Train bogie snaking state identification method based on wheel set motion state characteristics

Abstract

The invention discloses a train bogie snaking state identification method based on wheel set motion state characteristics, which comprises the following steps of: S1, acquiring acceleration signals of a train framework, and selecting transverse acceleration signals and longitudinal acceleration signals of the framework as signals for representing the wheel set motion state; s2, performing filtering and time-frequency transformation processing on the signal representing the motion state of the wheel set to obtain snaking motion characteristics; and S3, inputting the snaking motion features into a snaking monitoring network, and outputting a snaking state classification result. According to the method, a framework acceleration signal capable of indirectly reflecting the motion state of the wheel set is selected as a feature extraction target, and meanwhile, a CNN feature fusion network is adopted as a snaking monitoring network to extract feature information; according to the method, snaking states of different degrees can be identified by combining time domain and time frequency domain information characteristics of transverse and longitudinal acceleration of a framework, and snaking motion state identification with higher fine granularity is realized.

Description

Technical Field

[0001] The invention belongs to the technical field of high-speed train running status monitoring, and in particular relates to a train bogie snaking state identification method based on wheelset motion state characteristics. Background Art

[0002] In recent years, the rapid development of railway transportation has significantly promoted the world's economic prosperity and social progress, and has become an important pillar of the comprehensive transportation system at home and abroad. With the widespread operation of high-speed EMUs across the country, ensuring rail transportation safety to meet the growing demand for train transportation has become a research focus for railway-related workers. Snake motion stability is the key and basic requirement to ensure the stable and safe operation of high-speed trains. Snake motion of trains during high-speed driving will deteriorate the stability of vehicle operation and aggravate wheel-rail wear. Severe snake instability may even lead to derailment accidents. In order to ensure the safety and comfort of train operation, it is necessary to monitor the snake motion of high-speed trains online.

[0003] At present, the methods used to evaluate the snaking stability of vehicles on the line are mainly based on wheel-rail force and frame lateral acceleration. The method based on frame lateral acceleration can effectively reflect the snaking motion characteristics of the bogie and is easy to operate and low in cost, so it is often used for online monitoring of vehicle snaking. In China, the GB / T 5599-2019 standard is used to judge the snaking stability. After the vibration acceleration of the frame above the axle box is band-pass filtered at 0.5Hz to 10Hz, its peak value reaches or exceeds 8m / s for 6 or more consecutive times. 2 In actual operation, the wear of wheel-rail profiles, line conditions, and changes in vehicle parameters make the snaking stability of high-speed trains a complex system problem. If the snaking stability is judged by a single theoretical formula or a fixed judgment standard, it will inevitably have flaws.

[0004] Therefore, relevant workers have begun to further explore ways to improve the accuracy of snaking monitoring. Wang et al. proposed a snaking coefficient (HC) indicator related to the amplitude and frequency of the wheelset lateral displacement, and established a mapping relationship between the snaking state and the HC index. This method provides a means for monitoring the snaking state of trucks. Zeng et al. proposed a method based on the maximum Lyapunov exponent, using the periodic characteristics of the snaking signal to identify the snaking motion of the vehicle. Kulkarni et al. proposed an unsupervised anomaly detection and clustering algorithm based on the IVRIDA framework to detect and identify snaking motion and the corresponding causes of snaking. Ning et al. studied the evolution process of small snaking and proposed an improved prediction method for the amplitude interval of small snaking evolution. The performance on high-speed train data is better than that of the existing interval prediction method. In response to the problem of scarce measured snaking data, Wang et al. proposed a high-speed train snaking motion recognition method based on transfer learning. This method combines simulation data and measured data and has certain engineering application value. Ning et al. proposed a feature extraction method based on multiscale permutation entropy (MPE) and local tangent space alignment (LTSA), which can effectively extract the relevant features of vehicle small snaking. Guo et al. proposed a small snaking identification method with autocorrelation coefficient and spectral frequency expansion as key features, and sensitivity analysis showed that it has good results. Liu Xin et al. proposed a high-speed train small snaking identification method based on multi-source difference confrontation, which achieved good results in the task of identifying the snaking state of high-speed trains under variable working conditions. Since the snaking motion stability of high-speed trains is a complex system problem, using only a single signal data is not enough for effective fault tracking and diagnosis. Since the snaking motion stability of high-speed trains is a complex system problem, using only a single signal data is not enough for effective fault tracking and diagnosis. Therefore, many researchers have carried out relevant research on information fusion technology. Kulkarni et al. used the consistency between the lateral and longitudinal axle box accelerations of the snaking motion to introduce an indicator to detect the coupling mode of the lateral and yaw directions of the vehicle, and proposed a snaking monitoring program based on the axle box acceleration signal. Sun et al. collected multiple vibration signals from different parts and directions of the train frame and car body, and used cross-correlation technology to determine the index with the best correlation with the snaking stability. This method can effectively identify the small snaking state. Ye et al. combined the lateral-longitudinal-vertical acceleration signal data fusion of the bogie frame with the feature extraction method of independent mode function reconstruction-linear local tangent space alignment (IMFR-LLTSA) to identify the snaking motion of high-speed trains. The verification results of real data showed that this method is better than the snaking identification method that only uses the lateral acceleration signal.

