A Deep Learning-Based Method for Identifying the Coupler Force Across the Entire Domain of Heavy-Haul Trains

By using deep learning technology, a dynamic adjacency matrix is ​​generated using multi-scale one-dimensional convolution, bidirectional LSTM, and graph convolution. This solves the problems of high cost and insufficient reliability of traditional sensor monitoring, and enables accurate prediction of the coupler force across the entire range of heavy-haul trains, reducing operation and maintenance costs and improving safety.

CN123115709APending Publication Date: 2026-11-17SOUTHWEST JIAOTONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611633427.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Traditional solutions for monitoring coupler forces by deploying sensors across the entire area are costly and unreliable. Furthermore, the strong nonlinearity and strong correlation of coupler forces make it difficult to accurately capture and predict them in advance, which affects the safety management and operation and maintenance costs of heavy-haul trains.

Method used

By employing a deep learning-based approach, a dynamic adjacency matrix is ​​generated through multi-scale one-dimensional convolutional feature extraction, bidirectional LSTM network, and multi-head self-attention mechanism combined with relative position encoding and graph convolution technology, thereby achieving accurate prediction of coupler force from finite measurement points to the entire field.

Benefits of technology

It enables accurate prediction of coupler force across the entire range under limited measurement points, reducing the cost of monitoring equipment and improving the overall efficiency and safety of heavy-duty transportation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115709A_ABST
    Figure CN123115709A_ABST
Patent Text Reader

Abstract

This invention provides a deep learning-based method for identifying the full-domain coupler force of heavy-haul trains, relating to the field of rail transit technology. The method includes: extracting multi-scale one-dimensional convolutional features from temporal data of coupler force collected by sensors, outputting multi-scale local temporal features; fusing the bidirectional LSTM output and multi-head self-attention output through residual connections and layer normalization, then concatenating them with the multi-scale local temporal features and mapping them to the target dimension; generating a dynamic adjacency matrix based on the physical topology of the train formation and the index of the measuring point positions, and aggregating spatial features through graph convolution guided by the dynamic adjacency matrix to output the spatiotemporal fusion features of known measuring points; and adaptively allocating contribution weights based on the spatiotemporal fusion features, position coordinates, and target position coordinates of known measuring points using a spatial attention mechanism, interpolating to generate the full-domain coupler force. This invention achieves accurate prediction of the full-domain coupler force from key measuring points by using temporal data and spatial position information from a limited number of measuring points as input.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of rail transit technology, and in particular to a deep learning-based method for identifying the coupler force across the entire range of heavy-haul trains. Background Technology

[0002] With the continuous development of heavy-haul trains in my country and the sustained increase in train load and formation length, the longitudinal interaction forces between trains have intensified exponentially, posing unprecedented and stringent challenges to the longitudinal dynamics control of trains. Coupler force, as a core indicator directly reflecting the longitudinal dynamics state of a train, not only serves as a "barometer" for judging train operational safety but is also a crucial basis for achieving precise control of longitudinal impulses. Real-time monitoring of the entire range of coupler force distribution characteristics can provide data support for the dynamic adjustment of the traction and braking system, preventing component damage and safety accidents caused by sudden force changes from the outset.

[0003] Traditional methods of deploying sensors across the entire area to monitor coupler forces suffer from drawbacks such as high cost, insufficient reliability in harsh environments, and poor maintenance compatibility. Furthermore, the strong nonlinearity and strong correlation of coupler forces caused by multi-factor coupling mean that the dynamic distribution, transmission attenuation characteristics, and triggering and evolution mechanisms of coupler forces along the train formation have not been fully quantified and analyzed, making it difficult to accurately capture and predict coupler forces. Therefore, systematic mechanistic research and model building of coupler forces in heavy-haul trains to achieve real-time, accurate, and low-cost monitoring of coupler forces is not only an urgent need to solve the safety management challenges of heavy-haul trains, but also of irreplaceable importance for reducing operation and maintenance costs and improving the overall efficiency of heavy-haul transportation.

