Steel rail corrugation recognition method and device based on neural network fusion model

Abstract

The invention provides a steel rail corrugation identification method and device based on a neural network fusion model, and relates to the technical field of railway facility detection, and the method comprises the steps: collecting the vertical acceleration data of an axle box of a train; discrete processing is carried out through Fourier transform, low-frequency signals are filtered out, and time domain axle box acceleration data after trend filtering out are reduced through inverse Fourier transform; decomposing the time domain axle box acceleration data through a variational mode decomposition processing method to obtain multi-resolution intrinsic mode function components, and selecting a plurality of components with the highest signal energy ratio as input samples of the model; synchronously inputting the input sample into a convolutional neural network and a bidirectional long-short-term memory network, capturing local features through convolution operation, performing bidirectional time sequence data dependency capturing on axle box acceleration data in combination with a bidirectional mechanism, and combining the convolutional local features with time sequence features through a feature fusion layer of a neural network fusion model to obtain a time sequence data dependency relationship; and obtaining an identification result of the steel rail corrugation section.

Description

Technical Field

[0001] This invention relates to the field of railway facility inspection technology, and more particularly to a method and device for identifying rail corrugation based on a neural network fusion model. Background Technology

[0002] This section is intended to provide background or context for embodiments of the present invention. The description herein is not intended to imply that it is prior art simply because it is included in this section.

[0003] Currently, the high-speed railway backbone network is basically completed, and the operating mileage of high-speed railways is increasing. Various track defects are occurring, among which wavy wear of rails (referred to as "rail corrugation") is a significant defect affecting track health. Rail corrugation refers to the wear on the rail working surface caused by prolonged poor wheel-rail contact under train loads, resulting in a wavy, uneven wear pattern with similar spacing along the longitudinal direction of the rail surface. Rail corrugation not only causes abnormal vibrations and noise during train operation but also deteriorates the performance of both vehicles and tracks.

[0004] Accurate identification of rail corrugation is a key concern in railway track maintenance. Several methods have been proposed to address this issue, including: 1. Combining variational mode decomposition and correlation coefficients to decompose and select the axle box acceleration IMF component with the highest correlation coefficient, and further applying a smoothed pseudo-Wigner-Willy distribution to calculate rail corrugation-related parameters for identification. 2. Decomposing the train axle box vibration acceleration signal into relatively stable IMFs using an adaptive noise empirical mode decomposition method, introducing MPE and correlation coefficients to clean low signal-to-noise ratio components, extracting IMF entropy features, and combining the energy values ​​of the remaining IMFs and the total energy value to determine rail corrugation through a threshold. 3. Studying the correlation between noise and corrugation by decomposing the axle box acceleration signal, and proposing a corrugation index for high-speed rail corrugation identification. 4. Processing train bogie acceleration data using the WT method to achieve corrugation detection. 5. Analyzing the amplitude, power spectral density, and frequency distribution range of axle box acceleration using Fourier transform to achieve rail corrugation detection. 6. Simulate the coupled operation of the train and rails to obtain simulated axle box acceleration data and analyze the spectrum of the corrugated section. However, the detection results obtained by the above method have a large error compared with the actual situation.

[0005] With the development of artificial intelligence, machine learning and deep learning methods have been gradually applied to the field of corrugation recognition. For example, 7. To address the problem of insufficient axle box acceleration data for intercity railway rail corrugation, ID-GAN was used to expand the corrugation dataset. 8. 1D-CNN was used for subway rail corrugation recognition; however, this method only considers local features of axle box acceleration and does not consider the successive dependency characteristics of axle box acceleration time-series data. 9. Wheel-rail vibration acceleration simulation data was generated through a coupled model, and bispectral features of the signal were extracted by ensemble empirical mode decomposition. Based on the bispectral features, support vector machine was used to identify the corrugation depth. 10. IDCNN was used to classify and identify the axle box acceleration of subway trains to determine the presence of rail corrugation. Simulation data was used to simulate axle box acceleration of fixed depth and length of corrugation, and KSM and PSO algorithms were used to identify the corrugation length and depth. The above research methods mainly rely on simulation data to achieve high recognition accuracy, and there is relatively little effective research based on measured axle box acceleration data during high-speed rail vehicle operation. Furthermore, wave-grinding identification methods based on signal processing require manual feature extraction and manual threshold setting, making the feature extraction process cumbersome.

[0006] In summary, there is an urgent need for a technical solution that can overcome the above-mentioned defects, improve the rail corrugation identification method, and enhance the identification accuracy and processing efficiency. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention proposes a method and device for identifying rail corrugation based on a neural network fusion model. This method can monitor and comprehensively detect the axle box vibration acceleration of trains in real time, and quickly and accurately identify corrugated sections of high-speed rail tracks.

[0008] In a first aspect of the present invention, a method for identifying rail corrugation based on a neural network fusion model is proposed, comprising:

[0009] Collect vertical acceleration data of the train's axle boxes;

[0010] The vertical vibration acceleration data of the axle box is discretized in the frequency domain by Fourier transform, low-frequency signals are filtered out, and the time-domain axle box acceleration data after filtering out the trend is restored by inverse Fourier transform.

[0011] The time-domain axle box acceleration data is decomposed using variational mode decomposition to obtain multi-resolution intrinsic mode function components. Several components with the highest signal energy proportion among these intrinsic mode function components are selected and used as input samples for a neural network fusion model. This neural network fusion model is composed of a convolutional neural network and a bidirectional long short-term memory network, and is trained using axle box acceleration data corresponding to corrugated and non-corrugated rail sections.

[0012] The input samples are synchronously input into a convolutional neural network and a bidirectional long short-term memory network. Local features are captured through convolution operations, and the bidirectional mechanism is used to capture the bidirectional temporal data dependency of the axle box acceleration data. The feature fusion layer of the neural network fusion model combines the convolutional local features output by the convolutional neural network with the temporal features output by the bidirectional long short-term memory network to obtain the identification result of the rail corrugation section.

[0013] In a second aspect of the present invention, a rail corrugation identification device based on a neural network fusion model is proposed, comprising:

[0014] The data acquisition module is used to collect the vertical acceleration data of the train's axle boxes;

[0015] The data processing module is used to discretize the vertical vibration acceleration data of the axle box in the frequency domain through Fourier transform, filter out low-frequency signals, and restore the time-domain axle box acceleration data after filtering out the trend using inverse Fourier transform.

[0016] The variational mode decomposition module is used to decompose the time-domain axle box acceleration data using the variational mode decomposition processing method to obtain multi-resolution intrinsic mode function components. Several components with the highest signal energy proportion among the intrinsic mode function components are selected, and these selected components are used as input samples for the neural network fusion model. The neural network fusion model is composed of a convolutional neural network and a bidirectional long short-term memory network, and is trained using axle box acceleration data corresponding to corrugated and non-corrugated rail sections.

[0017] The fusion recognition module is used to synchronously input the input samples into a convolutional neural network and a bidirectional long short-term memory network. It captures local features through convolution operations, and captures bidirectional temporal data dependencies of axle box acceleration data through a bidirectional mechanism. The feature fusion layer of the neural network fusion model combines the convolutional local features output by the convolutional neural network with the temporal features output by the bidirectional long short-term memory network to obtain the recognition result of the rail corrugation section.

[0018] In a third aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a rail corrugation identification method based on a neural network fusion model.

[0019] In a fourth aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements a rail corrugation identification method based on a neural network fusion model.

[0020] In a fifth aspect of the present invention, a computer program product is provided, the computer program product comprising a computer program that, when executed by a processor, implements a rail corrugation identification method based on a neural network fusion model.

[0021] Compared with existing technologies, the rail corrugation identification method and device based on a neural network fusion model proposed in this invention have at least the following technical advantages:

[0022] 1. By adopting the axle box acceleration numerical restoration method, the overall drift of acceleration data caused by factors such as acceleration sensors is eliminated, the trend of axle box acceleration data is filtered out, and the accuracy of axle box acceleration data is significantly improved.

[0023] 2. By using the VMD method to extract the vertical acceleration feature components of the axle box, the feature dimension of the one-dimensional axle box acceleration is effectively expanded, eliminating the influence of factors such as vertical excitation, environmental noise, and resonance, improving the signal-to-noise ratio of the axle box acceleration feature components, and providing high-quality, multi-resolution input features for the wave wear identification model.