[0005] The above studies have made significant progress in defining the stability of snaking motion and identifying the snaking state, effectively improving the accuracy of snaking monitoring. However, these studies are still limited in evaluating the impact of snaking motion on the actual operation of the vehicle. In order to more comprehensively and accurately reflect the running safety status of the train, the wheel-rail contact state has gradually become a new perspective for evaluating the stability of snaking motion. Sun et al. constructed a vehicle snaking monitoring system based on the singular value decomposition (SVD) method of the Hankel matrix that can predict the lateral displacement of the wheelset and the wheel-rail contact angle, and verified a certain accuracy and effectiveness in the experiment, but the local prediction deviation is large when the track is uneven. Ye et al. proposed a virtual point tracking (VPT) method based on computer vision, which monitors the snaking motion of the vehicle by processing the dynamic wheel-rail contact video. Compared with the snaking monitoring method based on the acceleration of the bogie frame, it can more directly and reliably reflect the safety status of the vehicle, but the monitoring accuracy may be reduced due to environmental influences such as light. In summary, evaluating the snaking motion stability by wheel-rail contact status has become a new research trend. However, under the complex operating conditions of high-speed trains, how to improve the robustness and adaptability of the monitoring method is still an important direction that needs to be explored in depth in future research. Summary of the invention

[0006] The wheel-rail contact state can directly reflect the degree of danger of snaking motion, but it is difficult to accurately measure the wheel-rail contact state during operation. The wheel-rail clearance margin and the wheelset lateral displacement reflect the dynamic position and state of the wheelset on the track, and are important parameters for describing the wheelset motion state. They are closely related to the wheel-rail contact state. In view of the deficiencies in the prior art, the train bogie snaking state identification method based on the wheelset motion state characteristics provided by the present invention uses the wheelset lateral displacement and the wheel-rail clearance margin as the key basis for snaking motion state identification, which solves the problem that the existing snaking identification method based on the solidified judgment standard is difficult to reflect the safety state of the vehicle.

[0007] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a train bogie snaking state identification method based on the wheelset motion state characteristics, comprising the following steps:

[0008] S1, collecting the train frame acceleration signal, and selecting the frame lateral acceleration signal and longitudinal acceleration signal as the signal representing the motion state of the wheelset;

[0009] S2, filtering and time-frequency transforming the signal representing the motion state of the wheelset to obtain the snaking motion characteristics;

[0010] The snaking motion characteristics include the time-frequency domain characteristics of the lateral movement and the lateral shaking coupling characteristics;

[0011] S3, inputting the snaking motion feature into a snaking monitoring network, and outputting a snaking state classification result;

[0012] The snaking monitoring network includes a one-dimensional convolutional neural network module, a coordinate convolutional neural network module, a feature fusion module, a three-dimensional convolutional neural network module and a fully connected module;

[0013] The one-dimensional convolutional neural network module is used to extract the timing features of the input serpentine motion features in the time domain, and the coordinate convolutional neural network module is used to extract the position features of the input serpentine motion features in the frequency domain by introducing the coordinate channel; the feature fusion module is used to splice and fuse the extracted timing features and position features to obtain a coupled feature vector; the three-dimensional convolutional neural network module is used to extract the spatiotemporal information of multiple coupled feature vectors; the fully connected module is used to classify and identify the spatiotemporal information and output the serpentine state recognition result.

[0014] Furthermore, the step S2 includes the following sub-steps:

[0015] S21, performing FFT filtering on the frame lateral acceleration signal to obtain a lateral acceleration bandpass filtered signal;

[0016] Perform FFT filtering on the longitudinal acceleration signal of the frame to obtain a longitudinal acceleration bandpass filter signal, and extract the amplitude and frequency information of the longitudinal acceleration signal;

[0017] S22, performing synchronous compression wavelet transform processing on the lateral acceleration bandpass filter signal to obtain a synchronous compression time-frequency diagram;

[0018] S23, extracting the time-frequency information of the lateral acceleration signal in the synchronous compression time-frequency diagram as a lateral motion time-frequency domain signal;

[0019] S24, extracting the amplitude and frequency information from the lateral acceleration bandpass filter signal, and enhancing it using the extracted time-frequency information to obtain the amplitude and frequency information of the lateral acceleration signal;

[0020] S25, fusing the longitudinal acceleration signal amplitude and frequency information with the lateral acceleration signal amplitude and frequency information to obtain the lateral movement and shaking head coupling characteristics.