[0004] Therefore, there is an urgent need for a deep learning-based method for identifying the coupler force across the entire range of heavy-haul trains. This method would use time-series data and spatial location information from a limited number of measuring points as input to achieve accurate prediction of the coupler force across the entire range from key measuring points. This would reduce the installation of monitoring equipment, lower operation and maintenance costs, and improve the overall efficiency of heavy-haul transportation. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a deep learning-based method for identifying the coupler force across the entire range of heavy-haul trains, so as to achieve accurate prediction of the coupler force across the entire range from key measuring points by using time-series data and spatial location information of limited measuring points as input.

[0006] In a first aspect, embodiments of the present invention provide a method for identifying the full-domain coupler force of heavy-haul trains based on deep learning, including: Multi-scale one-dimensional convolutional feature extraction is performed on the time-series data of coupler force collected by the sensor, and the multi-scale local time-series features are output after weighted fusion by learnable weight coefficients. A bidirectional LSTM network is used to model the forward and backward long-range dependencies of temporal features. Multi-subspace attention is calculated by combining a multi-head self-attention mechanism with relative position encoding. The bidirectional LSTM output and the multi-head self-attention output are fused through residual connection and layer normalization, and then concatenated with multi-scale local temporal features and mapped to the target dimension. Constrained by the physical topology of train formation, the spatial correlation between measurement points is mined through multi-scale spatial attention and the attention score is corrected by physical prior with relative distance. A dynamic adjacency matrix is ​​generated based on the measurement point position index. Guided by the dynamic adjacency matrix, spatial features are aggregated through graph convolution and the spatiotemporal fusion features of known measurement points are output. Based on the spatial attention mechanism, contribution weights are adaptively assigned according to the known spatiotemporal fusion features of the measuring points, their position coordinates, and the target position coordinates, and the global coupler force is generated by interpolation.

[0007] Optionally, multi-scale one-dimensional convolution employs three parallel convolution branches to capture the instantaneous change features, short-term fluctuation features, and medium-term trend features of the time-series signal, respectively.

[0008] Optionally, the bidirectional LSTM network has a two-layer structure, with the hidden layer dimension set to half the input feature dimension, and the forward and backward outputs concatenated to restore the input feature dimension.

[0009] Optionally, the multi-head self-attention mechanism maps features to H subspaces and computes attention weights in parallel, where H is a preset positive integer.

[0010] Optionally, multi-scale spatial attention uses depthwise convolutional branches with different kernel sizes to extract measurement point features at different spatial scales in parallel, and globally models the correlation between measurement points through the attention mechanism branches. The outputs of each branch are then fused to obtain multi-scale spatial attention weighted features.

[0011] Optionally, the graph convolution adopts a two-level structure. The first level performs matrix multiplication with the spatial attention output using the dynamic adjacency matrix as the left multiplication factor to obtain preliminary fusion features. The second level concatenates the preliminary fusion features with the spatial attention output along the feature dimension. After linear layer mapping and activation function processing, the spatiotemporal fusion features of known measurement points with the same output dimension as the spatial attention output dimension are obtained.

[0012] Optionally, the contribution weight is generated by concatenating the difference and variance between the target location and the locations of each measurement point into a two-dimensional feature, mapping it into a high-dimensional feature through MLP, and taking the mean along the feature dimension to convert it into a scalar position influence factor. The scalar position influence factor is then generated by adding the basic attention score element by element and normalizing it.

[0013] Optionally, multi-scale spatial attention uses numerical clamping to constrain the relative distance difference between measurement points within a preset range to adapt to different train formation lengths.

