Track irregularity identification method and system based on discrete wavelet filter network

Abstract

This invention provides a method and system for identifying track irregularities based on a discrete wavelet filter network, belonging to the field of track irregularity identification technology. The method includes installing acceleration sensors at the front and rear ends of the train body to acquire vertical acceleration signals generated during operation. The signals are preprocessed, and the delay is calculated based on the running speed and a fixed distance. The signals from the front and rear ends are aligned and stitched together on the time axis to construct a dual-channel sample matrix. A recognition model is constructed and trained. This model performs multi-scale decomposition of each channel signal using a learnable discrete wavelet filter network, applies learnable scalar weights to each coefficient sub-band, and then reconstructs the signal to selectively enhance time-domain features. The reconstructed signal is then normalized, and through linear combination and projection, an estimated track irregularity value corresponding to the end of the input signal time period is output. The trained model is used to identify real-time samples. This method achieves continuous and efficient identification of track irregularities.

Description

Technical Field

[0001] This invention relates to the field of track irregularity identification technology, and more specifically, to a method and system for track irregularity identification based on discrete wavelet filter networks. Background Technology

[0002] Railway tracks, subjected to a complex interplay of factors including long-term train dynamic loads, environmental temperature variations, and structural aging, are prone to varying degrees of geometric degradation, leading to track irregularities. The continued development of track irregularities not only exacerbates vibration and noise levels during train operation, reducing passenger comfort, but also causes fluctuations in wheel-rail forces, adversely affecting the operational safety and service life of the vehicle-track system. Therefore, achieving efficient and stable track irregularity identification under operational conditions is of significant engineering importance for improving railway line condition monitoring capabilities and maintenance efficiency.

[0003] Currently, research on track irregularity identification largely focuses on data-driven methods based on the vibration response of operating trains. These methods collect dynamic response signals by deploying acceleration sensors at key points on the train and then directly learn the mapping relationship between the train response and track irregularities using neural networks. While data-driven methods reduce reliance on physical models, they generally employ complex neural network models with large parameter scales, resulting in high computational costs and sensitivity to noise and changes in operating conditions, which hinders online applications and engineering deployment.

[0004] To address the aforementioned issues, there is a need for a method and system that can effectively reduce the size of model parameters and computational complexity while ensuring recognition accuracy, improve stability and engineering applicability under complex operating conditions, and be suitable for continuous monitoring and online identification of track irregularities, in order to solve the above-mentioned technical problems. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for identifying track irregularities based on discrete wavelet filtering networks, in order to improve the aforementioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows:

[0006] Firstly, this application provides a method for identifying track irregularities based on a discrete wavelet filter network, including:

[0007] An acceleration sensor is installed at a predetermined position on the vehicle body, and the front vertical acceleration signal and rear vertical acceleration signal generated by the sensor are acquired when the train is running on the track.

[0008] Preprocessing is performed on all vertical acceleration signals, in which signal segments of a preset fixed time length are aligned and combined to obtain a sample matrix containing dual-channel vibration information;

[0009] A preset historical sample matrix is ​​sent to a preset track irregularity recognition model for training to obtain a trained track irregularity recognition model. The preset historical sample matrix is ​​a sample matrix of historical data collected by the accelerometer after preprocessing.

[0010] The sample matrix of the dual-channel vibration information is sent to the track irregularity recognition model after training to identify track irregularities and obtain the real-time recognition result of track irregularities.

[0011] Secondly, this application also provides a track irregularity identification system based on a discrete wavelet filter network, comprising:

[0012] The acquisition module is used to set up acceleration sensors at preset positions on the vehicle body and acquire the front vertical acceleration signal and rear vertical acceleration signal generated by the sensors when the train is running on the track.

[0013] The processing module is used to preprocess all vertical acceleration signals, wherein a sample matrix containing dual-channel vibration information is obtained by aligning and combining signal segments of preset fixed time lengths.

[0014] The training module is used to send a preset historical sample matrix to a preset track irregularity recognition model for training, and obtain a track irregularity recognition model after training. The preset historical sample matrix is ​​a sample matrix after preprocessing the historical data collected by the accelerometer.

[0015] The processing module is used to send the sample matrix of the dual-channel vibration information to the track irregularity recognition model after training to identify track irregularities and obtain the real-time recognition result of track irregularities.

[0016] The beneficial effects of this invention are as follows:

[0017] This invention introduces a learnable wavelet multi-scale filtering structure to decompose, weight, and reconstruct train body acceleration signals, thereby achieving adaptive adjustment of features across different frequency bands during model training. Based on this, a lightweight feedforward neural network is used to establish an end-to-end mapping relationship between train body acceleration and track irregularities. This invention effectively reduces model parameter size and computational complexity while maintaining recognition accuracy, improving stability and engineering applicability under complex operating conditions, and is suitable for continuous monitoring and online identification of track irregularities.