[0021] Furthermore, the step S22 includes the following sub-steps:

[0022] S22-1, performing continuous wavelet transform processing on the lateral acceleration bandpass filter signal;

[0023] S22-2, determining the instantaneous frequency of the signal according to the signal processed by continuous wavelet transform;

[0024] S22-3. According to the instantaneous frequency of the signal and the signal after continuous wavelet transformation, the wavelet coefficients are redistributed to obtain a signal with a more concentrated time-frequency representation after synchronous compression, and then a synchronous compression wavelet transform time-frequency diagram is obtained.

[0025] Furthermore, in step S22-3, the time-frequency representation of the signal SWT after synchronization compression is more concentrated y (ω,b) is expressed as:

[0026]

[0027] Where W y (a, b) represents the signal after continuous wavelet transform, ω y (a,b) represents the instantaneous frequency of the signal, a and b represent the scale and translation parameters respectively, ω represents the frequency, and δ represents the Dirac function.

[0028] Furthermore, in step S3, the one-dimensional convolutional neural network module includes a first branch and a second branch with the same structure, which are respectively used to extract the time-frequency domain characteristics of the input lateral motion and the timing characteristics corresponding to the lateral movement and head shaking coupling characteristics.

[0029] Furthermore, in step S3, the method for extracting position features by the coordinate convolutional neural network module is specifically as follows:

[0030] The synchronous compressed time-frequency graph is used as the input data of the coordinate convolutional neural network module, and two coordinate channels are added to represent the horizontal and vertical coordinates of each pixel respectively. The horizontal and vertical coordinates are subjected to 2D convolution operation together with the input data to access the position information of each pixel as the position feature of the serpentine running signal in the frequency domain.

[0031] Furthermore, the output of the coordinate convolutional neural network module is expressed as:

[0032]

[0033] Where Y[i,j] represents the output value of the feature map processed by the coordinate convolutional neural network module at position (i,j), (m,n) represents the convolution kernel index, c represents the channel index, W, k, b, C represent the weight, convolution kernel size, bias and original channel number, respectively, and X coord Represents the input data after adding the coordinate channel.

[0034] The beneficial effects of the present invention are:

[0035] (1) The method of the present invention considers the wheelset motion state mainly from the two perspectives of wheelset lateral displacement and wheel-rail clearance. By analyzing the relationship between wheelset lateral displacement and equivalent taper and rolling circle wheel diameter difference, and considering the influence of snaking frequency and acceleration amplitude, the indirect characterization information of the influence of snaking motion on vehicle dynamic performance is determined, namely, the frame lateral and longitudinal accelerations and the frame lateral acceleration time-frequency information.

[0036] (2) The method of the present invention proposes a snaking monitoring network structure, in which the 1DCNN module extracts the time domain features of the lateral and longitudinal accelerations of the frame, the SWT map is combined with the Coordconv-2D module to extract the lateral time-frequency features of the frame, and the 3DCNN and Fully-connected modules fuse the time domain and time-frequency domain features. The network model can efficiently learn the coupling characteristics of snaking motion and complete the vehicle snaking state recognition task.

[0037] (3) In the method of the present invention, the performance of the proposed network was tested using real high-speed train snaking data. At the same time, four networks with different input characteristics were designed for comparison. The experimental results verified the rationality and effectiveness of the proposed network and signal preprocessing method. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A flow chart of a train bogie snaking state identification method based on wheelset motion state characteristics provided by the present invention.

[0039] Figure 2 The present invention provides a curve showing the difference in equivalent taper and rolling circle radius as a function of lateral displacement.

[0040] Figure 3 A logical evolution framework diagram of the snaking monitoring method based on wheelset motion state characteristics provided by the present invention.

[0041] Figure 4 This is a time-frequency effect comparison diagram provided by the present invention.

[0042] Figure 5 This is the overall structure diagram of the HMN network provided by the present invention.

[0043] Figure 6 This is a data label division flow chart provided by the present invention.