[0014] Secondly, embodiments of the present invention also provide a heavy-haul train full-range coupler force identification system, the system comprising: The first extraction unit is used to extract multi-scale one-dimensional convolutional features from the time-series data of the coupler force collected by the sensor, and output multi-scale local time-series features after weighted fusion by learnable weight coefficients. The data processing unit is used to model the forward and backward long-range dependencies of temporal features using a bidirectional LSTM network, and to perform multi-subspace attention calculation by combining a multi-head self-attention mechanism with relative position encoding. The bidirectional LSTM output and the multi-head self-attention output are fused through residual connection and layer normalization, and then concatenated with multi-scale local temporal features and mapped to the target dimension. The second extraction unit is used to mine the spatial relationship between measurement points through multi-scale spatial attention and correct the attention score with relative distance physical prior, constrained by the physical topology of train formation. It generates a dynamic adjacency matrix based on the measurement point position index, and aggregates spatial features through graph convolution guided by the dynamic adjacency matrix, outputting the spatiotemporal fusion features of known measurement points. The generation unit is used to adaptively allocate contribution weights based on the known spatiotemporal fusion features of the measuring points, their position coordinates, and the target position coordinates, and then interpolate to generate the global coupler force.

[0015] Thirdly, embodiments of the present invention also provide a computer-readable storage medium storing computer instructions that cause a computer to execute a deep learning-based method for identifying the full-domain coupler force of a heavy-haul train, as described in the first aspect or any embodiment of the first aspect.

[0016] The embodiments of the present invention bring the following beneficial effects: This invention establishes a model that uses time-series data and spatial location information from limited measurement points as input, integrates the advantages of prior knowledge of train longitudinal dynamics physics and data-driven approaches, and establishes a time-domain feature encoder, a spatial feature encoder, and a coupler force spatial feature interpolation model. Through time-series feature extraction, spatial dependency modeling, and spatiotemporal fusion architecture design, it achieves accurate extrapolation estimation of coupler force from key measurement points to the entire domain.

[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating a deep learning-based method for identifying the full-domain coupler force of a heavy-haul train, as provided in an embodiment of the present invention. Figure 2 This is a structural block diagram of a heavy-haul train full-range coupler force identification system provided in an embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Traditional methods of deploying sensors across the entire area to monitor coupler forces suffer from drawbacks such as high cost, insufficient reliability in harsh environments, and poor maintenance compatibility. Furthermore, the strong nonlinearity and strong correlation of coupler forces caused by multi-factor coupling mean that the dynamic distribution, transmission attenuation characteristics, and triggering and evolution mechanisms of coupler forces along the train formation have not been fully quantified and analyzed, making it difficult to accurately capture and predict coupler forces. Therefore, systematic mechanistic research and model building of coupler forces in heavy-haul trains to achieve real-time, accurate, and low-cost monitoring of coupler forces is not only an urgent need to solve the safety management challenges of heavy-haul trains, but also of irreplaceable importance for reducing operation and maintenance costs and improving the overall efficiency of heavy-haul transportation.

[0023] Based on this, the present invention provides a deep learning-based method for identifying the coupler force of heavy-haul trains across the entire region. This method can use time-series data and spatial location information of limited measurement points as input to achieve accurate prediction of the coupler force from key measurement points across the entire region, thereby reducing the installation of monitoring equipment, lowering operation and maintenance costs, and improving the overall efficiency of heavy-haul transportation.

[0024] To facilitate understanding of this embodiment, a deep learning-based method for identifying the full-domain coupler force of heavy-haul trains will first be described in detail.

[0025] Figure 1 This is a flowchart illustrating a deep learning-based method for identifying the full-domain coupler force of heavy-haul trains, as provided in an embodiment of the present invention. Figure 1 As shown, the method may include: S102: Multi-scale one-dimensional convolutional feature extraction is performed on the time-series data of coupler force collected by the sensor, and multi-scale local time-series features are output after weighted fusion by learnable weight coefficients.

[0026] The time-series data of train coupler forces includes millisecond-level instantaneous impacts caused by track irregularities and traction / braking operations, second-level oscillations caused by the transmission of vehicle coupling gaps, and minute-level trends formed by changes in train operating conditions. A single-scale convolutional kernel cannot simultaneously capture feature information at these three time granularities. Therefore, in this embodiment, multi-scale one-dimensional convolution is used to extract features at the three time scales in parallel, and then adaptively fused using learnable weight coefficients. This provides a complete and granular local temporal representation for subsequent long-range dependency modeling of bidirectional LSTM and temporal association enhancement of multi-head self-attention.