[0024] 3. The end-to-end VMD-CNN-Bi-LSTM corrugation recognition method can learn the complex nonlinear relationship of axle box acceleration and has a strong representation learning ability. It breaks through the feature learning limitations of single CNN and Bi-LSTM models, and fully learns the signal characteristics, temporal characteristics and data characteristics of different spatial points of corrugation and normal axle box acceleration. It has excellent recognition ability for corrugated sections of high-speed rail rails, with high degree of automation, superior performance and strong generalization ability.

[0025] 4. This invention provides a new approach for identifying other types of track geometric shortwave defects. By deeply analyzing and comprehensively examining the measured high-speed rail corrugation and axle box acceleration in normal sections from actual train inspections, it mines the characteristics and spatiotemporal features of axle box acceleration signals and designs a highly efficient algorithm to quickly and accurately identify high-speed rail corrugation sections. In this innovative research, key parameters and thresholds have undergone in-depth experimental analysis and precise adjustment to ensure that the algorithm achieves optimal performance in various practical application scenarios. Based on the optimized parameters, the efficiency, accuracy, stability, and reliability of the algorithm have been significantly improved. This innovative method effectively improves the accuracy and efficiency of high-speed rail corrugation identification, simplifies the axle box acceleration feature mining process, provides strong technical support for railway track defect detection, and provides a scientific basis for track maintenance and repair. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a schematic flowchart of a rail corrugation identification method based on a neural network fusion model according to an embodiment of the present invention.

[0028] Figure 2 This is a schematic diagram of the overall process for identifying rail corrugation according to a specific embodiment of the present invention.

[0029] Figure 3 This is a diagram of the original axle box vertical acceleration according to a specific embodiment of the present invention.

[0030] Figure 4 This is a vertical acceleration diagram of the axle box after trend filtering according to a specific embodiment of the present invention.

[0031] Figure 5 This is a VMD decomposition diagram of the vertical acceleration of the axle box in a normal section according to a specific embodiment of the present invention.

[0032] Figure 6 This is a characteristic component diagram of the vertical acceleration IMF of the axle box in the corrugated section according to a specific embodiment of the present invention.

[0033] Figure 7 This is a characteristic component diagram of the vertical acceleration IMF of the axle box in the corrugated section according to a specific embodiment of the present invention.

[0034] Figure 8 This is a characteristic component diagram of the vertical acceleration IMF of the axle box in the normal section according to a specific embodiment of the present invention.

[0035] Figure 9 This is a schematic diagram of the relationship between the VMD-CNN-Bi-LSTM wave-moistening recognition model according to a specific embodiment of the present invention.

[0036] Figure 10 This is a confusion matrix diagram of the VMD-CNN-Bi-LSTM wave polishing recognition model in a specific embodiment of the present invention on the test set.

[0037] Figure 11 This is a schematic diagram illustrating the test results of VMD-CNN-Bi-LSTM, Support Vector Machine, and Gaussian Naive Bayes according to a specific embodiment of the present invention.

[0038] Figure 12 This is a schematic diagram of the architecture of a rail corrugation identification device based on a neural network fusion model according to an embodiment of the present invention.

[0039] Figure 13 This is a schematic diagram of a computer device structure according to an embodiment of the present invention. Detailed Implementation

[0040] The principles and spirit of the invention will now be described with reference to several exemplary embodiments. It should be understood that these embodiments are given merely to enable those skilled in the art to better understand and implement the invention, and are not intended to limit the scope of the invention in any way. Rather, these embodiments are provided to make this disclosure more thorough and complete, and to fully convey the scope of this disclosure to those skilled in the art.

[0041] Those skilled in the art will recognize that embodiments of the present invention can be implemented as a system, apparatus, device, method, or computer program product. Therefore, this disclosure can be specifically implemented in the following forms: entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0042] According to an embodiment of the present invention, a method and device for identifying rail corrugation based on a neural network fusion model are proposed, which relates to the field of railway facility detection technology.

[0043] The principles and spirit of the present invention will be explained in detail below with reference to several representative embodiments.

[0044] In the embodiments of the present invention, the following terms need to be explained:

[0045] VMD: Variational Mode Decomposition. IMF: Intrinsic Mode Function. CNN: Convolutional Neural Network. Bi-LSTM: Bidirectional Long Short-Term Memory Network. LSTM: Long Short-Term Memory Network. RNN: Recurrent Neural Network.

[0046] Figure 1 This is a schematic flowchart of a rail corrugation identification method based on a neural network fusion model according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0047] S101, collects vertical acceleration data of the train's axle boxes;

[0048] S102, the vertical vibration acceleration data of the axle box is discretized in the frequency domain by Fourier transform, low-frequency signals are filtered out, and the time-domain axle box acceleration data after filtering out the trend is restored by inverse Fourier transform.

[0049] S103, the time-domain axle box acceleration data is decomposed using variational mode decomposition to obtain multi-resolution intrinsic mode function components. Several components with the highest signal energy proportion among the intrinsic mode function components are selected, and these selected components are used as input samples for the neural network fusion model. The neural network fusion model is composed of a convolutional neural network and a bidirectional long short-term memory network, and is trained using axle box acceleration data corresponding to corrugated and non-corrugated rail sections.

[0050] S104, the input samples are synchronously input into the convolutional neural network and the bidirectional long short-term memory network. Local features are captured through convolution operations. The bidirectional mechanism is combined to capture the bidirectional temporal data dependency of the axle box acceleration data. The feature fusion layer of the neural network fusion model combines the convolutional local features output by the convolutional neural network with the temporal features output by the bidirectional long short-term memory network to obtain the identification result of the rail corrugation section.

[0051] To provide a clearer explanation of the rail corrugation identification method based on the neural network fusion model, each step will be described in detail below.

[0052] In one embodiment, for S101, the vertical acceleration data of the train's axle boxes is collected, and the specific process is as follows:

[0053] A sliding window is used to retrieve vertical vibration acceleration data of the train's axle boxes. Each preset number of axle box vertical vibration acceleration data constitutes a sample data.

[0054] In one embodiment, for S102, the vertical vibration acceleration data of the axle box is discretized in the frequency domain by Fourier transform, low-frequency signals are filtered out, and the time-domain axle box acceleration data after filtering out the trend is restored by inverse Fourier transform.

[0055] The Fourier transform converts the time-domain signal to the frequency domain, and the transformation formula is as follows:

[0056]

[0057] In the formula, X[k] is a discrete sample of the frequency domain signal, representing the complex amplitude of the k-th frequency component; x[n] is a discrete sample of the original time domain signal; N is the number of sample points, representing the length of the signal; k is a discrete variable in the frequency domain; j is the imaginary unit. It is a complex exponential function; n is a discrete variable in the time domain;

[0058] The inverse Fourier transform is used to restore the time-domain axle box acceleration data after filtering out the trend. The transformation formula is as follows:

[0059]

[0060] In the formula, x[n] is a discrete sample of the reconstructed time-domain signal; X[k] is a discrete sample of the frequency-domain signal; k and n are discrete variables in the frequency domain and time domain, respectively; j is the imaginary unit; and N is the number of sample points.

[0061] In one embodiment, for S103, the time-domain axle box acceleration data is decomposed using a variational mode decomposition method to obtain multi-resolution intrinsic mode function components. Several components with the highest signal energy proportion among the intrinsic mode function components are selected, and these selected components are used as input samples for the neural network fusion model. The specific process is as follows:

[0062] Construct a variational model, let the original time series signal be f(t), and use the Hilbert transform to solve for the K intrinsic mode function components. The calculation formula is as follows:

[0063]

[0064] In the formula, u k It is the k-th intrinsic mode function component; t represents time; w k It is the frequency center of the k-th mode, and K is the total number of modes; represents the derivative of the function with respect to time t; ||·||2 represents the L2 norm; i is the imaginary unit; st represents the constraint condition; f(t) represents the original time series signal;

[0065] To solve the variational model, we transform it using the augmented Lagrange function, as shown in the following expression:

[0066]

[0067] In the formula, L represents the augmented Lagrange function; α is the quadratic penalty factor; and λ is the Lagrange multiplier.

[0068] Based on the variational model, the time-domain axle box acceleration data is decomposed using the variational mode decomposition method to obtain multi-order intrinsic mode function components.