[0018] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the track irregularity identification method based on discrete wavelet filter network described in an embodiment of the present invention;

[0021] Figure 2 This is a schematic diagram of the track irregularity identification system based on discrete wavelet filter network described in an embodiment of the present invention.

[0022] Figure 3 This is a model structure diagram of the track irregularity identification method based on discrete wavelet filter network described in this embodiment of the invention;

[0023] Figure 4 This is a time-domain comparison diagram of the track irregularity identification results described in this embodiment of the invention.

[0024] In the diagram: 701, Acquisition module; 702, Processing module; 703, Training module; 704, Recognition module. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0026] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0027] Example 1:

[0028] This embodiment provides a method for identifying track irregularities based on a discrete wavelet filter network.

[0029] See Figure 1 , Figure 3 and Figure 4 The figure shows that the method includes steps S1, S2, S3, and S4.

[0030] Step S1: Install an acceleration sensor at a preset position on the vehicle body, and acquire the front vertical acceleration signal and rear vertical acceleration signal generated by the sensor when the train is running on the track.

[0031] It is understandable that this step involves installing accelerometers at two specific locations on the same side of the train body—the front and rear ends—to simultaneously collect two vertical acceleration signals as the train travels along the track. This arrangement essentially captures the dynamic response of the train body to the same track excitation at both ends. These two raw time-domain signals form the direct data basis for subsequent analysis. Their inherent time correlation and amplitude differences imply crucial information about the location and severity of track irregularities. Specifically, the accelerometers are installed at the front and rear corners on the left (or right) side of the train carriage to collect the train body's vertical acceleration.

[0032] Step S2: Preprocess all vertical acceleration signals, wherein signal segments of a preset fixed time length are aligned and combined to obtain a sample matrix containing dual-channel vibration information.

[0033] Understandably, this step uses a sliding window of preset time length to synchronously capture two parallel acceleration signal streams. The key operation involves aligning and correcting the front and rear signal segments along the time axis based on the fixed distance and speed of the vehicle body, and then combining them into a dual-channel sample matrix. This process essentially transforms the data from a "time stream" to "spatial segment corresponding samples," ensuring that each sample matrix encapsulates the vibration history of the front and rear ends of the vehicle body as they traverse the same track segment in the time dimension. This establishes a structured input format for the model that is associated with a specific track segment.

[0034] It is understood that in this process, step S2 includes steps S21, S22 and S23.

[0035] Step S21: Based on the acquired front-end vertical acceleration signal, multiple continuous front-end vibration signal sample segments are extracted from the front-end signal sequence through a predetermined time window that slides with a fixed step size.

[0036] It is understandable that this step extracts a series of temporally continuous front-end vibration signal sample segments from the continuous front-end signal sequence through a predetermined time window that slides with a fixed step size. This operation transforms the infinite real-time signal stream into signal segments of finite length that can be analyzed independently, providing basic units for subsequent processing.

[0037] Step S22: Based on the acquired rear vertical acceleration signal and the front vibration signal sample segment, by calculating the delay caused by the train passing through a fixed distance of its body at the running speed, determine the rear signal starting point corresponding to each front sample segment in the track space, and extract signals of the same time length to obtain the rear vibration signal sample segment corresponding to the front vibration signal sample segment.

[0038] It is understandable that this step, acquiring the vibration response at another location on the train body corresponding to these front-end sample segments in space, requires accurately calculating the time delay of the rear-end signal relative to the front-end signal based on the train's real-time operating speed and the fixed distance between the front and rear acceleration sensors on the train body. This delay is then used to determine the starting point of the rear-end signal corresponding to each front-end sample segment, and subsequently, rear-end vibration signal sample segments of the same time length are extracted. This ensures that although each pair of front-end and rear-end sample segments has a time domain offset, they correspond to the exact same segment in track space, which is crucial for eliminating the influence of the train body's geometry and achieving spatial alignment of the vibration signals. Specifically, this step uses acceleration sensors arranged at the front and rear ends on the same side of the train body to collect vertical acceleration signals from the front and rear of the train body, respectively, denoted as... and ,in The sampling time is indicated. The two acceleration signals constitute a dual-channel input, used to characterize the vertical vibration characteristics of the train body during operation.

[0039] Step S23: Based on the front-end vibration signal sample segment and the rear-end vibration signal sample segment, align the front-end and rear-end signal sample segments representing the vibration of the same track segment on the time axis and splice them in the channel dimension to construct the sample matrix containing dual-channel vibration information.