[0027] It should be noted that the coupler force time-series data in this application embodiment is data from a finite number of measuring points. The finite number of measuring points can be couplers arranged at the rear of a finite number of cars in the train; alternatively, the finite number of measuring points can be obtained by searching using a greedy algorithm based on the correlation coefficient between each coupler in the train and the correlation coefficients of various working conditions, with a certain limit. In the scenario of monitoring coupler force in heavy-haul trains, measuring point sensors are usually installed at the coupler of each car. The measuring points are numbered sequentially from the front to the rear of the train according to the physical order of the train formation. It should be noted that the rear of the car refers to the end located according to the train's direction of travel; the end in the direction of travel is usually considered the front, and the opposite side is called the rear. The finite number of measuring points used in this application embodiment can be distributed at the front, middle, and rear of the car. The finite number of measuring points in this application embodiment can be multiple discontinuous measuring points. For example, this application embodiment can select measuring points numbered 1, 54, 108, 109, 135, 163, 210, and 215. The advantages of using the above-mentioned limited measurement point distribution are as follows: on the one hand, high-precision interpolation of the coupler force across the entire range can be achieved with a small number of key measurement points, significantly reducing the cost of sensor deployment; on the other hand, the selected measurement points cover the front-end excitation source, the middle section of the violent oscillation zone, and the rear-end transmission end of the coupler force longitudinal propagation, and the measurement points are kept discontinuous to reduce information redundancy, thereby maximizing information coverage under limited measurement point conditions and ensuring the accuracy and robustness of the coupler force identification across the entire range.

[0028] The multi-scale one-dimensional convolution feature extraction in this embodiment is implemented using a multi-scale convolution module. The multi-scale convolution module processes the input time-series signal through three parallel one-dimensional convolution branches, respectively capturing the instantaneous changes, short-term fluctuations, and medium-term trend changes of the coupler force time-series signal. Based on learnable weight coefficients, the module performs feature fusion and batch normalization on each convolution branch to eliminate dimensional differences.

[0029] The normalization process is as follows:

[0030] in, and Var[ ] are the mean and variance of the features within the batch, respectively. To prevent small constants with a denominator of zero ( =10 5 ), γ and β are learnable scaling factors and offsets, which enhance the model's ability to adapt to feature distribution.

[0031] The three parallel one-dimensional convolutional models of the multi-scale convolutional model capture the instantaneous changes, short-term fluctuations, and medium-term trends of the coupler force time series signal by using convolutional kernel sizes of 1, 3, and 5, respectively.

[0032] Specifically, in this embodiment, a time-domain feature encoder model is used to process the coupler force time series data.

[0033] The collected coupler force time-series data is reshaped to adapt its dimensions from [B, N, L] to [B×N, 1, L]. The adapted data is then processed through three parallel one-dimensional convolutional branches to extract features from each branch and weighted fusion. This process involves different convolutional kernel sizes and pooling operations, with kernel sizes of 1, 3, and 5 to capture instantaneous changes, short-term fluctuations, and medium-term trends in the coupler force. Finally, learnable weight coefficients are used to batch normalize the fused features, yielding local features of the coupler force and eliminating dimensional differences. Here, B represents the number of data batches loaded into GPU memory during model training or prediction inference, N is the number of known measurement points, and L is the sampling window length of the coupler force time-series signal at each measurement point. This can be expressed by the following formula:

[0034] in, This provides the time-series data of the coupler force sensor at the finite measuring points of the coupler force within the [t-L+1,t] time window. The feature tensor after convolution with uniform dimensions. These are learnable weight coefficients; This is the i-th one-dimensional convolution branch; For one-dimensional batch normalization.