[0069] Furthermore, for the aforementioned multi-order intrinsic mode function components, the corresponding energy percentage is calculated; wherein, the energy of the intrinsic mode function components is calculated using the signal energy formula, the calculation formula is as follows:

[0070]

[0071] In the formula, E i S represents the energy of the i-th intrinsic mode function component; j is the j-th sample value in the intrinsic mode function component; N′ is the number of samples in the intrinsic mode function component; P iE represents the proportion of the energy of the i-th intrinsic mode function component to the total energy; j Let M represent the energy of the j-th intrinsic mode function component, and M be the total number of intrinsic mode function components.

[0072] Variational mode decomposition was performed on multiple sets of axle box acceleration data that were both corrugated and non-corrugated. The data were sorted according to their energy proportions, and the components with the highest signal energy proportions were selected as input samples for the neural network fusion model.

[0073] The neural network fusion model is composed of a convolutional neural network and a bidirectional long short-term memory network, and is trained using axle box acceleration data corresponding to the corrugated and non-corrugated sections of the rail.

[0074] In one embodiment, for S104, the input samples are simultaneously input into a convolutional neural network and a bidirectional long short-term memory network. Local features are captured through convolution operations, and the bidirectional mechanism is used to capture the bidirectional temporal data dependency of the axle box acceleration data. The feature fusion layer of the neural network fusion model combines the convolutional local features output by the convolutional neural network with the temporal features output by the bidirectional long short-term memory network to obtain the identification result of the rail corrugation section. The specific process is as follows:

[0075] The selected components are simultaneously input into a convolutional neural network and a bidirectional long short-term memory network. The convolutional neural network uses four layers of convolution to capture the intrinsic mode function component features across spatiotemporal points. The bidirectional long short-term memory network extracts the forward and backward temporal dependency features of the intrinsic mode function components of the axle box acceleration through a bidirectional mechanism.

[0076] A fusion processing layer of convolutional neural network and bidirectional long short-term memory network is constructed. The output features of the convolutional neural network and bidirectional long short-term memory network are fused together. Four fully connected layers are superimposed to perform feature fusion and dimensionality reduction calculation. Spatiotemporal features under different feature frequency patterns are extracted. The output layer uses the Sigmoid function to return the data category probability to identify the rail corrugated section and obtain the rail corrugated section and the rail non-corrugated section.

[0077] For the neural network fusion model, the fusion method is as follows:

[0078] The high-dimensional features obtained from the convolutional neural network and the bidirectional long short-term memory network are flattened into one-dimensional features. The output feature dimensions of the convolutional neural network and the bidirectional long short-term memory network are aligned, and the output features of the convolutional neural network and the bidirectional long short-term memory network are concatenated in the vertical dimension to form multi-dimensional features. Four fully connected layers are used for fusion calculation and feature dimensionality reduction. The hyperbolic tangent activation function is selected as the activation function for each fully connected layer, and the sigmoid function is selected in the output layer to make a probability judgment on the data category to obtain the recognition result.

[0079] Specifically, the convolutional neural network consists of multiple stacked convolutional layers, pooling layers, fully connected layers, and an output layer. The convolutional layers extract multi-class feature information from the original data through convolutional kernels and perform dimensionality reduction processing on the original high-dimensional input data. The calculation formula is as follows:

[0080]

[0081] RELU(x) = max(0,x); (9)

[0082] pooling(x) = max(x); (10)

[0083] In the formula, It is the calculated value of the i-th neuron in the k-th layer of the convolutional layer; This is the output of neuron j in the local region of the k-1th convolutional layer above; The parameters are the kernel parameters of the k-th layer; For bias terms; M i f(x) is the output of the convolution in the (k-1)th layer above; f(x) is the non-linear activation function, using the ReLU function; x represents the neuron output; pooling is the pooling operation, and pooling(x) = max(x) is the max pooling operation.

[0084] Specifically, the bidirectional long short-term memory network adds bidirectionality to the long short-term memory network; the long short-term memory network is based on a recurrent neural network, and by adding forget gates, input gates, and output gates, it establishes dependencies between time-series data, where the state information of one unit is:

[0085] f t =σ(W f ·[h t-1 ,x t ]+b f (11)

[0086] i t =σ(W i ·[h t-1 ,x t ]+b i (12)

[0087] g t =Tanh(W g ·[h t-1 ,x t ]+b g (13)

[0088] o t =σ(W o ·[h t-1 ,xt ]+b o (14)

[0089] C t =f t *C t-1 +i t *g t (15)

[0090] h t =Tanh(C t )*o t (16)

[0091] In the formula, f t It is the output of the forget gate at time step t; σ is the Sigmoid activation function; W f h is the state transition matrix for the forget gate; t-1 It is the hidden state of the previous time step; x t b is the input for the current time step; f It is the bias term of the forget gate; i t It is the input gate output of time step t; W i b is the state transition matrix of the input gate; i It is the bias term of the input gate; g t Candidates at time step t are candidate memory units; tanh is the hyperbolic tangent activation function; W g It is the weight matrix of the candidate memory units; b g It is the bias term for candidate memory units; o t It is the output gate output at time step t; W o b is the state transition matrix of the output gate; o It is the bias term of the output gate; C t It is the cell state at time step t; C t-1 It is the cell state of the time step preceding time step t; h t It is the hidden state of time step t.

[0092] It should be noted that although the operation of the method of the present invention has been described in a specific order in the above embodiments and figures, this does not require or imply that the operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0093] This invention uses the measured and analyzed vertical vibration acceleration of the axle box, and leverages VMD-CNN-Bi-LSTM feature self-learning to intelligently identify corrugation, referencing... Figure 2 The specific process is as follows:

[0094] 210, Vertical vibration acceleration data of the axle box;

[0095] A sliding window was used to retrieve data from the vertical axle box vibration acceleration data of the comprehensive test train, with each 1000 axle box acceleration data points constituting one sample.

[0096] 220, Axle box acceleration trend filtering based on Fourier transform;

[0097] After obtaining the discrete representation of the axle box acceleration data in the frequency domain through Fast Fourier Transform, the low-frequency signal is filtered out, and the time-domain axle box acceleration data after filtering out the trend is restored by Inverse Fourier Transform.

[0098] 230, Axle box acceleration decomposition based on VMD, forming multi-resolution IMF components;

[0099] Multi-resolution IMF components are obtained by decomposing the acceleration signal using VMD, and the IMF characteristic components of corrugated and normal rails are extracted by combining signal energy.

[0100] 240. Calculate the signal energy of the IMF components of the axle box acceleration of each order, and select the top 4 components with the largest energy; form the [4,1000]-dimensional IMF result components as the input of the CNN-Bi-LSTM model.

[0101] 250, input into CNN, mine convolutional features at different spatial locations;

[0102] 250', input Bi-LSTM, to mine comprehensive temporal features; 250 and 250' can be performed simultaneously.

[0103] 260, Input to CNN-Bi-LSTM feature fusion layer;

[0104] 270, Identify corrugated sections and normal sections (non-corrugated sections).

[0105] The IMF is simultaneously input into CNN and Bi-LSTM. Local features are captured through convolution operations, and the Bi-LSTM is used to capture bidirectional temporal data dependencies of the axle box acceleration data.

[0106] This method combines the convolutional local features output by CNN with the temporal features of Bi-LSTM using a CNN-Bi-LSTM feature fusion layer to achieve high-speed railway rail corrugation identification. Experiments were conducted using a large amount of vertical vibration acceleration data of axle boxes collected from a high-speed railway, and the results were compared with mainstream classification models. The experimental results show that the proposed model achieves a corrugation identification accuracy of 98.4% and a recall rate of 96.9%, comprehensively outperforming classifiers such as Support Vector Machine and Gaussian Naive Bayes. The model demonstrates significant classification performance, strong robustness, and simple and efficient feature extraction. This method can comprehensively monitor and detect train axle box vibration acceleration in real time, quickly and easily learn the axle box acceleration features of corrugated and non-corrugated sections, and promptly and effectively identify rail corrugated sections, providing a scientific basis for railway track maintenance.

[0107] The rail corrugation identification method based on a neural network fusion model of the present invention will be described below with reference to an exemplary scenario.