[0044] Figure 7 This is a schematic diagram of the model training loss provided by the present invention. DETAILED DESCRIPTION

[0045] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0046] The embodiment of the present invention provides a train bogie snaking state recognition method based on the wheelset motion state characteristics, such as Figure 1 As shown, the following steps are included:

[0047] S1, collecting the train frame acceleration signal, and selecting the frame lateral acceleration signal and longitudinal acceleration signal as the signal representing the motion state of the wheelset;

[0048] S2, filtering and time-frequency transforming the signal representing the motion state of the wheelset to obtain the snaking motion characteristics;

[0049] The snaking motion characteristics include the time-frequency domain characteristics of the lateral movement and the lateral shaking coupling characteristics;

[0050] S3, inputting the snaking motion feature into a snaking monitoring network, and outputting a snaking state classification result;

[0051] The snaking monitoring network includes a one-dimensional convolutional neural network module, a coordinate convolutional neural network module, a feature fusion module, a three-dimensional convolutional neural network module and a fully connected module;

[0052] The one-dimensional convolutional neural network module is used to extract the timing features of the input serpentine motion features in the time domain, and the coordinate convolutional neural network module is used to extract the position features of the input serpentine motion features in the frequency domain by introducing the coordinate channel; the feature fusion module is used to splice and fuse the extracted timing features and position features to obtain a coupled feature vector; the three-dimensional convolutional neural network module is used to extract the spatiotemporal information of multiple coupled feature vectors; the fully connected module is used to classify and identify the spatiotemporal information and output the serpentine state recognition result.

[0053] In the embodiment of the present invention, in the above step S1, a specific cause analysis process is given for selecting the frame lateral acceleration signal and the longitudinal acceleration signal as the wheel pair motion state characteristics.

[0054] Specifically, for the joint influence of the meandering frequency and the acceleration signal amplitude on the meandering operation, when the wheelset meanders but does not cause the wheel flange to collide with the rail, the frame undergoes simple harmonic motion, and its lateral acceleration amplitude can be approximately expressed as:

[0055]

[0056] Where Y W and f 0 They respectively represent the lateral displacement of the wheelset and the main frequency of snaking.

[0057] Considering the cost and ease of operation, the snake instability assessment method based on frame acceleration is usually adopted during the operation of high-speed trains, such as GB / T 5599-2019 and TSI standards. Taking the TSI standard (European standard) as an example, the standard stipulates that the peak value of the acceleration at the end of the frame is band-pass filtered at 3 to 9 Hz, and its peak value exceeds 8 m / s for 10 consecutive times. 2 , it is determined to be snaking instability. The wheel-rail clearance of European railways is about 5mm. In order to avoid wheel flange contact when the train is snaking, the allowed snaking amplitude is 2.5mm (with a 2.5mm safety margin). According to the above formula, when the snaking main frequency is 3 to 9Hz, the corresponding acceleration range is 0.1g to 0.8g (g is the acceleration of gravity). Therefore, the frame lateral acceleration limit in the TSI standard is set to 8m / s 2 The setting logic of the lateral acceleration limit in the GB / T 5599-2019 standard is similar to that of the TSI standard. However, the method of setting the alarm peak value by reserving a safety margin alone cannot accurately reflect the degree of vehicle snaking. It is easy to see from the above formula that the lateral displacement of the wheelset is affected by the snaking frequency and the lateral acceleration peak value. When the snaking motion of different snaking frequencies triggers the same acceleration limit alarm condition, the wheelset lateral displacement represented is not the same. Table 1 shows the estimated lateral displacement when the alarm is triggered under three different snaking frequencies.

[0058] Table 1 Comparison of wheelset lateral displacement at different snaking frequencies when the alarm is triggered

[0059]

[0060] If a fixed acceleration peak value is used as the only criterion for snaking instability, the operating risk of the vehicle may be misjudged, resulting in conservative or aggressive treatment. In the actual operation of high-speed trains, the generation of snaking motion is affected by many factors, and the snaking frequencies of snaking motion caused by different factors are also different. Therefore, in the subsequent process of this embodiment, the amplitude and frequency of the acceleration signal are comprehensively considered when judging the stability of snaking motion.

[0061] Regarding the impact of wheel-rail clearance on serpentine operation, in order to ensure the safe operation of the vehicle, the train needs to avoid wheel-rail collision as much as possible during operation. Different treads and rails will produce different wheel-rail clearances, which will directly affect the range of wheelset lateral displacement allowed for train operation. Generally, the larger the wheel-rail clearance, the greater the allowable wheelset lateral displacement, and vice versa. The equivalent taper and rolling circle radius difference can characterize the wheel-rail contact state to a certain extent. Take the equivalent taper and rolling circle radius difference curves of the LMA and LMB standard treads matched with the CHN60 rails as an example, as shown in Figure 2 shown.