[0035] S104: A bidirectional LSTM network is used to model the forward and backward long-range dependencies of temporal features. Multi-subspace attention is calculated by combining a multi-head self-attention mechanism with relative position encoding. The bidirectional LSTM output and the multi-head self-attention output are fused through residual connection and layer normalization, and then concatenated with multi-scale local temporal features and mapped to the target dimension. In this embodiment, a bidirectional LSTM network module is used to model the forward and backward long-range dependencies of temporal features. The bidirectional LSTM network captures the long-range temporal dependencies of the coupler force signal from the past to the present and from the future to the present. By introducing a multi-head self-attention mechanism, the temporal correlation between key time steps and different measurement points is explicitly strengthened. The features are mapped to H subspaces and attention weights are calculated in parallel to capture multi-dimensional temporal dependencies.

[0036] By introducing residual connections and layer normalization into the deep network architecture of the temporal feature encoder, the stability of model training is ensured and the gradient vanishing problem is alleviated. The results of the multi-head self-attention branch and the bidirectional LSTM network branch are cross-fused.

[0037] Feature concatenation and linear projection are used to concatenate the results of multi-scale one-dimensional convolution with the results of cross-fusion and map them to the target dimension.

[0038] The original data is input into a bidirectional LSTM network, which consists of a forward LSTM and a backward LSTM. The forward LSTM is used to capture the temporal dependencies from the past to the present, and the backward LSTM is used to capture the inverse dependencies from the future to the present. The hidden layers of both the forward and backward LSTM networks are set to D / 2 dimensions, and the bidirectional outputs are concatenated to D dimensions. The final bidirectional fused output can be expressed by the following formula:

[0039] in, The feature tensor after bidirectional fusion, α∈[0,1] represents the adaptive fusion weights, and the forward hidden layer state... Carrying time-series information, reverse-hiding state. It carries future time series information after the current time, where D is the dimension of the extracted data features.

[0040] Simultaneously, a multi-head self-attention mechanism is introduced, comprising three independent linear layers. Features are mapped to H subspaces, and attention weights are computed in parallel to capture multi-dimensional temporal dependencies. Relative position encoding is also incorporated to enhance the physical plausibility of association modeling. This can be expressed by the following formula:

[0041] in, The basic temporal features output by the multi-head self-attention mechanism module are H, where H is the number of attention heads, and W is the number of attention heads.O A multi-head fusion matrix; This is the function for feature splicing operations.

[0042] The attention of the h-th head is calculated as follows:

[0043] in, Let h be the query tensor for the head attention. ; Let h be the key tensor of the attention head. ; Let h be the tensor of the attention value of the h-th head. ; For relative time encoding tensors, Feature dimensions for each head; softmax ( ) is the normalization function.

[0044] Relative time-encoded tensor The calculation formula is:

[0045] in, Let h be the projection matrix of the h-th head. The position-encoded projection matrix; RelPosEnc( ) represents the relative position encoding function, t represents the query time step index, the current time step in which attention weights are being calculated, and the corresponding query matrix. The t-th row in the matrix; k is the key time step index, corresponding to the key matrix. The kth row in the array.

[0046] To achieve adaptive complementarity between LSTM and multi-head attention features, residual connections were designed. Layer normalization was used to constrain feature distribution, accelerating model convergence and ensuring the efficient and stable cascaded feature extraction process of multi-scale convolution, LSTM, and multi-head attention. This operation yields the globally fused features of the coupler force. Residual connections, through "direct cross-layer feature connections," allow the network to learn the "residual between input and output" rather than directly learning the output, reducing learning difficulty. Furthermore, fitting the residuals converges more easily than directly fitting the target features. This process can be represented by the following formula:

[0047] in, The output characteristics of weighted residual connections, The basic timing characteristics output by the bidirectional LSTM module; The basic temporal features output by the multi-head self-attention mechanism module; These are learnable weighting coefficients. For the linear projection matrix from the original input features to the LSTM features and the linear projection matrix from the original input features to the features of the multi-head self-attention mechanism module; It is a temporal feature tensor.