[0108] 1. Collect vertical vibration acceleration data of the axle box:

[0109] To expand the dataset and unify the data dimensions of the model input, a sliding window was used to slide and extract data on the corrugated axle box acceleration data and the normal axle box acceleration data. Every 1,000 axle box acceleration data constituted one sample, resulting in a total of 13,102 samples for training and testing.

[0110] 2. Filter out the vertical vibration acceleration trend term of the axle box:

[0111] For 13,102 samples, the total number of samples for the acceleration of the grinding wheel box is 7,092, of which 6,001 are used for training and 1,091 are used for testing.

[0112] The total number of samples for (non-wave-patterned) normal axle box acceleration is 6010, of which 5009 are used for training and 1000 are used for testing.

[0113] To reduce the overall drift of acceleration data caused by factors such as acceleration sensors, trend filtering is performed using fast Fourier transform and inverse Fourier transform to restore the vertical vibration acceleration data of the axle box.

[0114] The Fast Fourier Transform converts a time-domain signal to the frequency domain, and the transformation formula (1) is as follows:

[0115]

[0116] In the formula, X[k] is a discrete sample of the frequency domain signal, representing the complex amplitude of the k-th frequency component; x[n] is a discrete sample of the original time domain signal; N is the number of sample points, representing the length of the signal; k is a discrete variable in the frequency domain; j is the imaginary unit. It is a complex exponential function; n is a discrete variable in the time domain.

[0117] The inverse Fourier transform converts the frequency domain signal back to the time domain, as shown in formula (2):

[0118]

[0119] In the formula, x[n] is a discrete sample of the reconstructed time-domain signal; X[k] is a discrete sample of the frequency-domain signal; k and n are discrete variables in the frequency domain and time domain, respectively; j is the imaginary unit; and N is the number of sample points.

[0120] After obtaining the discrete representation of the axle box acceleration data in the frequency domain through Fast Fourier Transform, low-frequency signals are filtered out, and the time-domain axle box acceleration data after filtering out the trend is restored using Inverse Fourier Transform. For example... Figure 3 The diagram shown is a diagram of the original axle box vertical acceleration according to a specific embodiment of the present invention; as follows: Figure 4 The figure shown is a diagram of the vertical acceleration of the axle box after trend filtering according to a specific embodiment of the present invention. The vibration acceleration of the axle box after trend filtering oscillates around 0g.

[0121] 3. Extraction of vertical vibration acceleration feature components of axle box based on VMD:

[0122] Variational Mode Decomposition (VMD) is an adaptive signal decomposition method that effectively handles the nonlinearity and non-stationarity of axle box acceleration data, providing a more stable data foundation for the training of subsequent CNN-Bi-LSTM models. VMD decomposes axle box acceleration data into a series of Intrinsic Mode Functions (IMFs) with specific center frequencies and frequency bands. Iterative updates of the IMFs are achieved using the alternating direction multiplier method, adaptively searching for the optimal center frequency and frequency band of the IMFs. The core of VMD lies in constructing a variational model and solving the variational problem.

[0123] A variational model is constructed, using the sum of the IMFs as the input signal and minimizing the sum of the frequency bands of the IMFs as the objective function. Let the original time series signal be f(t), and the Hilbert transform is used to solve for the K IMFs. The calculation formula is as follows:

[0124]

[0125] In the formula, u k It is the k-th intrinsic mode function component; t represents time; w k It is the frequency center of the k-th mode, and K is the total number of modes; represents the derivative of the function with respect to time t; ||·||2 represents the L2 norm; i is the imaginary unit; st represents the constraint condition; f(t) represents the original time series signal;

[0126] To solve the variational model, the augmented Lagrange function can be used to transform the above model into an equivalent unconstrained optimization problem, as shown in the following expression:

[0127]

[0128] In the formula, L represents the augmented Lagrange function; α is the quadratic penalty factor; and λ is the Lagrange multiplier.

[0129] Substituting the axle box acceleration into VMD decomposition to obtain a stationary characteristic sequence, the first 6 IMF components of the axle box acceleration in the corrugated and normal sections are as follows: Figure 5 and Figure 6 As shown, where, Figure 5 This is a VMD decomposition diagram of the vertical acceleration of the axle box in a normal section according to a specific embodiment of the present invention. Figure 6 This is a characteristic component diagram of the vertical acceleration IMF of the axle box in the corrugated section according to a specific embodiment of the present invention.

[0130] Based on the variational model, the time-domain axle box acceleration data is decomposed using the variational mode decomposition method to obtain multi-order intrinsic mode function components.

[0131] Furthermore, the energy proportion of each order of IMF components after VMD decomposition is calculated, and the energy of the IMF is calculated using the signal energy formula. The calculation formula is as follows:

[0132]

[0133] In the formula, E i S represents the energy of the i-th intrinsic mode function component; j is the j-th sample value in the intrinsic mode function component; N′ is the number of samples in the intrinsic mode function component; P i E represents the proportion of the energy of the i-th intrinsic mode function component to the total energy; j Let M represent the energy of the j-th intrinsic mode function component, and M be the total number of intrinsic mode function components.

[0134] After performing VMD decomposition on the axle box accelerations of multiple sets of corrugated and non-corrugated axle boxes, the IMF energy proportions were calculated and sorted according to equations (6) and (7), and the top four IMF component energy proportions were selected. Figure 7 , Figure 8 As shown, these are the four high-energy IMF components of the corrugated / normal axle box acceleration after IMF energy sorting; among them, Figure 7 This is a characteristic component diagram of the vertical acceleration IMF of the axle box in the corrugated section according to a specific embodiment of the present invention. Figure 8 This is a characteristic component diagram of the vertical acceleration IMF of the axle box in the normal section according to a specific embodiment of the present invention.

[0135] 4. Identification of rail corrugation sections based on VMD-CNN-Bi-LSTM model:

[0136] Combining VMD, convolutional neural networks (CNN), and bidirectional long short-term memory networks (Bi-LSTM), a method for identifying rail corrugation in high-speed railways is designed, such as... Figure 9 The diagram illustrates the relationship between the VMD-CNN-Bi-LSTM wave-grinding recognition model according to a specific embodiment of the present invention. When processing time-series data, one-dimensional convolutional neural networks are often used for feature extraction and pattern recognition. However, one-dimensional data mining has limited information and feature dimensions. VMD decomposes axle box acceleration into intrinsic mode components (IMFs) at different feature frequencies, and selects the top four IMFs based on signal energy. These components are then input into CNN and Bi-LSTM, respectively. The CNN model uses four convolutional layers to capture IMF component features across spatiotemporal points; the Bi-LSTM extracts the forward and inverse temporal dependencies of the axle box acceleration IMF components through a bidirectional mechanism. Finally, a CNN-Bi-LSTM network fusion processing layer is constructed, fusing the output features of the CNN and Bi-LSTM neural networks. Four fully connected layers are then superimposed for feature fusion and dimensionality reduction calculations, extracting spatiotemporal features under different feature frequency modes. The output layer is a binary classification problem, using the Sigmoid function to return the data category probability, ultimately achieving wave-grinding recognition.

[0137] In one embodiment, a convolutional neural network (CNN) consists of multiple stacked convolutional layers, pooling layers, fully connected layers, and an output layer. The convolutional layers extract multi-class feature information from the original data through convolutional kernels and perform dimensionality reduction processing on the original high-dimensional input data, as shown in the following formula:

[0138]

[0139] RELU(x) = max(0,x); (9)

[0140] pooling(x) = max(x); (10)

[0141] In the formula, It is the calculated value of the i-th neuron in the k-th layer of the convolutional layer; This is the output of neuron j in the local region of the k-1th convolutional layer above; The parameters are the kernel parameters of the k-th layer; For bias terms; M iThe output of the (k-1)th convolutional layer is denoted by f(x); f(x) is a non-linear activation function, specifically the ReLU function, which enables the neural network to learn more complex patterns (Equation 9); x represents the neuron output; pooling is the pooling operation, and pooling(x) = max(x) is the max pooling operation. Pooling layers help extract higher-level feature representations. While ensuring local linear invariance, they can significantly reduce the spatial dimension of the data, allowing the neural network to focus more on learning more abstract features and improving the model's generalization ability. This model uses a max pooling layer to learn the most salient local feature values ​​(Equation 10).