[0040] Understandably, this step precisely aligns the spatially aligned front and rear vibration signal sample segments on the time axis, ensuring they represent the responses of different vehicle positions to the same track excitation at the same moment. These two aligned signals are then stitched together along the channel dimension to construct a sample matrix containing dual-channel vibration information that comprehensively characterizes the combined dynamic response of the front and rear ends of the vehicle to the same track segment. This matrix serves as the input for subsequent deep feature extraction, and its structure effectively preserves the spatial correlation information of the track excitation during its transmission within the vehicle. Specifically, this step uses a length of [missing information - likely a unit of length] for the continuously acquired dual-channel vertical acceleration signals of the vehicle. The sliding time window is used to segment the data and construct the model input samples. The input sample is represented as:

[0041]

[0042] in, Indicates the first The sample matrix of input samples, and They represent the first Vertical acceleration signals from the front and rear of the train body for each input sample. Indicates the length of the time window. Represents the real number field.

[0043] Step S3: Send the preset historical sample matrix to the preset track irregularity recognition model for training to obtain the track irregularity recognition model after training. The preset historical sample matrix is ​​the sample matrix after preprocessing the historical data collected by the accelerometer.

[0044] Understandably, this step uses a large matrix of preprocessed historical samples as input and the corresponding real track irregularity data as the learning target to train the pre-defined track irregularity recognition model. This process enables the model to adaptively adjust its internal parameters, autonomously learning and summarizing deep features and mapping patterns that are sensitive to changes in track condition but relatively stable to changes in vehicle speed, load, and other factors from a massive number of "vibration-irregularity" correspondence examples. Ultimately, this results in a trained model with fixed parameters and generalized recognition capabilities.

[0045] It is understood that in this process, step S3 includes steps S31 and S32.

[0046] Step S31: Based on the sample matrix, scale decompose the signal of each channel, and by applying a learnable adjustment parameter to each coefficient sub-band obtained by decomposition, synthesize the adjusted coefficient sub-band to obtain a reconstructed signal whose features are selectively enhanced in the time domain.

[0047] Understandably, this step processes the sample matrix containing dual-channel vibration information. First, scale decomposition is performed on each channel signal in the matrix. This operation uses discrete wavelet transform to recursively decompose the waveform of each vibration signal in the time domain into a series of components representing different frequency ranges, thus obtaining a set of coefficient subbands containing low-frequency approximation coefficients and high-frequency detail coefficients. The purpose is to separate the mixed time-domain vibrations into different frequency subbands for subsequent targeted analysis. Next, by applying a learnable adjustment parameter to each coefficient sub-band obtained from the decomposition and performing a uniform scalar multiplication on all coefficients within the sub-band, the essence of this operation is to allow the model to autonomously learn and assign weights to different frequency sub-bands during the training process. This enables it to adaptively enhance frequency features sensitive to track irregularities (such as vibration components related to specific wavelength imbalances) while suppressing the influence of irrelevant noise or the vehicle's own vibration modes. Subsequently, the weighted coefficient sub-bands are used for signal synthesis, i.e., they are re-merged through inverse discrete wavelet transform, resulting in a reconstructed signal with selectively enhanced features in the time domain. This reconstructed signal retains the main form of the original signal, but highlights the feature components useful for the recognition task.

[0048] In this step, the constructed input samples are fed into a trained learnable discrete wavelet filter network (LDWF-Net) for processing. The LDWF-Net includes a learnable discrete wavelet filter module and a feedforward neural network module. The learnable discrete wavelet filter module performs multi-scale discrete wavelet decomposition on the input dual-channel acceleration signal, applies learnable weights to the wavelet coefficients at different scales, and then reconstructs the signal, thereby achieving adaptive enhancement or suppression of features in different frequency bands.

[0049] In this step, step S31 includes steps S311, S312 and S313.

[0050] Step S311: Separate each channel signal in the sample matrix at multiple scales using a preset learnable discrete wavelet filter network to obtain low-frequency approximation coefficients and high-frequency detail coefficients at each scale;

[0051] Understandably, this step uses a pre-defined learnable discrete wavelet filter network to separate each channel signal in the sample matrix at multiple scales. Here, "scale separation" specifically refers to using the mathematical tool of discrete wavelet transform to recursively decompose the signal into frequency components at different levels (scales). Each level of decomposition generates a set of low-frequency approximation coefficients that characterize the general appearance of the signal at that scale and a set of high-frequency detail coefficients that characterize the detailed changes of the signal at that scale. This transforms the mixed time-domain vibrations from the "time domain" to the "time-frequency domain," enabling subsequent analysis of different frequency bands.