[0062] According to TB / T 449-2016 standard, the wheel-rail clearance when the standard LMA and LMB treads match the CNH60 rails is approximately 8.1mm and 6.5mm. Figure 2 In the case of LMA and CHN60, when the wheelset lateral displacement is less than 7mm, the difference between the equivalent taper and the rolling circle radius increases slowly. When the wheelset lateral displacement is between 7 and 9mm, the difference between the equivalent taper and the rolling circle radius increases rapidly. When the wheelset lateral displacement is greater than 9mm, the difference between the equivalent taper and the rolling circle radius is large and shows a strong nonlinear growth characteristic. When LMB and CHN60 are matched, when the wheelset lateral displacement is less than 5mm, the difference between the equivalent taper and the rolling circle radius changes little. When the wheelset lateral displacement is between 5 and 7.5mm, the difference between the equivalent taper and the rolling circle radius increases rapidly. When the wheelset lateral displacement is greater than 7.5mm, the difference between the equivalent taper and the rolling circle radius is large and shows a strong nonlinear growth characteristic.

[0063] When the wheelset lateral displacement is close to the wheel-rail clearance, the equivalent taper and the rolling circle radius difference both show a rapid growth trend, especially when the wheelset lateral displacement exceeds the wheel-rail clearance to a certain extent, the equivalent taper and the rolling circle radius difference show a strong nonlinear growth trend, which means that the vehicle faces a high risk of contact between the wheel flange and the rail. Therefore, in order to avoid wheel-rail collision as much as possible, this embodiment takes into account the influence of the wheel-rail clearance margin when performing real-time monitoring of the vehicle's operating status in the subsequent process.

[0064] It is difficult to directly measure the wheel-rail contact state of a high-speed train during operation. Therefore, in this embodiment, a hunting monitoring network (Hunting Montior Net) is constructed by fusion of CNN features. The acceleration sensor signal is used to indirectly evaluate the vehicle running state and realize the hunting motion state recognition based on the wheel set motion state characteristics. The logical evolution framework of this monitoring method is shown in Figure 3. Normally, high-speed trains use high-rigidity positioning devices to connect the wheelset and bogie to ensure the stable operation of the wheelset driving the bogie. In theory, the motion state of the bogie and wheelset of a high-speed train will have a strong consistency, so the frame acceleration can be used as a reliable mapping of the motion state of the wheelset. The frame lateral acceleration contains the core information of the snaking frequency and acceleration amplitude, and is an important indicator for judging the snaking of the train. In addition, when the lateral displacement of the wheelset approaches or exceeds the wheel-rail gap, the rolling circle radius difference changes significantly, and this change can be mapped through the frame longitudinal acceleration.

[0065] exist Figure 3 In order to fully capture the complex dynamics of wheelset motion, this paper uses the filtered longitudinal and lateral accelerations of the frame as the input of HMN, and introduces the synchronous compressed time-frequency diagram of the lateral acceleration to enhance the model's ability to capture the frequency position characteristics of the snake. This design strategy aims to enable the model to more completely learn the coupling characteristics of the vehicle's lateral and yaw directions, and more accurately distinguish the characteristics of the wheelset's lateral displacement in different areas, thereby achieving an accurate assessment of the running status of the high-speed train.

[0066] based on Figure 3 As shown in the figure, step S2 of the embodiment of the present invention includes the following sub-steps:

[0067] S21, performing FFT filtering on the frame lateral acceleration signal to obtain a lateral acceleration bandpass filtered signal;

[0068] Perform FFT filtering on the longitudinal acceleration signal of the frame to obtain a longitudinal acceleration bandpass filter signal, and extract the amplitude and frequency information of the longitudinal acceleration signal;

[0069] S22, performing synchronous compression wavelet transform processing on the lateral acceleration bandpass filter signal to obtain a synchronous compression time-frequency diagram;

[0070] S23, extracting the time-frequency information of the lateral acceleration signal in the synchronous compression time-frequency diagram as a lateral motion time-frequency domain signal;

[0071] S24, extracting the amplitude and frequency information from the lateral acceleration bandpass filter signal, and enhancing it using the extracted time-frequency information to obtain the amplitude and frequency information of the lateral acceleration signal;

[0072] S25, fusing the longitudinal acceleration signal amplitude and frequency information with the lateral acceleration signal amplitude and frequency information to obtain the lateral movement and shaking head coupling characteristics.

[0073] In this embodiment, the vehicle's snaking signal has an obvious main frequency in the snaking frequency range of 0.5 to 10 Hz, and the position information of the main frequency in the snaking frequency range is very important for judging the snaking motion on the vehicle's dynamic performance. Although CoordConv performs well in tasks that require precise position information, its ability to extract time-frequency features is also affected by the time-frequency resolution of the input time-frequency graph.

[0074] In step S22 of this embodiment, in order to better extract the concentration degree and relative position information of the main frequency of the acceleration signal in the snaking frequency range, the synchronous compression wavelet transform time-frequency analysis method is selected. The effect comparison between the wavelet transform and the synchronous compression wavelet transform time-frequency analysis method (SWT) is as follows: Figure 4 It shows that, in addition to the advantages of wavelet transform, the SWT method is also superior in frequency resolution and has better time-frequency analysis display effect.