[0142] VMD decomposes time-series data into multiple intrinsic IMFs, each reflecting different frequency components, thus addressing the issues of nonlinearity and nonstationarity of the data. This model decomposes one-dimensional box acceleration data into [4, 1000]-dimensional IMF feature vectors. To more comprehensively capture the box acceleration features at different spatial locations, this model further transforms the [4, 1000]-dimensional IMF components into [4, 25, 40]-dimensional features. These IMFs are then used for high-dimensional feature extraction via CNN.

[0143] The CNN framework proposed in this paper is as follows: Figure 9 As shown in the upper middle part, this model uses four convolutional layers, employs the ReLU activation function, and uses max pooling for all pooling layers. After four convolutional layers and ReLU activation, the 4×1000-dimensional IMF feature vector is converted into a set of feature maps, which are then sent to the max pooling layer for downsampling. After downsampling, the feature maps are unfolded to form a high-dimensional convolutional feature map of [250, 1, 1]. This model mainly uses 3×3 small convolutional kernels. 3×3 convolutional kernels have fewer parameters, which helps reduce the complexity and computational cost of the model, reduces the risk of overfitting, and has good adaptability to various tasks. They are suitable for expanding the depth and width of neural networks. In addition, Google's Inception network achieves excellent recognition accuracy by extracting features of different scales in parallel using 3×3 convolutional kernels, proving the importance of 3×3 convolutional kernels in feature extraction. Therefore, this study mainly uses 3×3 small convolutional kernels.

[0144] In one embodiment, for a bidirectional long short-term memory network (Bi-LSTM), in order to capture the temporal dependencies of axle box acceleration data, Bi-LSTM is combined to capture the bidirectional temporal data dependency features of the 4-dimensional IMF temporal components of axle box acceleration.

[0145] Specifically, LSTM neural networks are an improvement on recurrent neural networks (RNNs). By adding forget gates, input gates, and output gates, important information can be remembered while unimportant information can be selectively forgotten, thus establishing long-distance dependencies between time-series data.

[0146] The state information of a cell in an LSTM network is as follows:

[0147] f t =σ(W f ·[h t-1 ,x t ]+b f (11)

[0148] i t =σ(W i ·[h t-1 ,x t ]+b i (12)

[0149] g t =Tanh(W g ·[h t-1 ,x t ]+b g (13)

[0150] o t =σ(W o ·[h t-1 ,x t ]+b o (14)

[0151] C t =f t *C t-1 +i t *g t (15)

[0152] h t =Tanh(C t )*o t (16)

[0153] In the formula, formula (11) is the forgetting gate, f t It is the output of the forget gate at time step t; σ is the Sigmoid activation function; W f h is the state transition matrix for the forget gate; t-1 It is the hidden state of the previous time step; x t b is the input for the current time step; f It is the bias term of the forget gate;

[0154] Equation (12) is the input gate, i t It is the input gate output of time step t; W i b is the state transition matrix of the input gate; i It is the bias term of the input gate;

[0155] Equation (13) is the candidate memory unit, g tCandidates at time step t are candidate memory units; tanh is the hyperbolic tangent activation function; W g It is the weight matrix of the candidate memory units; b g It is the bias term for candidate memory units;

[0156] Equation (14) is the output gate, o t It is the output gate output at time step t; W o b is the state transition matrix of the output gate; o It is the bias term of the output gate;

[0157] Equation (15) represents the updated cell state, determined by both the forget gate and the input gate, C t It is the cell state at time step t; C t-1 It is the cell state of the time step preceding time step t;

[0158] Equation (16) represents the final hidden state, which is determined by the output gate and the updated cell state. t It is the hidden state of time step t.

[0159] Bi-LSTM adds bidirectionality to LSTM, enabling it to consider both forward and backward information when processing sequence data. Therefore, this invention employs a bidirectional LSTM for bidirectional temporal feature capture. This model consists of a 4-layer Bi-LSTM structure, as shown below. Figure 9 In the lower middle part, the hidden layer dimension is set to 20. Since each node of the Bi-LSTM considers the features of the forward information in its computation, in order to capture the full temporal result features and reduce computational cost, this model selects the output features of the last node to form the overall temporal features of the IMF. Because a bidirectional LSTM model is used, the final output is a temporal feature of [2 × hidden layer dimension = 40] dimensions.

[0160] In one embodiment, for the CNN-Bi-LSTM feature fusion layer, the high-dimensional features obtained from CNN and Bi-LSTM are flattened into one-dimensional features. The output feature dimensions of CNN and Bi-LSTM are aligned, and the output features of CNN and Bi-LSTM are concatenated vertically to form 280-dimensional features. Four fully connected layers are used for fusion calculation and feature dimensionality reduction. Each fully connected layer selects tanh as the activation function, and the output layer uses the sigmoid function to make probabilistic judgments on the data categories, ultimately identifying whether there is ripple in the axle box acceleration. The collaborative reuse of the CNN and Bi-LSTM models not only captures the temporal characteristics of the axle box acceleration IMF components but also considers the data characteristics of the axle box acceleration IMF components across spatial points.

[0161] Three metrics were used to evaluate the VMD-CNN-Bi-LSTM wave damage classification model: accuracy, recall, and precision. Figure 10 This is a confusion matrix diagram of the VMD-CNN-Bi-LSTM wave polishing recognition model of a specific embodiment of the present invention on the test set. The experimental results are as follows: Figure 10 As shown in Table 1, the confusion matrix indicates that the model's accuracy, recall, and precision are 98.4%, 96.9%, and 100%, respectively. The evaluation metrics demonstrate that this model, based on axle box acceleration data, effectively identifies high-speed rail rail corrugation, exhibiting significant classification performance. It comprehensively identifies corrugation with a low false negative rate and strong robustness, making it suitable for real-world high-speed rail corrugation detection scenarios.

[0162] Table 1 Experimental Results of VMD-CNN-Bi-LSTM Wave Polishing Recognition Model

[0163] Evaluation indicators accuracy Recall rate Accuracy result 98.4％ 96.9％ 100％

[0164] This model is compared with two other classification models: Support Vector Machine (SVM) and Gaussian Naive Bayes (GaussianNB). SVM and Gaussian Naive Bayes were trained and tested using the same axle box acceleration dataset. The experimental results are as follows: Figure 11 As shown, Figure 11 This is a schematic diagram illustrating the test results of VMD-CNN-Bi-LSTM, Support Vector Machine, and Gaussian Naive Bayes according to a specific embodiment of the present invention.

[0165] The Support Vector Machine (SVM) model achieved the worst performance among the three models in terms of wave wear recognition accuracy (66%), recall (66%), and precision (69%), making it unsuitable for wave wear classification in axle box acceleration data with non-linear relationships. For data with non-linear relationships, SVM relies heavily on forward feature extraction. Further mining the feature rules of wave wear / normal axle box acceleration and then combining it with SVM for classification may yield better results. The Gaussian Naive Bayes model achieved an accuracy of 87%, recall of 86%, and precision of 90% in wave wear recognition, performing well in wave wear recognition scenarios and outperforming SVM. It can self-learn some feature rules from axle box acceleration data; however, it is still weaker than the VMD-CNN-Bi-LSTM model.

[0166] The VMD-CNN-Bi-LSTM proposed in this invention outperforms support vector machines and Gaussian Naive Bayes models in terms of accuracy, recall, and precision in corrugation recognition. It can learn complex nonlinear relationships and has strong representation learning capabilities. It fully learns the signal features, temporal features, and data features of different spatial points of corrugation and normal axle box acceleration, and has a strong recognition effect in high-speed rail corrugation scenarios.

[0167] After introducing the method of exemplary embodiments of the present invention, the following references are made. Figure 12 This invention provides an exemplary embodiment of a rail corrugation identification device based on a neural network fusion model.

[0168] The implementation of the rail corrugation identification device based on the neural network fusion model can refer to the implementation of the above method, and the repeated parts will not be described again. The term "module" or "unit" used below can refer to a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0169] Based on the same inventive concept, this invention also proposes a rail corrugation identification device based on a neural network fusion model, such as... Figure 12 As shown, the device includes:

[0170] The data acquisition module 1210 is used to collect the vertical acceleration data of the axle boxes of the train;

[0171] The data processing module 1220 is used to discretize the vertical vibration acceleration data of the axle box in the frequency domain through Fourier transform, filter out low-frequency signals, and restore the time-domain axle box acceleration data after filtering out the trend using inverse Fourier transform.

