-domain signals.

[0035] In this embodiment, the minimum-maximum normalization formula in step S24 is as follows:

[0036] in, This represents the data after min-max normalization; Indicates the actual value; Represents the maximum value in the sequence; This represents the minimum value in the sequence.

[0037] In this embodiment, the time-frequency feature fusion variable-dimensional convolutional neural network diagnostic model described in step S3 includes a one-dimensional convolutional neural network (1 DCNN ) and two-dimensional convolutional neural networks (2 DCNN ), and 1 DCNN and 2 DCNN Connected via a fully connected layer; the 1 DCNN It includes 3 convolutional layers, 3 pooling layers, and 1 batch normalization layer; the 2 DCNN It consists of only 4 convolutional layers and 4 pooling layers.

[0038] Parameter selection has a significant impact on convolutional neural networks (CNNs). Appropriate parameters can fully utilize the feature extraction capabilities of CNNs. The choice of parameters such as batch size, learning rate, number of training iterations, loss function, and optimizer greatly affects the convergence speed and accuracy of CNNs. In this embodiment, the batch size is set to 64, the learning rate to 0.00001, and the number of training iterations to 1500. To suit the classification task, the cross-entropy loss function is chosen. The optimizer is selected... Adam To alleviate network overfitting, the network model is modified. L 2. Regularization. To improve the nonlinearity of the network, regularization is adopted. ReLU Activation function.

[0039] In this embodiment, step S4 specifically includes the following steps: S41. Divide the wheel-rail defect dataset into a training set and a test set, with 450 samples for each wheel-rail defect class, totaling 1350 samples across the three sequence classes. Each sample contains 5000 points in the time domain and 6000 points in the frequency domain. Randomly select 80% of the samples as the training set and the remaining 20% ​​as the test set, resulting in 1080 samples in the training set and 270 samples in the test set. S42. Input the training set into the variable-dimensional convolutional neural network diagnostic model and train it for wheel-track state recognition; S43. Input the test set into the variable-dimensional convolutional neural network diagnostic model after each training session, test the variable-dimensional convolutional neural network diagnostic model after each training session, and determine whether the training number has been reached. If yes, the training is complete; otherwise, return to step S42.

[0040] In this embodiment, step S5 specifically includes the following steps: S51. Input the wheel-rail time-domain signal and frequency-domain signal into 1 DCNN In, each 1 DCNN The output, after being flattened, contains 256 feature points; S52. Connect the feature points end to end to form a 512-dimensional feature vector, and use a fully connected neural network to reduce the dimensionality; S53. Increase the dimensionality of the reduced feature vectors and rearrange them to form a 16×16 time-frequency feature map; S54. Input the time-frequency feature map into 2 DCNN Perform wheel-rail status identification and output the wheel-rail status identification results.

[0041] In this embodiment, the wheel-rail state described in step S54 includes three states: wheel-rail health state, wheel polygonal state, and rail corrugation state.

[0042] In this embodiment, the descriptive formula for the wheel polygon is:

[0043]

[0044] in, This represents the passing frequency corresponding to each order of polygon; Indicates the wheel is Temporal roughness; Indicates the train's speed; Indicates the order of the polygon; Indicates the wheel radius; Indicates the first The magnitude of the polygon; Indicates the first Two-dimensional polygons; Indicates the first Phase of a polygon.

[0045] In this embodiment, the descriptive formula for rail corrugation is:

[0046] in, This represents the frequency of the wave-damping response at a single wavelength. Indicates the wavelength of the wave mill.

[0047] The specific working principle and process of this invention are as follows: According to the present invention, a variable-dimensional convolutional neural network that fuses time and frequency features is used for wheel-track condition recognition. The model converges after about 300 iterations, and the accuracy of the test set after convergence is about 95%, with the highest accuracy reaching 95.93%. The loss curve has small oscillations and is relatively stable after convergence, which can achieve a good wheel-track defect condition recognition effect.

[0048] Measured axle box vibration acceleration signals and their spectra under different operating conditions were used as inputs to verify the wheel-rail condition identification capability. The measured vertical vibration acceleration signals of the axle boxes were obtained from type tests of a certain type of high-speed train, with a wheel diameter of 0.92 mm. m Under the polygonal wheel condition, the track condition is good; under the rail corrugation condition, the wheel condition is good; and under the healthy condition, both wheel and rail conditions are good. 360 signal segments were randomly extracted for each condition, and their spectra were plotted. The corresponding sequences for each condition were input into the model to identify wheel-rail defects. As shown in Table 1, it can be seen that the present invention can achieve effective identification with high accuracy, and the response time meets the requirements of real-time online monitoring. This demonstrates that the model has a certain generalization ability and the ability to diagnose wheel-rail defects under complex conditions, possessing theoretical significance and engineering value for online monitoring of wheel-rail conditions.