[0052] Specifically, for each acceleration signal, the Discrete Wavelet Transform (DWT) is used to decompose it into a series of low-frequency approximation coefficients. and high frequency detail coefficient ,in This indicates the scale (number of layers) of the decomposition. The approximation coefficients of each layer are processed by high-pass and low-pass filters, and then downsampled to obtain the corresponding high-frequency detail coefficients and low-frequency approximation coefficients, specifically expressed as follows:

[0053]

[0054]

[0055] in, These are low-frequency approximation coefficients. For high-frequency detail coefficients, The scale (number of layers) for decomposition. Indicates the first The value of the low-frequency approximation coefficient at the k-th position, i.e., the current low-frequency approximation coefficient at the k-th position. The input signal for layer decomposition. When hour, The original input signal, Indicates the first The wavelet coefficient index after layer decomposition, i.e. or The position number at the current scale. Since wavelet decomposition involves double downsampling, therefore... This indicates the position of the coefficients after downsampling, rather than the sampling point number of the original signal. [·] represents the filter coefficients of the high-pass decomposition filter, used to extract high-frequency detail components; These represent the filter coefficients of the low-pass decomposition filter, used to extract low-frequency approximation components. Therefore, Indicates the calculation of the first High-frequency detail factor At that time, the high-pass filter approximates the coefficients of the previous layer. The applied filter weights.

[0056] Step S312: By applying a learnable scalar weight to the low-frequency approximation coefficients and high-frequency detail coefficients at each scale, and performing a uniform scalar multiplication on all coefficients, the weighted low-frequency approximation coefficients and high-frequency detail coefficients are obtained.

[0057] Understandably, this step involves applying a learnable scalar weight to the low-frequency approximation coefficients and high-frequency detail coefficients obtained at each scale, and performing a uniform scalar multiplication on all coefficients within that sub-band. Here, "learnable scalar weight" means that each frequency sub-band (such as the high-frequency detail sub-band at the same scale) shares the same weight value. This value is not preset but is obtained through backpropagation and automatic optimization during model training based on the performance of the final recognition task. Furthermore, features related to track irregularities in vehicle vibration are often concentrated in specific frequency bands, while other frequency bands may contain vehicle vibrations or random noise. This "sub-band level" rather than "coefficient level" weighting method allows the network to adaptively enhance the overall amplitude of key frequency band signals beneficial to recognition with a small parameter cost, while suppressing irrelevant or interfering frequency bands. In the specific task of track irregularity recognition, this helps the model robustly focus on vibration modes dominated by track geometric excitation, rather than individual vehicle differences or instantaneous impacts.

[0058] Among them, for each layer This process is repeated recursively until a predetermined maximum number of decomposition layers is reached. After the signal is decomposed into multiple sub-bands, each sub-band (including approximation coefficients and detail coefficients) is applied with a learnable weight matrix. , so as to perform weighted processing. For the first Layer detail factor Approximation coefficients of the final layer Use the corresponding learning weights and To perform scalar multiplication, we obtain the weighted coefficients:

[0059] ;

[0060] in, For the first The weighted coefficients of the layers, For the first Layer learning weights For the first The detail factor of the layer, The weighted coefficients of the final layer. The learning weights for the final layer, These are the approximation coefficients for the final layer.

[0061] This weighting mechanism allows the signal in each sub-band to be flexibly enhanced or suppressed based on its frequency characteristics. All signals are in the same decomposition layer. The coefficients in the model share a common weight. Instead of assigning weights individually to each coefficient within that layer, this approach not only stabilizes the training process and reduces parameter fluctuations, but also effectively suppresses noise and outliers while lowering computational costs.

[0062] Step S313: The weighted low-frequency approximation coefficients and high-frequency detail coefficients are recombined by inverse discrete wavelet transform to obtain a reconstructed signal with selectively enhanced features in the time domain.

[0063] It is understandable that the weighted low-frequency approximation coefficients and high-frequency detail coefficients of each scale are re-merged by inverse discrete wavelet transform. This "reconstruction" step is the inverse process of "decomposition". It re-integrates the coefficients of each frequency band after selective enhancement in the time-frequency domain and transforms them back into a time-domain signal, thereby obtaining a reconstructed signal with selectively enhanced features in the time domain. This output signal retains the temporal structure of the original signal, but its internal vibration feature components that are sensitive to the orbital state have been significantly enhanced by the weighting mechanism learned by the network, providing higher quality and more targeted input features for subsequent normalization and regression mapping steps.

[0064] In this step, the inverse discrete wavelet transform (IDWT) is applied to the weighted detail coefficients and approximation coefficients to reconstruct them back into the time-domain signal. The formula is as follows:

[0065]

[0066] in, The weighted coefficients of the final layer. This represents the signal reconstructed by the inverse discrete wavelet transform. Represents the inverse discrete wavelet transform; Indicates the first floor to the second floor. The set of high-frequency detail coefficients of the layer, For the first High-frequency detail coefficients after weighted processing of the layer.