[0172] The variational mode decomposition module 1230 is used to decompose the time-domain axle box acceleration data through the variational mode decomposition processing method to obtain multi-resolution intrinsic mode function components. It selects several components with the highest signal energy proportion among the intrinsic mode function components and uses these selected components as input samples for a neural network fusion model. The neural network fusion model is composed of a convolutional neural network and a bidirectional long short-term memory network, and is trained using axle box acceleration data corresponding to corrugated and non-corrugated rail sections.

[0173] The fusion recognition module 1240 is used to synchronously input the input samples into a convolutional neural network and a bidirectional long short-term memory network. It captures local features through convolution operations, and captures bidirectional temporal data dependencies of axle box acceleration data through a bidirectional mechanism. The feature fusion layer of the neural network fusion model combines the convolutional local features output by the convolutional neural network with the temporal features output by the bidirectional long short-term memory network to obtain the recognition result of the rail corrugation section.

[0174] In one embodiment, the data acquisition module 1210 acquires vertical acceleration data of the train's axle boxes, including:

[0175] A sliding window is used to retrieve vertical vibration acceleration data of the train's axle boxes. Each preset number of axle box vertical vibration acceleration data constitutes a sample data.

[0176] In one embodiment, the data processing module 1220 performs discrete processing on the vertical vibration acceleration data of the axle box in the frequency domain using Fourier transform, filters out low-frequency signals, and restores the time-domain axle box acceleration data after filtering using inverse Fourier transform, including:

[0177] The Fourier transform converts a time-domain signal to the frequency domain. The transformation formula is as follows:

[0178]

[0179] In the formula, X[k] is a discrete sample of the frequency domain signal, representing the complex amplitude of the k-th frequency component; x[n] is a discrete sample of the original time domain signal; N is the number of sample points, representing the length of the signal; k is a discrete variable in the frequency domain; j is the imaginary unit. It is a complex exponential function; n is a discrete variable in the time domain;

[0180] The inverse Fourier transform is used to restore the time-domain axle box acceleration data after filtering out the trend. The transformation formula is as follows:

[0181]

[0182] In the formula, x[n] is a discrete sample of the reconstructed time-domain signal; X[k] is a discrete sample of the frequency-domain signal; k and n are discrete variables in the frequency domain and time domain, respectively; j is the imaginary unit; and N is the number of sample points.

[0183] In one embodiment, the variational mode decomposition module 1230 decomposes the time-domain axle box acceleration data using a variational mode decomposition processing method to obtain multi-resolution intrinsic mode function components. It then selects several components with the highest signal energy proportion among the intrinsic mode function components and uses these selected components as input samples for a neural network fusion model, including:

[0184] Construct a variational model, let the original time series signal be f(t), and use the Hilbert transform to solve for the K intrinsic mode function components. The calculation formula is as follows:

[0185]

[0186] In the formula, u k It is the k-th intrinsic mode function component; t represents time; w k It is the frequency center of the k-th mode, and K is the total number of modes; represents the derivative of the function with respect to time t; ||·||2 represents the L2 norm; i is the imaginary unit; st represents the constraint condition; f(t) represents the original time series signal;

[0187] To solve the variational model, we transform it using the augmented Lagrange function, as shown in the following expression:

[0188]

[0189] In the formula, L represents the augmented Lagrange function; α is the quadratic penalty factor; and λ is the Lagrange multiplier.

[0190] Based on the variational model, the time-domain axle box acceleration data is decomposed using the variational mode decomposition method to obtain multi-order intrinsic mode function components.

[0191] In one embodiment, the variational mode decomposition module 1230 decomposes the time-domain axle box acceleration data using a variational mode decomposition processing method to obtain multi-resolution intrinsic mode function components. It then selects several components with the highest signal energy proportion among the intrinsic mode function components and uses these selected components as input samples for a neural network fusion model, including:

[0192] For the aforementioned multi-order intrinsic mode function components, the corresponding energy percentage is calculated; wherein, the energy of the intrinsic mode function components is calculated using the signal energy formula, the calculation formula is as follows:

[0193]

[0194] In the formula, E i S represents the energy of the i-th intrinsic mode function component; j is the j-th sample value in the intrinsic mode function component; N′ is the number of samples in the intrinsic mode function component; P i E represents the proportion of the energy of the i-th intrinsic mode function component to the total energy; j Let M represent the energy of the j-th intrinsic mode function component, and M be the total number of intrinsic mode function components.

[0195] Variational mode decomposition was performed on multiple sets of axle box acceleration data that were both corrugated and non-corrugated. The data were sorted according to their energy proportions, and the components with the highest signal energy proportions were selected as input samples for the neural network fusion model.

[0196] In one embodiment, the fusion recognition module 1240 synchronously inputs the input samples into a convolutional neural network and a bidirectional long short-term memory network. It captures local features through convolution operations, and uses a bidirectional mechanism to capture the bidirectional temporal data dependencies of the axle box acceleration data. The feature fusion layer of the neural network fusion model combines the convolutional local features output by the convolutional neural network with the temporal features output by the bidirectional long short-term memory network to obtain the recognition result of the rail corrugated section, including:

[0197] The selected components are simultaneously input into a convolutional neural network and a bidirectional long short-term memory network. The convolutional neural network uses four layers of convolution to capture the intrinsic mode function component features across spatiotemporal points. The bidirectional long short-term memory network extracts the forward and backward temporal dependency features of the intrinsic mode function components of the axle box acceleration through a bidirectional mechanism.

[0198] A fusion processing layer of convolutional neural network and bidirectional long short-term memory network is constructed. The output features of the convolutional neural network and bidirectional long short-term memory network are fused together. Four fully connected layers are superimposed to perform feature fusion and dimensionality reduction calculation. Spatiotemporal features under different feature frequency patterns are extracted. The output layer uses the Sigmoid function to return the data category probability to identify the rail corrugated section and obtain the rail corrugated section and the rail non-corrugated section.

[0199] In one embodiment, the fusion recognition module 1240 is further configured to:

[0200] The high-dimensional features obtained from the convolutional neural network and the bidirectional long short-term memory network are flattened into one-dimensional features. The output feature dimensions of the convolutional neural network and the bidirectional long short-term memory network are aligned, and the output features of the convolutional neural network and the bidirectional long short-term memory network are concatenated in the vertical dimension to form multi-dimensional features. Four fully connected layers are used for fusion calculation and feature dimensionality reduction. The hyperbolic tangent activation function is selected as the activation function for each fully connected layer, and the sigmoid function is selected in the output layer to make a probability judgment on the data category to obtain the recognition result.

[0201] In one embodiment, the convolutional neural network consists of multiple stacked convolutional layers, pooling layers, fully connected layers, and an output layer. The convolutional layers extract multi-class feature information from the original data through convolutional kernels and perform dimensionality reduction processing on the original high-dimensional input data. The calculation formula is as follows:

[0202]

[0203] RELU(x) = max(0,x); (9)

[0204] pooling(x) = max(x); (10)

[0205] In the formula, It is the calculated value of the i-th neuron in the k-th layer of the convolutional layer; This is the output of neuron j in the local region of the k-1th convolutional layer above; The parameters are the kernel parameters of the k-th layer; For bias terms; M i f(x) is the output of the convolution in the (k-1)th layer above; f(x) is the non-linear activation function, using the ReLU function; x represents the neuron output; pooling is the pooling operation, and pooling(x) = max(x) is the max pooling operation.

[0206] In one embodiment, the bidirectional long short-term memory network is a long short-term memory network with added bidirectionality; the long short-term memory network is based on a recurrent neural network, and by adding forget gates, input gates, and output gates, it establishes dependencies between time-series data, where the state information of one unit is:

[0207] f t =σ(W f ·[h t-1 ,x t ]+b f (11)

[0208] i t =σ(W i ·[h t-1 ,x t ]+b i (12)

[0209] g t =Tanh(W g ·[h t-1 ,x t ]+b g (13)

[0210] o t =σ(W o ·[h t-1 ,x t ]+b o (14)

[0211] C t =f t *C t-1 +i t *g t (15)

[0212] h t =Tanh(C t )*o t (16)

[0213] In the formula, f tIt is the output of the forget gate at time step t; σ is the Sigmoid activation function; W f h is the state transition matrix for the forget gate; t-1 It is the hidden state of the previous time step; x t b is the input for the current time step; f It is the bias term of the forget gate; i t It is the input gate output of time step t; W i b is the state transition matrix of the input gate; i It is the bias term of the input gate; g t Candidates at time step t are candidate memory units; tanh is the hyperbolic tangent activation function; W g It is the weight matrix of the candidate memory units; b g It is the bias term for candidate memory units; o t It is the output gate output at time step t; W o b is the state transition matrix of the output gate; o It is the bias term of the output gate; C t It is the cell state at time step t; C t-1 It is the cell state of the time step preceding time step t; h t It is the hidden state of time step t.