[0075] Based on this, step S22 of this embodiment includes the following sub-steps:

[0076] S22-1, performing continuous wavelet transform processing on the lateral acceleration bandpass filter signal;

[0077] Specifically, for a given time series signal, the continuous wavelet transform formula is expressed as:

[0078]

[0079] In the formula, a and b represent the scale and translation parameters respectively, ψ * represents the conjugate of the wavelet function;

[0080] S22-2, determining the instantaneous frequency of the signal according to the signal processed by continuous wavelet transform;

[0081] Specifically, the instantaneous frequency of the signal is expressed as:

[0082]

[0083] S22-3, redistributing the wavelet coefficients according to the instantaneous frequency of the signal and the signal after continuous wavelet transformation, obtaining a signal with a more concentrated time-frequency representation after synchronous compression, and then obtaining a synchronous compression wavelet transform time-frequency diagram;

[0084] Specifically, after synchronization compression, the time-frequency representation of the signal SWT is more concentrated y (ω,b) is expressed as:

[0085]

[0086] Where W y (a, b) represents the signal after continuous wavelet transform, ω y(a,b) represents the instantaneous frequency of the signal, a and b represent the scale and translation parameters respectively, ω represents the frequency, and δ represents the Dirac function.

[0087] In step S3 of the embodiment of the present invention, Figure 5 As shown, the snake detection network (HMN network) includes a one-dimensional convolutional neural network module (CoordConv), a coordinate convolutional neural network module (CoordConv), a feature fusion module (CoordConv), a three-dimensional convolutional neural network module (CoordConv) and a fully connected module (CoordConv).

[0088] In this embodiment, if Figure 5 As shown, the one-dimensional convolutional neural network module includes a first branch and a second branch with the same structure, which are respectively used to extract the time-frequency domain characteristics of the input lateral motion and the timing characteristics corresponding to the lateral movement and head shaking coupling characteristics.

[0089] Specifically, the amplitude and duration of the lateral acceleration of the frame are important references for identifying vehicle snaking instability, and the longitudinal acceleration signal contains the vehicle rolling wheel diameter difference information (mapping the position interval of wheel-rail contact). From the above analysis, it can be seen that the time domain information of the lateral and longitudinal acceleration signals of the frame can comprehensively reflect the characteristics of snaking motion. 1DCNN can effectively extract local features from time series signals and has high efficiency and generalization ability. Therefore, 1DCNN is selected to perform preliminary feature extraction on the lateral and longitudinal acceleration signals of the frame. The specific parameters of the above 1DCNN are shown in Table 2;

[0090] Table 2: 1DCNN parameter settings

[0091]

[0092] In the embodiment of the present invention, in the process of identifying the snaking state, the time domain characteristics of the signal contain limited information and cannot accurately describe the dynamic stability of the vehicle, so time-frequency analysis can be used as a supplement. The relative position information of the main frequency in the time-frequency diagram is very important for judging the degree of snaking instability, and the traditional 2DCNN has limitations in extracting position information. Therefore, in this embodiment, a coordinate convolutional neural network module is used to extract position features. By adding zero values ​​to the edges of the input data, it can ensure that the edge pixels also participate in the convolution operation and increase the receptive field of the convolution kernel, so that the convolution operation can better process the boundary information of the input data.

[0093] Specifically, in this embodiment, the method for extracting position features by the coordinate convolutional neural network module is as follows:

[0094] The synchronous compressed time-frequency graph is used as the input data of the coordinate convolutional neural network module, and two coordinate channels are added to represent the horizontal and vertical coordinates of each pixel respectively. The horizontal and vertical coordinates are subjected to 2D convolution operation together with the input data to access the position information of each pixel as the position feature of the serpentine running signal in the frequency domain.

[0095] Among them, the output of the coordinate convolutional neural network module is expressed as:

[0096]

[0097] Where Y[i,j] represents the output value of the feature map processed by the coordinate convolutional neural network module at position (i,j), (m,n) represents the convolution kernel index, c represents the channel index, W, k, b, C represent the weight, convolution kernel size, bias and original channel number, respectively, and X coord Represents the input data after adding the coordinate channel.

[0098] In this embodiment, the parameter settings of the coordinate convolutional neural network module (CoordConv) are shown in Table 3;

[0099] Table 3: CoordConv parameter settings

[0100]

[0101] In the embodiment of the present invention, Figure 5 In the snaking monitoring network shown in the figure, in order to better capture the complex relationships and patterns in the signal features, the feature fusion module (Merge+Reshape) is used to concatenate and fuse the time domain and frequency domain features into a single coupled feature vector with richer information to improve the performance and generalization ability of the model.