[0067] Step S32: Based on the numerical distribution of the reconstructed signal, perform normalization processing, and then perform linear combination and projection of the normalized signal features to obtain the track irregularity estimate corresponding to the end of the input signal time period.

[0068] Understandably, this step involves normalizing and mapping the reconstructed signal. First, the numerical distribution is normalized. This is typically done by calculating and utilizing the mean and variance of data within a batch, shifting and scaling the signal to convert it into a zero-mean, unit-variance distribution. This aims to eliminate the dimensional influence caused by differences in overall vibration energy between different samples, improving the stability of model training. Finally, the normalized signal undergoes linear combination and projection of features. This involves using a learnable weight matrix and bias vector to linearly transform the signal, mapping it from the normalized feature space to a target space that directly represents the track state, thus directly outputting the track irregularity estimate corresponding to the end of the input signal time period. The entire process, from multi-scale decomposition and feature separation to adaptive weighting to highlight key information, then to normalizing and stabilizing the data distribution, and finally completing regression estimation through linear transformation, gradually transforms the original vibration signal into an accurate track geometric parameter estimate. In this step, step S32 includes steps S321 and S322.

[0069] Step S321: Input the reconstructed signal into the normalization layer of a preset feedforward neural network for processing, wherein the normalization process is performed by a preset normalization formula to obtain the normalized reconstructed signal;

[0070] It is understood that this step inputs the reconstructed signal into the normalization layer of a pre-defined feedforward neural network for processing. The "normalization layer" specifically refers to layer normalization technology. This processing calculates the mean and variance of the feature vectors of the reconstructed signal at each independent time step, and performs centering, translation, and scaling on the feature vectors based on these statistics to obtain the normalized reconstructed signal. The core purpose of this operation is to eliminate the changes in internal data distribution caused by the differences in the overall energy of the vibration signals between samples (such as due to differences in train speed or load) during training, so that the data processed by subsequent network layers is within a stable numerical range. This can improve the model training convergence speed and generalization stability under varying operating conditions.

[0071] Among them, after completing wavelet reconstruction and obtaining the signal The reconstructed signal will then be further processed by a feedforward neural network. The feedforward neural network (FNN) includes a LayerNorm layer for normalization and a Dense fully connected layer. The specific processing flow is as follows: First, the reconstructed signal... It is input into the LayerNorm layer for normalization. For each time step eigenvectors of the signal Normalization will be performed independently on the feature dimension. LayerNorm applies the following formula to each... Vector normalization:

[0072]

[0073] Among them, LayerNorm represents the layer normalization layer processing. Indicates the first The input feature vector at each time step; and These represent the mean and variance of the feature vector along its feature dimension, respectively. To prevent extremely small constants with a denominator of zero; ⊙ represents element-wise multiplication; and These represent the learnable scaling and translation parameters, respectively.

[0074] Step S322: Input the normalized reconstructed signal into the fully connected layer of the feedforward neural network for processing. The normalized reconstructed signal is flattened into a fixed-length vector, and the fixed-length vector is linearly transformed to obtain the track irregularity estimate corresponding to the end of all vertical acceleration signal time periods.

[0075] Understandably, this step inputs the normalized reconstructed signal into the fully connected layer of the feedforward neural network for processing. Here, the "fully connected layer" is a linear transformation layer. Its processing first concatenates the normalized reconstructed signal, which still retains its time-step structure, along the time dimension, flattening it into a single feature vector of fixed length to integrate information from all time steps. Then, this fixed-length feature vector undergoes a linear transformation, i.e., a weighted summation using a learnable weight matrix, plus a learnable bias vector. Finally, the information in the high-dimensional feature space is mapped to a scalar output, yielding an estimate of the track irregularity corresponding to the end of the input signal's time period. The entire process begins with data normalization to stabilize the input distribution and ends with the precise mapping from features to the target through linear transformation.

[0076] The signal after LayerNorm normalization It is fed into a fully connected layer for processing. First, the input normalized signal... It will be flattened into a vector of fixed length. Then, these flattened vectors undergo a linear transformation through a fully connected layer to obtain the output. :

[0077]

[0078] in, This is the estimated track irregularity value corresponding to the end of the time window. For normalized signal A vector of fixed length for the exhibition platform. This represents the learnable weight parameters of the fully connected output layer, used to map feature vectors to track irregularity estimates; This represents a learnable bias term.

[0079] In the model training phase, measured or reference track irregularity data are used as supervision signals. The network parameters are trained by minimizing the mean square error between the predicted results and the reference values. The loss function is defined as follows:

[0080]

[0081] in, For loss function, For the sample size, For reference, the track is not smooth. Output results for the model. This indicates the length of the time window.