[0214] It should be noted that although several modules of the rail corrugation identification device based on the neural network fusion model have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of the present invention, the features and functions of two or more modules described above can be embodied in one module. Conversely, the features and functions of one module described above can be further divided and embodied by multiple modules.

[0215] Based on the aforementioned inventive concept, such as Figure 13 As shown, the present invention also proposes a computer device 1300, including a memory 1310, a processor 1320, and a computer program 1330 stored in the memory 1310 and executable on the processor 1320. When the processor 1320 executes the computer program 1330, it implements the aforementioned rail corrugation identification method based on a neural network fusion model.

[0216] Based on the aforementioned inventive concept, this invention proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned rail corrugation identification method based on a neural network fusion model.

[0217] Based on the aforementioned inventive concept, this invention proposes a computer program product, which includes a computer program that, when executed by a processor, implements a rail corrugation identification method based on a neural network fusion model.

[0218] This invention addresses the problem of overall drift in acceleration data caused by factors such as axle box acceleration sensors. It uses Fast Fourier Transform and Inverse Fourier Transform to filter out trends and restore the vertical vibration acceleration data of the axle box.

[0219] To address the issues of limited information and feature dimensions, as well as non-stationarity of time-series data in one-dimensional axle box acceleration data mining, VMD is used to decompose axle box acceleration into intrinsic mode components at different characteristic frequencies. The top four intrinsic mode components with the highest energy values ​​are selected based on signal energy to expand the axle box acceleration features and solve the problems of nonlinearity and non-stationarity of axle box acceleration.

[0220] To address the issues of cumbersome feature extraction in corrugation identification technology and the fact that current corrugation identification methods mostly rely on simulation data to achieve high identification accuracy, this paper combines signal processing methods with deep learning to automatically and synchronously capture the local features of measured axle box acceleration and the successive dependency characteristics of time-series data. This fully explores the axle box acceleration features and efficiently and accurately identifies corrugated sections of high-speed railway rails.

[0221] To address the limitations of CNNs in extracting only spatial convolutional features and Bi-LSTMs in extracting only temporal features, this paper innovatively proposes a CNN-Bi-LSTM fusion network. This network connects CNN and Bi-LSTM networks in parallel and overlays a fully connected network to learn the spatial and temporal features of axle box acceleration, thereby improving the network's generalization and feature learning capabilities.

[0222] The axle boxes of a train are connected to the train wheelsets and can directly feel the vibrations and impacts caused by track irregularities. When a train travels through a corrugated section, the axle boxes are subjected to periodic impacts, which manifest as axle box acceleration. Therefore, data analysis and feature extraction of train axle box acceleration can identify corrugated sections of the rail. However, traditional corrugation identification methods rely mainly on signal processing, requiring manual feature extraction and threshold setting, making the feature extraction process cumbersome. Deep learning-based corrugation identification methods mostly achieve high accuracy based on subway simulation data, with limited research on measured axle box acceleration data from high-speed rail operations. This invention creatively proposes an end-to-end intelligent identification method and terminal device for high-speed rail rail corrugation. Based on measured high-speed rail axle box acceleration, it combines signal processing methods with deep learning, and applies CNN and Bi-LSTM in synergy to automatically and synchronously capture local and temporal features of axle box acceleration, effectively and promptly identifying rail corrugated sections and providing a scientific basis for railway track maintenance.

[0223] Compared with existing technologies, the rail corrugation identification method and device based on a neural network fusion model proposed in this invention have at least the following technical advantages:

[0224] 1. By adopting the axle box acceleration numerical restoration method, the overall drift of acceleration data caused by factors such as acceleration sensors is eliminated, the trend of axle box acceleration data is filtered out, and the accuracy of axle box acceleration data is significantly improved.

[0225] 2. By using the VMD method to extract the vertical acceleration feature components of the axle box, the feature dimension of the one-dimensional axle box acceleration is effectively expanded, eliminating the influence of factors such as vertical excitation, environmental noise, and resonance, improving the signal-to-noise ratio of the axle box acceleration feature components, and providing high-quality, multi-resolution input features for the wave wear identification model.

[0226] 3. The end-to-end VMD-CNN-Bi-LSTM corrugation recognition method can learn the complex nonlinear relationship of axle box acceleration and has a strong representation learning ability. It breaks through the feature learning limitations of single CNN and Bi-LSTM models, and fully learns the signal characteristics, temporal characteristics and data characteristics of different spatial points of corrugation and normal axle box acceleration. It has excellent recognition ability for corrugated sections of high-speed rail rails, with high degree of automation, superior performance and strong generalization ability.

[0227] 4. This invention provides a new approach for identifying other types of track geometric shortwave defects. By deeply analyzing and comprehensively examining the measured high-speed rail corrugation and axle box acceleration in normal sections from actual train inspections, it mines the characteristics and spatiotemporal features of axle box acceleration signals and designs a highly efficient algorithm to quickly and accurately identify high-speed rail corrugation sections. In this innovative research, key parameters and thresholds have undergone in-depth experimental analysis and precise adjustment to ensure that the algorithm achieves optimal performance in various practical application scenarios. Based on the optimized parameters, the efficiency, accuracy, stability, and reliability of the algorithm have been significantly improved. This innovative method effectively improves the accuracy and efficiency of high-speed rail corrugation identification, simplifies the axle box acceleration feature mining process, provides strong technical support for railway track defect detection, and provides a scientific basis for track maintenance and repair.

[0228] The acquisition, storage, use, and processing of data in this application comply with relevant laws and regulations.

[0229] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0230] This invention is described with reference to flowchart illustrations and / or block diagrams of methods and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0231] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0232] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0233] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered 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 method for identifying rail corrugation based on a neural network fusion model, characterized in that, include: Collect vertical acceleration data of the train's axle boxes; The vertical vibration acceleration data of the axle box is discretized in the frequency domain by Fourier transform, low-frequency signals are filtered out, and the time-domain axle box acceleration data after filtering out the trend is restored by inverse Fourier transform. The time-domain axle box acceleration data is decomposed using variational mode decomposition to obtain multi-resolution intrinsic mode function components. Several components with the highest signal energy proportion among these intrinsic mode function components are selected and used as input samples for a neural network fusion model. This neural network fusion model is composed of a convolutional neural network and a bidirectional long short-term memory network, and is trained using axle box acceleration data corresponding to corrugated and non-corrugated rail sections. The input samples are synchronously input into a convolutional neural network and a bidirectional long short-term memory network. Local features are captured through convolution operations, and the bidirectional mechanism is used to capture the bidirectional temporal data dependency of the axle box acceleration data. The feature fusion layer of the neural network fusion model combines the convolutional local features output by the convolutional neural network with the temporal features output by the bidirectional long short-term memory network to obtain the identification result of the rail corrugation section.

2. The rail corrugation identification method based on a neural network fusion model according to claim 1, characterized in that, Collect vertical acceleration data of the train's axle boxes, including: A sliding window is used to retrieve vertical vibration acceleration data of the train's axle boxes. Each preset number of axle box vertical vibration acceleration data constitutes a sample data.