[0048] S106: With the physical topology of train formation as a constraint, the spatial correlation between measurement points is mined through multi-scale spatial attention and the attention score is corrected with physical priors of relative distance. A dynamic adjacency matrix is ​​generated based on the measurement point position index. The spatial features are aggregated through graph convolution guided by the dynamic adjacency matrix, and the spatiotemporal fusion features of known measurement points are output.

[0049] First, index the tensor based on the known positions of the measurement points. The relative distance encoding tensor between any two measurement points is obtained. To avoid excessively large absolute values ​​of relative position differences due to excessively long train formation lengths, and to ensure the model's adaptability to different train formation lengths, numerical clamping is used to constrain the relative position differences to a certain range. Within this range, the processing method is as follows:

[0050] Where i and j are the measurement point indices; The preset maximum distance threshold; This is the position difference matrix for all pairs of measuring points. The clamp represents the positional difference between the j-th measuring point and the i-th measuring point in the b-th batch; ) is a numerical clamping function; The relative distance value after clamping.

[0051] By transforming discrete relative position indices into high-dimensional dense features using a learnable distance coding matrix, the distance coding can be obtained. The expression is:

[0052] Where i and j are the measurement point indices; The learnable distance encoding matrix; This is the position difference matrix for all pairs of measuring points; The preset maximum distance threshold; clamp( ) is a numerical clamping function.

[0053] It receives temporal features output from the temporal feature encoder. By designing a depthwise convolutional branch that matches the scale list scales=[1,3,5], it captures associated features of different spatial ranges channel by channel. This can be expressed by the following formula:

[0054] Where k is the scale branch index; This is the output of the temporal feature encoder; Conv1d( ) represents a one-dimensional convolution operation; kernel=saclers[k] is the kernel size of the k-th branch; padding=saclers[k] / 2 is the padding amount.

[0055] Introducing learnable scale weights This achieves adaptive fusion of multi-scale features. After softmax normalization, the features at each scale are weighted to obtain the multi-scale convolutional fusion features. This can be expressed by the following formula:

[0056] Where R is the total number of scale branches; These are features resulting from multi-scale convolutional fusion. Learnable parameters initialized to 1.

[0057] By incorporating distance weights into multi-head attention computation, spatial correlation modeling under physical constraints is achieved. By concatenating the multi-head attention layers, the spatial encoding at known measurement points can be obtained. Encode the relative distance Averaging along dimension 2 ,and After concatenation and linear layer fusion, the feature representation is optimized through residual connections and layer normalization, which can be expressed by the following formula:

[0058] To adapt to topological changes in different heavy-haul train formations and reflect the dynamic attenuation characteristics of coupler force transmission, a physically guided dynamic adjacency matrix construction module was designed, based on the end-to-end learning amplitude of the measurement point position encoding and the association weights of physical laws. First, the measurement point position index is converted into high-dimensional features through the position encoding module, preserving position order information and relative distance characteristics. Then, the position features of the measurement point pairs are concatenated and mapped to the association strength in the [0,1] interval through a feedforward network, generating a dynamic adjacency matrix. This can be expressed by the following formula:

[0059] in, The positional encoding feature of measurement point i relative to measurement point j; The position encoding feature of measurement point j relative to measurement point i; concat( ) is the feature splicing function; The physical correlation strength between measurement points i and j in batch b is represented by Linear3, which maps the output to D dimensions, and Linear4 maps it to 1 dimension.

[0060] in, AIt is a dynamic adjacency matrix. Encode the position of the coupler force measuring point, concat( The output dimension is Linear3 maps the output to a D-dimensional plane, while Linear4 maps it to a 1-dimensional plane, σ( ) is the Sigmoid activation function.