[0082] During the model application phase, the real-time collected vertical acceleration signal of the vehicle body is input into the trained LDWF-Net, and the track irregularity recognition results are output point by point through a sliding time window, thereby realizing the continuous recognition of track irregularities in the target track section.

[0083] Step S4: Send the sample matrix of the dual-channel vibration information to the track irregularity recognition model after training to perform track irregularity recognition and obtain the real-time recognition result of track irregularity.

[0084] Understandably, this step involves real-time acquisition of the current dual-channel vibration sample matrix, which is preprocessed in the same manner as in step S2, and input into the trained model. Based on the learned patterns, the model analyzes and infers the vibration characteristics of the current track segment represented by the input samples, and outputs a corresponding track irregularity estimate. This process achieves end-to-end, continuous identification of the track state, transforming the real-time vibration signal stream into intuitively understandable track quality indicators, thus completing the final closed loop from data perception to state evaluation.

[0085] Example 2:

[0086] This embodiment provides a specific set of data, in which LDWF-Net is used to describe the non-roughness recognition model.

[0087] The dataset, consisting of vehicle vertical acceleration and corresponding track irregularities, was divided into training, validation, and test sets in a 7:1:2 ratio for each vehicle speed. The training set covered an 800-meter track length, while the test set covered 200 meters. To ensure rigorous evaluation, the test set was completely separated from the training set in terms of track intervals. Considering that vehicle vibrations are mainly concentrated in the low-frequency range, the acceleration signals and corresponding track irregularity data were low-pass filtered using cutoff frequencies of 2Hz, 10Hz, and 20Hz, respectively. The sampling frequencies were adjusted accordingly to 4Hz, 20Hz, and 40Hz. This preprocessing step helps improve the training efficiency of the neural network. The network input consisted of 32-sequence long sequences from two channels, corresponding to the acceleration signals measured at the front, rear, and left sides of the vehicle body, respectively. These signals were normalized to the [0,1] range using Min-Max scaling. The number of decomposition layers in the LDWF module was determined by the user-specified frequency of interest (1Hz in this study), which corresponds to the dominant frequency band of the vehicle acceleration response. The output of the LDWF module, after being normalized by LayerNorm, is fed into a feedforward neural network containing a fully connected layer with 32 ReLU units and a dropout rate of 0.3. The final linear output neuron is used to estimate the vertical orbital displacement. The model is trained for 100 epochs using the Adam optimizer with a learning rate of 0.001 and a batch size of 16.

[0088] To quantitatively evaluate model performance, normalized root mean square error (NRMSE) is used to measure the difference between the identified track irregularities and the true values.

[0089] To demonstrate the superiority of our model, we compared it with two benchmark models, LSTM and CNN-BiLSTM, both using the front and rear accelerations on the left side of the vehicle as input and track irregularities as output. To ensure a fair comparison, all models were trained using the same data partitioning, optimization strategy, and training protocol, and hyperparameters were tuned on the validation set.

[0090] The LSTM baseline model employs a two-layer stacked structure, each layer containing 32 hidden units, followed by a 16-unit ReLU fully connected layer, ultimately outputting a linear layer with a sequence-to-single output configuration. The CNN-BiLSTM baseline model consists of two one-dimensional convolutional layers (16 kernels, kernel size 7, ReLU activation, identical padding), each followed by batch normalization and dropout (dropout rate = 0.35), and two stacked bidirectional LSTM layers with 32 hidden units each. Extracted features are further processed through two 16-unit ReLU fully connected layers, using L2 regularization (weight decay coefficient = 0.001) and dropout (dropout rate = 0.30), before outputting the final linear result. The final configurations of both baseline models were obtained through extensive validation tuning.

[0091] The number of trainable parameters for LDWF-Net, LSTM, and CNN-BiLSTM are 787, 1.3 × 10⁻⁶, and 1.3 × 10⁻⁶, respectively. 4 and 4.1×10 4 This indicates that the proposed model is lightweight. Under the same conditions (80 km / h), the time-domain components of track irregularities identified by different models are as follows: Figure 4 As shown. (Through) Figure 4 As can be seen, LDWF-Net performs significantly better. In the time domain, its predicted values ​​closely match the true values ​​across the entire distance, with only a small deviation near steep peaks, indicating that it effectively captures the dominant low-frequency components. To quantitatively compare the recognition performance of different methods under different conditions, we have summarized the results in the model data analysis table below. This table includes the recognition of the three models for the three frequencies, along with the training time for each session and the normalized root mean square error (NRMSE) to quantitatively evaluate the model performance. The model data analysis table is shown below:

[0092]

[0093] As shown in the table above, due to the significantly fewer trainable parameters, LDWF-Net requires the shortest training time among the three models, reducing it by 25%-57%. In terms of accuracy, LDWF-Net consistently achieves the lowest NRMSE across all cutoff frequencies, followed by LSTM, while CNN-BiLSTM produces the largest error. Notably, this performance ranking remains constant across all cutoff frequencies and operating speeds, with NRMSE decreasing by 0.04-0.47. At the three upper cutoff frequencies, CNN-BiLSTM exhibits the largest error increase (from 1.47 to 1.70, then to 2.06), indicating its high sensitivity to high-frequency noise under the limited training data conditions in this study. In contrast, LDWF-Net demonstrates the smallest performance degradation (from 1.29 to 1.50, then to 1.59), indicating its superior robustness and stability when introducing higher-frequency components.