3. The rail corrugation identification method based on a neural network fusion model according to claim 1, characterized in that, The vertical vibration acceleration data of the axle box is discretized in the frequency domain using Fourier transform, low-frequency signals are filtered out, and the time-domain axle box acceleration data after filtering out the trend is restored using inverse Fourier transform, including: The Fourier transform converts a time-domain signal to the frequency domain. The transformation formula is as follows: In the formula, X[k] is a discrete sample of the frequency domain signal, representing the complex amplitude of the k-th frequency component; x[n] is a discrete sample of the original time domain signal; N is the number of sample points, representing the length of the signal; k is a discrete variable in the frequency domain; j is the imaginary unit. It is a complex exponential function; n is a discrete variable in the time domain; The inverse Fourier transform is used to restore the time-domain axle box acceleration data after filtering out the trend. The transformation formula is as follows: In the formula, x[n] is a discrete sample of the reconstructed time-domain signal; X[k] is a discrete sample of the frequency-domain signal; k and n are discrete variables in the frequency domain and time domain, respectively; j is the imaginary unit; and N is the number of sample points.

4. The rail corrugation identification method based on a neural network fusion model according to claim 1, characterized in that, The time-domain axle box acceleration data is decomposed using variational mode decomposition to obtain multi-resolution intrinsic mode function (IMF) components. Several IMF components with the highest signal energy proportion are selected and used as input samples for the neural network fusion model, including: Construct a variational model, let the original time series signal be f(t), and use the Hilbert transform to solve for the K intrinsic mode function components. The calculation formula is as follows: In the formula, u k It is the k-th intrinsic mode function component; t represents time; w k It is the frequency center of the k-th mode, and K is the total number of modes; represents the derivative of the function with respect to time t; ||·||2 represents the L2 norm; i is the imaginary unit; st represents the constraint condition; f(t) represents the original time series signal; To solve the variational model, we transform it using the augmented Lagrange function, as shown in the following expression: In the formula, L represents the augmented Lagrange function; α is the quadratic penalty factor; and λ is the Lagrange multiplier. Based on the variational model, the time-domain axle box acceleration data is decomposed using the variational mode decomposition method to obtain multi-order intrinsic mode function components.

5. The rail corrugation identification method based on a neural network fusion model according to claim 4, characterized in that, The time-domain axle box acceleration data is decomposed using variational mode decomposition to obtain multi-resolution intrinsic mode function (IMF) components. Several IMF components with the highest signal energy proportion are selected and used as input samples for the neural network fusion model, including: For the aforementioned multi-order intrinsic mode function components, the corresponding energy percentage is calculated; wherein, the energy of the intrinsic mode function components is calculated using the signal energy formula, the calculation formula is as follows: In the formula, E i S represents the energy of the i-th intrinsic mode function component; j is the j-th sample value in the intrinsic mode function component; N′ is the number of samples in the intrinsic mode function component; P i E represents the proportion of the energy of the i-th intrinsic mode function component to the total energy; j Let M represent the energy of the j-th intrinsic mode function component, and M be the total number of intrinsic mode function components. Variational mode decomposition was performed on multiple sets of axle box acceleration data that were both corrugated and non-corrugated. The data were sorted according to their energy proportions, and the components with the highest signal energy proportions were selected as input samples for the neural network fusion model.

6. The rail corrugation identification method based on a neural network fusion model according to claim 1, characterized in that, The input samples are synchronously fed into a convolutional neural network and a bidirectional long short-term memory network. Local features are captured through convolution operations, and the bidirectional mechanism is used to capture the bidirectional temporal data dependencies of the axle box acceleration data. The feature fusion layer of the neural network fusion model combines the convolutional local features output by the convolutional neural network with the temporal features output by the bidirectional long short-term memory network to obtain the identification result of the rail corrugation section, including: The selected components are simultaneously input into a convolutional neural network and a bidirectional long short-term memory network. The convolutional neural network uses four layers of convolution to capture the intrinsic mode function component features across spatiotemporal points. The bidirectional long short-term memory network extracts the forward and backward temporal dependency features of the intrinsic mode function components of the axle box acceleration through a bidirectional mechanism. A fusion processing layer of convolutional neural network and bidirectional long short-term memory network is constructed. The output features of the convolutional neural network and bidirectional long short-term memory network are fused together. Four fully connected layers are superimposed to perform feature fusion and dimensionality reduction calculation. Spatiotemporal features under different feature frequency patterns are extracted. The output layer uses the Sigmoid function to return the data category probability to identify the rail corrugated section and obtain the rail corrugated section and the rail non-corrugated section.

7. The rail corrugation identification method based on a neural network fusion model according to claim 6, characterized in that, The method also includes: The high-dimensional features obtained from the convolutional neural network and the bidirectional long short-term memory network are flattened into one-dimensional features. The output feature dimensions of the convolutional neural network and the bidirectional long short-term memory network are aligned, and the output features of the convolutional neural network and the bidirectional long short-term memory network are concatenated in the vertical dimension to form multi-dimensional features. Four fully connected layers are used for fusion calculation and feature dimensionality reduction. The hyperbolic tangent activation function is selected as the activation function for each fully connected layer, and the sigmoid function is selected in the output layer to make a probability judgment on the data category to obtain the recognition result.

8. The rail corrugation identification method based on a neural network fusion model according to claim 6, characterized in that, The convolutional neural network consists of multiple layers of stacked convolutional layers, pooling layers, fully connected layers, and an output layer. The convolutional layers extract multi-class feature information from the original data through convolutional kernels and perform dimensionality reduction processing on the original high-dimensional input data. The calculation formula is as follows: RELU(x) = max(0,x); pooling(x) = max(x); In the formula, It is the calculated value of the i-th neuron in the k-th layer of the convolutional layer; This is the output of neuron j in the local region of the k-1th convolutional layer above; The parameters are the kernel parameters of the k-th layer; For bias terms; M i f(x) is the output of the convolution in the (k-1)th layer above; f(x) is the non-linear activation function, using the ReLU function; x represents the neuron output; pooling is the pooling operation, and pooling(x) = max(x) is the max pooling operation.

9. The rail corrugation identification method based on a neural network fusion model according to claim 6, characterized in that, The bidirectional long short-term memory network is based on the long short-term memory network with added bidirectionality. The long short-term memory network is based on a recurrent neural network, and establishes dependencies between time-series data by adding forget gates, input gates, and output gates. The state information of one unit is: f t =σ(W f ·[h t-1 ,x t ]+b f ); i t =σ(W i ·[h t-1 ,x t ]+b i ); rice t = Tanh(W g ·[h t-1 ,x t ]+b g ); the t =σ(W o ·[h t-1 ,x t ]+b o ); C t =f t *C t-1 +i t *g t ; h t = Tanh(C) t )*o t ; In the formula, f t It is the output of the forget gate at time step t; σ is the Sigmoid activation function; W f h is the state transition matrix for the forget gate; t-1 It is the hidden state of the previous time step; x t b is the input for the current time step; f It is the bias term of the forget gate; i t It is the input gate output of time step t; W i b is the state transition matrix of the input gate; i It is the bias term of the input gate; g t Candidates at time step t are candidate memory units; tanh is the hyperbolic tangent activation function; W g It is the weight matrix of the candidate memory units; b g It is the bias term for candidate memory units; o t It is the output gate output at time step t; W o b is the state transition matrix of the output gate; o It is the bias term of the output gate; C t It is the cell state at time step t; C t-1 It is the cell state of the time step preceding time step t; h t It is the hidden state of time step t.

10. A rail corrugation identification device based on a neural network fusion model, characterized in that, include: The data acquisition module is used to collect the vertical acceleration data of the train's axle boxes; The data processing module is used to discretize the vertical vibration acceleration data of the axle box in the frequency domain through Fourier transform, filter out low-frequency signals, and restore the time-domain axle box acceleration data after filtering out the trend using inverse Fourier transform. The variational mode decomposition module is used to decompose the time-domain axle box acceleration data using the variational mode decomposition processing method to obtain multi-resolution intrinsic mode function components. Several components with the highest signal energy proportion among the intrinsic mode function components are selected, and these selected components are used as input samples for the neural network fusion model. The neural network fusion model is composed of a convolutional neural network and a bidirectional long short-term memory network, and is trained using axle box acceleration data corresponding to corrugated and non-corrugated rail sections. The fusion recognition module is used to synchronously input the input samples into a convolutional neural network and a bidirectional long short-term memory network. It captures local features through convolution operations, and captures bidirectional temporal data dependencies of axle box acceleration data through a bidirectional mechanism. The feature fusion layer of the neural network fusion model combines the convolutional local features output by the convolutional neural network with the temporal features output by the bidirectional long short-term memory network to obtain the recognition result of the rail corrugation section.

11. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 9.

13. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 9.

Images