[0061] Distance coding The expression is

[0062] Graph convolutional layers use dynamic adjacency matrices To guide the aggregation of spatial features based on physical associations, ensuring that the feature propagation process strictly follows the topological rules of coupler force transmission, a weighted fusion of the adjacency matrix and spatial features is first achieved through batch matrix multiplication, completing the propagation of physical constraints on the features. To enhance the nonlinear expressive power of the features, an activation function and a Dropout layer are introduced, designing a two-level graph convolutional structure, which can be expressed by the following formula:

[0063] in, For the graph features after initial aggregation, Linear( ) is the feature dimension calibration layer; ReLU( ) is the activation function; dropout( ) is the random deactivation function.

[0064] The final output is the spatiotemporal fusion feature of the known measurement points.

[0065] S108: Based on the spatial attention mechanism, contribution weights are adaptively allocated according to the known spatiotemporal fusion features of the measuring points, their position coordinates, and the target position coordinates, and the global coupler force is generated by interpolation.

[0066] This step uses adaptive encoding to concatenate the difference between the target location and the measurement point location with the orientation into a two-dimensional feature, which is then mapped to adaptive position features via MLP. The high-dimensional position features are averaged along their feature dimensions to transform into a scalar position influence factor, quantifying the positional association strength of each measurement point with the target location. The scalar position influence factor is then element-wise added to the base attention score to achieve adaptive correction of the attention score, resulting in a final attention score that incorporates the positional prior. To eliminate excessive weighting of distant points due to feature noise and to avoid ignoring true nonlinear associations in the data due to overemphasizing physical priors, the final attention score is normalized using softmax to generate attention weights that sum to 1. Finally, the interpolated features of the target location are obtained by weighted summation of the measurement point features using these attention weights.

[0067] Receive the spatiotemporal fusion features of known measurement points. First, the locations of the known measurement points are... Location of the target to be identified The relative positional relationships are taken as input and mapped into high-dimensional dense features through a learnable neural network, dynamically adapting to the nonlinear laws of the coupler force spatial distribution. This can be expressed by the following formula:

[0068] in, Characterizes the adaptive positional features of the m-th target position and the n-th measuring point in the b-th batch. For an adaptive coding network, θ is a learnable parameter, and the measurement point location is... With the target location .

[0069] The feedforward neural network structure of this network achieves a non-linear mapping from low-dimensional position indices to high-dimensional features, and concatenates the differences and variances into two-dimensional features, which are then mapped to D-dimensional adaptive position features via an MLP network. This can be expressed by the following formula:

[0070] in, It is an adaptive positional coding function; Encode the source location; Encode the target location; The position difference between the target location and the measuring point. The sign of the position difference can be used to reflect the relationship between the target position and the measuring point position within the train formation; It is a feedforward neural network.

[0071] To adapt to the dimensions of attention scores, the high-dimensional location features are averaged along the feature dimensions and transformed into a scalar location influence factor, quantifying the locational association strength of each measurement point to the target location. This can be expressed by the following formula:

[0072] in, S represents the mean along the last dimension (feature dimension). pos This is a scalar positional influence factor.

[0073] The spatiotemporal fusion features of known measurement points are linearly transformed to generate vectors Q, K, and V. Then, the basic attention score is calculated from Q and K using matrix multiplication. Finally, the scalar position influence factor is added element-wise to the basic attention score to achieve adaptive correction of the attention score, yielding the final attention score for the fusion position prior. This can be expressed by the following formula:

[0074] Where S is the final attention score fused with location priors; Q is the query matrix; K is the key matrix; and D is the feature dimension. This is a scalar positional influence factor.

[0075] To eliminate the influence of excessively high weights on distant measurement points due to feature noise and the neglect of true nonlinear correlations in the data due to overemphasizing physical priors, the final attention score, which incorporates location priors, is normalized using softmax to generate attention weights that sum to 1. Therefore, the final attention score can be expressed by the following formula:

[0076] Where Q is the query matrix; K is the key matrix; and D is the feature dimension; The final attention score; softmax( E( is the normalization function); ,dim= 1) The mean is taken along the feature dimension.

[0077] The coupler force spatial feature interpolation model adopts a position-encoded adaptive encoding method. It adaptively captures the spatial correlation between known measurement points and target positions through an end-to-end learnable network structure and uses a learnable neural network to map it into high-dimensional dense features.