[0049] Table 1. Verification of measured data

[0050] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A wheel-track state recognition method based on feature fusion variable-dimensional convolutional neural network, characterized in that, Includes the following steps: S1. Establish a vehicle-track coupled dynamic system model and obtain the time-domain response of the vertical vibration acceleration of the axle box through simulation; S2. The obtained time-domain response of the vertical vibration acceleration of the axle box is analyzed and preprocessed to construct a wheel-rail defect state dataset, specifically: S21. Take 0.5 to 1.0 mm portions respectively. s The time-domain response of the vertical vibration acceleration of the axle box is taken as the time-domain signal of each sample; S22. Perform a Fourier transform on the time-domain signal to obtain its corresponding spectrum; S23. Based on the obtained spectrum, select the first 1200. Hz As a frequency domain signal; S24. Perform min-max normalization on the time-domain and frequency-domain signals to obtain the processed time-domain and frequency-domain signals, and construct a wheel-rail defect state dataset; S3. Establish a diagnostic model for a variable-dimensional convolutional neural network that integrates time-frequency features; the diagnostic model for a variable-dimensional convolutional neural network that integrates time-frequency features includes 1 DCNN and 2 DCNN , and 1 DCNN and 2 DCNN Connected via a fully connected layer; the 1 DCNN It includes 3 convolutional layers, 3 pooling layers, and 1 batch normalization layer; the 2 DCNN It consists of only 4 convolutional layers and 4 pooling layers; S4. Input the wheel-rail defect data set into the neural network and train the neural network; S5. Use the trained variable-dimensional convolutional neural network diagnostic model to identify the wheel-rail status and output the wheel-rail status identification results; Step S5 specifically includes the following steps: S51. Input the wheel-rail time-domain signal and frequency-domain signal into 1 DCNN In the middle, and output the feature points; S52. Connect the feature points end to end to form a feature vector, and use a fully connected neural network for dimensionality reduction; S53. Increase the dimensionality of the reduced feature vectors and rearrange them to form a time-frequency feature map; S54. Input the time-frequency feature map into 2 DCNN Perform wheel-rail status identification and output the wheel-rail status identification results.

2. The wheel-track state recognition method based on feature fusion variable-dimensional convolutional neural network according to claim 1, characterized in that: The vehicle-track coupled dynamics model described in step S1 includes a vehicle subsystem and a track subsystem. The vehicle subsystem includes one car body and two bogies, with the two bogies located below both ends of the car body. Each bogie includes a frame, two wheelsets, and four axle boxes. The frame is connected to the two wheelsets below it via a primary suspension and to the car body above it via air springs. The primary suspension includes a primary steel spring and a primary vertical damper. The four axle boxes are fixed to both sides of the two wheelsets. The track subsystem includes rails and track slabs.

3. The wheel-track state recognition method based on feature fusion variable-dimensional convolutional neural network according to claim 1, characterized in that, The minimum-maximum normalization formula in step S24 is as follows: in, This represents the data after min-max normalization; Indicates the actual value; Represents the maximum value in the sequence; This represents the minimum value in the sequence.

4. The wheel-track state recognition method based on feature fusion variable-dimensional convolutional neural network according to claim 1, characterized in that, Step S4 specifically includes the following steps: S41. Divide the wheel-rail defective state dataset into a training set and a test set; S42. Input the training set into the variable-dimensional convolutional neural network diagnostic model and train it for wheel-track state recognition; S43. Input the test set into the variable-dimensional convolutional neural network diagnostic model after each training session, test the variable-dimensional convolutional neural network diagnostic model after each training session, and determine whether the training number has been reached. If yes, the training is complete; otherwise, return to step S42.

5. The wheel-track state recognition method based on feature fusion variable-dimensional convolutional neural network according to claim 1, characterized in that: The wheel-rail condition described in step S54 includes three states: wheel-rail health, wheel polygonal shape, and rail corrugation.

6. The wheel-track state recognition method based on feature fusion variable-dimensional convolutional neural network according to claim 5, characterized in that, The formula describing the polygon of the wheel is: in, This represents the passing frequency corresponding to each order of polygon; Indicates the wheel is Temporal roughness; Indicates the train's speed; Indicates the order of the polygon; Indicates the wheel radius; Indicates the first The magnitude of the polygon; Indicates the first Two-dimensional polygons; Indicates the first Phase of a polygon.

7. The wheel-track state recognition method based on feature fusion variable-dimensional convolutional neural network according to claim 5, characterized in that, The formula describing the rail corrugation is as follows: in, This represents the frequency of the single-wavelength wave milling response; Indicates the wavelength of the wave mill.

Images