[0094] Example 3:

[0095] like Figure 2 As shown, this embodiment provides a track irregularity identification system based on a discrete wavelet filter network. See [link to documentation]. Figure 2 The system includes an acquisition module 701, a processing module 702, a training module 703, and a processing module 704.

[0096] The acquisition module 701 is used to set an acceleration sensor at a preset position on the vehicle body and acquire the front vertical acceleration signal and the rear vertical acceleration signal generated by the sensor when the train is running on the track.

[0097] The processing module 702 is used to preprocess all the vertical acceleration signals, wherein a sample matrix containing dual-channel vibration information is obtained by aligning and combining signal segments of a preset fixed time length.

[0098] Training module 703 is used to send a preset historical sample matrix to a preset track irregularity recognition model for training, and obtain a track irregularity recognition model after training. The preset historical sample matrix is ​​a sample matrix after preprocessing the historical data collected by the accelerometer.

[0099] The processing module 704 is used to send the sample matrix of the dual-channel vibration information to the track irregularity recognition model after training to perform track irregularity recognition and obtain the real-time recognition result of track irregularity.

[0100] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0101] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0102] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A track irregularity identification method based on a discrete wavelet filter network, characterized by, include: An acceleration sensor is installed at a predetermined position on the vehicle body, and the front vertical acceleration signal and rear vertical acceleration signal generated by the sensor are acquired when the train is running on the track. Preprocessing is performed on all vertical acceleration signals, in which signal segments of a preset fixed time length are aligned and combined to obtain a sample matrix containing dual-channel vibration information; A preset historical sample matrix is ​​sent to a preset track irregularity recognition model for training to obtain a trained track irregularity recognition model. The preset historical sample matrix is ​​a sample matrix of historical data collected by the accelerometer after preprocessing. The sample matrix of the dual-channel vibration information is sent to the track irregularity recognition model after training to identify track irregularities and obtain the real-time recognition result of track irregularities.

2. The track irregularity identification method based on discrete wavelet filter network according to claim 1, characterized in that... The preprocessing based on all vertical acceleration signals includes aligning and combining signal segments of preset fixed time lengths, including: Based on the acquired front-end vertical acceleration signal, multiple continuous front-end vibration signal sample segments are extracted from the front-end signal sequence through a predetermined time window that slides with a fixed step size. Based on the acquired rear vertical acceleration signal and the front vibration signal sample segment, the delay caused by the train passing a fixed distance of its body at its running speed is calculated to determine the rear signal starting point corresponding to each front sample segment in the track space, and signals of the same time length are extracted to obtain the rear vibration signal sample segment corresponding to the front vibration signal sample segment. Based on the front-end vibration signal sample segment and the rear-end vibration signal sample segment, the front-end and rear-end signal sample segments representing the vibration of the same track segment are aligned on the time axis and spliced ​​in the channel dimension to construct the sample matrix containing dual-channel vibration information.

3. The track irregularity identification method based on discrete wavelet filter network according to claim 1, characterized in that... The preset historical sample matrix is ​​sent to the preset track irregularity recognition model for training, including: Based on the sample matrix, the signal of each channel is scaled and decomposed. By applying a learnable adjustment parameter to each coefficient sub-band obtained by decomposition, the adjusted coefficient sub-band is synthesized to obtain a reconstructed signal whose features are selectively enhanced in the time domain. The reconstructed signal is normalized based on its numerical distribution, and then the normalized signal is linearly combined and projected to obtain the track irregularity estimate corresponding to the end of the input signal time period.

4. The track irregularity identification method based on discrete wavelet filter network according to claim 3, characterized in that... The step of scaling the signal of each channel based on the sample matrix and synthesizing the adjusted coefficient subbands by applying learnable adjustment parameters to each coefficient subband obtained from the scaling includes: The sample matrix signal is separated at multiple scales by a preset learnable discrete wavelet filter network to obtain low-frequency approximation coefficients and high-frequency detail coefficients at each scale. By applying a learnable scalar weight to the low-frequency approximation coefficients and high-frequency detail coefficients at each scale, and performing a uniform scalar multiplication on all coefficients, the weighted low-frequency approximation coefficients and high-frequency detail coefficients are obtained. By recombining the weighted low-frequency approximation coefficients with the high-frequency detail coefficients through inverse discrete wavelet transform, a reconstructed signal with selectively enhanced features in the time domain is obtained.