[0102] Furthermore, the three-dimensional convolutional neural network module (3DCNN) is good at processing data with time series characteristics and can capture the spatiotemporal information in the data in both time and space dimensions. The coupling features are extracted through the three-dimensional convolutional neural network module (3DCNN), and then the classification results are output through the fully connected module (Fully-connected); the parameter settings of the three-dimensional convolutional neural network module (3DCNN) are shown in Table 4;

[0103] Table 4: 3DCNN parameter settings

[0104]

[0105] The embodiment of the present invention provides a verification experiment example of the above identification method.

[0106] In this embodiment, the snaking data used are the measured running data of the CRH2-380A model during driving on the Shanghai-Hangzhou line, which uses a new standard LMA tread with rail CHN60. In a vehicle abnormal vibration report, it was pointed out that the CRH380a-6001 bogie had experienced abnormal snaking vibration during operation on the Shanghai-Hangzhou line. In order to study the cause of snaking of the CRH380a-6001 bogie, the anti-snaking shock absorber was replaced, the semi-active control strategy was partially closed or upgraded, and line experiments were carried out to generate vibration data under different working conditions. The experiment selected the lateral and longitudinal acceleration signals collected by the acceleration sensor at the train frame, and the sensor sampling frequency was 2500Hz. The collected original signal is first processed uniformly, first the data is band-pass filtered at 0.5Hz to 10Hz, and then downsampled at 250Hz. For the longitudinal acceleration of the frame, the longitudinal acceleration of the frame is band-pass filtered again [f-2, f+2] according to the highest main frequency f of the lateral acceleration of the frame after filtering. Finally, the processing time of each data item is 1.5 seconds, and the data set contains 767 data items in total.

[0107] Considering the differences in rolling circle radius and equivalent taper characteristics in different lateral displacement ranges, the estimated value of wheelset lateral displacement is selected as the basis for the classification of snaking status (although the calculation result of the formula is an ideal value, it can still roughly reflect the lateral displacement of the wheelset). According to the characteristics of the above-mentioned standard LMA tread and rail CHN60, the real data is divided into 4 categories (dangerous snaking, warning snaking, small snaking, and normal operation). The detailed classification process and sample composition are as follows: Figure 6 As shown in Table 5:

[0108] Table 5: Sample composition

[0109]

[0110] The high-speed train snaking state recognition experiment in this embodiment aims to achieve effective classification of the snaking motion of high-speed vehicles by extracting and learning signal features. In this experiment, 70% of the real data set is used as the training set, and 30% of the data set is used as the test set to test the network. The classification loss function and optimization algorithm are the cross-entropy loss function (Cross-entropy Loss) and the Adam optimization algorithm (Adaptive Moment Estimation). In the experiment, the learning rate of Adam is set to 0.0001, and the first-order attenuation factor and the second-order attenuation factor are 0.9 and 0.999 respectively. Figure 7 As shown in Figure 2, the network proposed in this paper converges after about 1200 epochs.

[0111] At present, the method used for snaking recognition is mainly based on the frame lateral acceleration signal. In order to verify whether adding the time-frequency characteristics of the frame longitudinal acceleration and the frame lateral acceleration in the proposed network is conducive to the recognition of snaking state and the good effect of the proposed network, four experiments are set up by changing the input features and network structure for comparison. In order to make the experiment more reasonable, the network used in each case is Figure 5 In order to avoid randomness, each network structure was tested 10 times and the average accuracy was taken as the final result. Table 6 shows the classification effect of each network:

[0112] Table 6: Average accuracy of different experiments

[0113]

[0114] Based on the average accuracy of each experiment, it can be seen that the overall accuracy achieved by using only the frame lateral acceleration as an input feature is the lowest, at only 87.45%. Using the frame lateral acceleration input plus the frame lateral acceleration signal, or the frame lateral acceleration CWT map as input features, the overall accuracy achieved is improved to 94.81% and 91.34%, respectively, and there is also an improvement in the accurate prediction of various categories. It can be seen that both the frame longitudinal acceleration and the wavelet time-frequency map contain important information that helps improve the accuracy of model recognition. The frame lateral, longitudinal acceleration and CWT Figure 3 When the model is used as input, the overall recognition accuracy of the model is 93.51%. Although the recognition accuracy has improved, the effect is not ideal. Figure 3 When the proposed model is used as input, a better recognition accuracy of 98.7% is achieved. In summary, the experiments prove that the proposed model takes into account the lateral and longitudinal acceleration of the frame and the lateral time-frequency domain characteristics of the frame, has high accuracy, and can be used for the task of high-speed vehicle snaking state recognition with higher precision.