5. The track irregularity identification method based on discrete wavelet filter network according to claim 3, characterized in that... The reconstructed signal is normalized based on its numerical distribution, and then the normalized signal is subjected to linear combination and projection of its features, including: The reconstructed signal is input into the normalization layer of a preset feedforward neural network for processing, wherein the normalization process is performed by a preset normalization formula to obtain the normalized reconstructed signal. The normalized reconstructed signal is input into the fully connected layer of the feedforward neural network for processing. The normalized reconstructed signal is flattened into a fixed-length vector, and the fixed-length vector is linearly transformed to obtain the track irregularity estimate corresponding to the end of all vertical acceleration signal time periods.

6. A track irregularity identification system based on a discrete wavelet filter network, characterized in that, include: The acquisition module is used to set up acceleration sensors at preset positions on the vehicle body and acquire the front vertical acceleration signal and rear vertical acceleration signal generated by the sensors when the train is running on the track. The processing module is used to preprocess all vertical acceleration signals, wherein a sample matrix containing dual-channel vibration information is obtained by aligning and combining signal segments of preset fixed time lengths. The training module is used to send a preset historical sample matrix to a preset track irregularity recognition model for training, and obtain a track irregularity recognition model after training. The preset historical sample matrix is ​​a sample matrix after preprocessing the historical data collected by the accelerometer. The identification module is used to send the sample matrix of the dual-channel vibration information to the track irregularity identification model after training to identify track irregularities and obtain the real-time identification result of track irregularities.

7. The track irregularity identification system based on discrete wavelet filter network according to claim 6, characterized in that, The processing module includes: The first processing unit is used to extract multiple continuous front-end vibration signal sample segments from the front-end signal sequence based on the acquired front-end vertical acceleration signal and through a predetermined time window that slides with a fixed step size. The second processing unit is used to determine the starting point of the rear signal corresponding to each front sample segment in the track space by calculating the delay caused by the train passing a fixed distance of its body at the running speed, based on the acquired rear vertical acceleration signal and the front vibration signal sample segment, and to extract the signal of the same time length to obtain the rear vibration signal sample segment corresponding to the front vibration signal sample segment. The third processing unit is used to align the front and rear signal sample segments representing the vibration of the same track segment on the time axis and splice them in the channel dimension according to the front-end vibration signal sample segment and the rear-end vibration signal sample segment to construct the sample matrix containing dual-channel vibration information.

8. The track irregularity identification system based on discrete wavelet filter network according to claim 6, characterized in that, The training module includes: The first training unit is used to perform scale decomposition on the signal of each channel in the sample matrix, and to synthesize the adjusted coefficient sub-bands by applying learnable adjustment parameters to each coefficient sub-band obtained by decomposition, so as to obtain a reconstructed signal with selectively enhanced features in the time domain. The second training unit is used to perform normalization processing based on the numerical distribution of the reconstructed signal, and to perform linear combination and projection of the normalized signal features to obtain the track irregularity estimate corresponding to the end of the input signal time period.

9. The track irregularity identification system based on discrete wavelet filter network according to claim 8, characterized in that, The first training unit includes: The first training subunit is used to separate each channel signal in the sample matrix at multiple scales through a preset learnable discrete wavelet filter network to obtain low-frequency approximation coefficients and high-frequency detail coefficients at each scale. The second training subunit is used to obtain the weighted low-frequency approximation coefficients and high-frequency detail coefficients by applying a learnable scalar weight to the low-frequency approximation coefficients and high-frequency detail coefficients at each scale, and by performing a uniform scalar multiplication on all coefficients. The third training subunit is used to re-merge the weighted low-frequency approximation coefficients and high-frequency detail coefficients through inverse discrete wavelet transform, so as to obtain a reconstructed signal with selectively enhanced features in the time domain.

10. The track irregularity identification system based on discrete wavelet filter network according to claim 8, characterized in that, The second training unit includes: The fourth training subunit is used to input the reconstructed signal into the normalization layer of a preset feedforward neural network for processing, wherein the normalized signal is obtained by normalization processing through a preset normalization formula. The fifth training subunit is used to input the normalized reconstructed signal into the fully connected layer of the feedforward neural network for processing. The normalized reconstructed signal is flattened into a fixed-length vector, and the fixed-length vector is linearly transformed to obtain the track irregularity estimate corresponding to the end of all vertical acceleration signal time periods.

Images