[0115] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0116] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A train bogie snaking state recognition method based on wheelset motion state characteristics, characterized in that: The following steps are involved: S1, collecting the train frame acceleration signal, and selecting the frame lateral acceleration signal and longitudinal acceleration signal as the signal representing the motion state of the wheelset; S2, filtering and time-frequency transforming the signal representing the motion state of the wheelset to obtain the snaking motion characteristics; The snaking motion characteristics include the time-frequency domain characteristics of the lateral movement and the lateral shaking coupling characteristics; S3, inputting the snaking motion feature into a snaking monitoring network, and outputting a snaking state classification result; The snaking monitoring network includes a one-dimensional convolutional neural network module, a coordinate convolutional neural network module, a feature fusion module, a three-dimensional convolutional neural network module and a fully connected module; The one-dimensional convolutional neural network module is used to extract the timing features of the input serpentine motion features in the time domain, and the coordinate convolutional neural network module is used to extract the position features of the input serpentine motion features in the frequency domain by introducing the coordinate channel; the feature fusion module is used to splice and fuse the extracted timing features and position features to obtain a coupled feature vector; the three-dimensional convolutional neural network module is used to extract the spatiotemporal information of multiple coupled feature vectors; the fully connected module is used to classify and identify the spatiotemporal information and output the serpentine state recognition result.

2. The train bogie snaking state identification method based on wheelset motion state characteristics according to claim 1 is characterized in that: The step S2 comprises the following sub-steps: S21, performing FFT filtering on the frame lateral acceleration signal to obtain a lateral acceleration bandpass filtered signal; Perform FFT filtering on the longitudinal acceleration signal of the frame to obtain a longitudinal acceleration bandpass filter signal, and extract the amplitude and frequency information of the longitudinal acceleration signal; S22, performing synchronous compression wavelet transform processing on the lateral acceleration bandpass filter signal to obtain a synchronous compression time-frequency diagram; S23, extracting the time-frequency information of the lateral acceleration signal in the synchronous compression time-frequency diagram as a lateral motion time-frequency domain signal; S24, extracting the amplitude and frequency information from the lateral acceleration bandpass filter signal, and enhancing it using the extracted time-frequency information to obtain the amplitude and frequency information of the lateral acceleration signal; S25, fusing the longitudinal acceleration signal amplitude and frequency information with the lateral acceleration signal amplitude and frequency information to obtain the lateral movement and shaking head coupling characteristics.

3. The train bogie snaking state identification method based on wheelset motion state characteristics according to claim 2 is characterized in that: The step S22 comprises the following sub-steps: S22-1, performing continuous wavelet transform processing on the lateral acceleration bandpass filter signal; S22-2, determining the instantaneous frequency of the signal according to the signal processed by continuous wavelet transform; S22-3. According to the instantaneous frequency of the signal and the signal after continuous wavelet transformation, the wavelet coefficients are redistributed to obtain a signal with a more concentrated time-frequency representation after synchronous compression, and then a synchronous compression wavelet transform time-frequency diagram is obtained.

4. The train bogie snaking state identification method based on wheelset motion state characteristics according to claim 3 is characterized in that: In step S22-3, the time-frequency representation of the synchronously compressed signal SWT is more concentrated. y (ω,b) is expressed as: Where W y (a, b) represents the signal after continuous wavelet transform, ω y (a,b) represents the instantaneous frequency of the signal, a and b represent the scale and translation parameters respectively, ω represents the frequency, and δ represents the Dirac function.

5. The train bogie snaking state identification method based on wheelset motion state characteristics according to claim 1 is characterized in that: In step S3, the one-dimensional convolutional neural network module includes a first branch and a second branch with the same structure, which are respectively used to extract the time-frequency domain characteristics of the input lateral motion and the timing characteristics corresponding to the lateral movement and head shaking coupling characteristics.

6. The train bogie snaking state identification method based on wheelset motion state characteristics according to claim 2 is characterized in that: In step S3, the method for extracting position features by the coordinate convolutional neural network module is specifically as follows: The synchronous compressed time-frequency graph is used as the input data of the coordinate convolutional neural network module, and two coordinate channels are added to represent the horizontal and vertical coordinates of each pixel respectively. The horizontal and vertical coordinates are subjected to 2D convolution operation together with the input data to access the position information of each pixel as the position feature of the serpentine running signal in the frequency domain.

7. The train bogie snaking state identification method based on wheelset motion state characteristics according to claim 6 is characterized in that: The output of the coordinate convolutional neural network module is expressed as: Where Y[i,j] represents the output value of the feature map at position (i,j) after convolution by the coordinate convolutional neural network module, (m,n) represents the convolution kernel index, c represents the channel index, W, k, b, C represent the weight, convolution kernel size, bias and original channel number, respectively, and X coord Represents the input data after adding the coordinate channel.

Images