[0078] Finally, the interpolated features of the target location are obtained by weighted summation of the measurement point features using attention weights, which can be expressed by the following formula:

[0079] in, For the target position features after adaptive interpolation, bmm( ) represents batch matrix multiplication, V is the value vector, and W is the value vector. O This is for outputting the projection matrix.

[0080] The final prediction of the entire coupling force was obtained.

[0081] This application embodiment is based on a deep learning network and uses data from a limited number of measuring points to extrapolate and identify the coupler forces of other couplers, achieving online monitoring and accurate identification of coupler forces across the entire network. By acquiring coupler forces from a limited number of measuring points, this application embodiment reduces the cost of deploying sensors across the entire network to monitor coupler forces, as in traditional solutions, and improves the accuracy of coupler force monitoring.

[0082] The deep learning-based method for identifying the global coupler force of heavy-haul trains provided in this application predicts the global coupler force by passing the input limited coupler force data through a three-stage concatenation process: a temporal feature encoder, a spatial feature encoder, and a coupler force spatial feature interpolation. Specifically, the temporal feature encoder receives the collected time-series coupler force data from a limited number of measuring points and extracts multi-scale features and global long-range dependency features. The spatial feature encoder captures the spatial correlation and physical topological constraints of the coupler force. The coupler force spatial feature interpolation model accurately maps the features of the limited measuring points to the global target location features, ultimately achieving the prediction of the global coupler force.

[0083] Optionally, multi-scale one-dimensional convolution employs three parallel convolution branches to capture the instantaneous change features, short-term fluctuation features, and medium-term trend features of the time-series signal, respectively.

[0084] Optionally, the bidirectional LSTM network has a two-layer structure, with the hidden layer dimension set to half the input feature dimension, and the forward and backward outputs concatenated to restore the input feature dimension.

[0085] Optionally, the multi-head self-attention mechanism maps features to H subspaces and computes attention weights in parallel, where H is a preset positive integer.

[0086] Optionally, multi-scale spatial attention uses depthwise convolutional branches with different kernel sizes to extract measurement point features at different spatial scales in parallel, and globally models the correlation between measurement points through the attention mechanism branches. The outputs of each branch are then fused to obtain multi-scale spatial attention weighted features.

[0087] Optionally, the graph convolution adopts a two-level structure. The first level performs matrix multiplication with the spatial attention output using the dynamic adjacency matrix as the left multiplication factor to obtain preliminary fusion features. The second level concatenates the preliminary fusion features with the spatial attention output along the feature dimension. After linear layer mapping and activation function processing, the spatiotemporal fusion features of known measurement points with the same output dimension as the spatial attention output dimension are obtained.

[0088] Optionally, the contribution weight is generated by concatenating the difference and variance between the target location and the locations of each measurement point into a two-dimensional feature, mapping it into a high-dimensional feature through MLP, and taking the mean along the feature dimension to convert it into a scalar position influence factor. The scalar position influence factor is then generated by adding the basic attention score element by element and normalizing it.

[0089] Optionally, multi-scale spatial attention uses numerical clamping to constrain the relative distance difference between measurement points within a preset range to adapt to different train formation lengths.

[0090] Figure 2 A structural block diagram of a heavy-haul train full-range coupler force identification system provided in this application embodiment is shown below. Figure 2As shown, the system may include: The first extraction unit 201 is used to extract multi-scale one-dimensional convolutional features from the time-series data of the coupler force collected by the sensor, and output multi-scale local time-series features after weighted fusion by learnable weight coefficients. The data processing unit 203 is used to perform forward and backward long-range dependency modeling on temporal features using a bidirectional LSTM network, and to perform multi-subspace attention calculation by combining a multi-head self-attention mechanism with relative position encoding. The bidirectional LSTM output and the multi-head self-attention output are fused through residual connection and layer normalization, and then concatenated with multi-scale local temporal features and mapped to the target dimension. The second extraction unit 205 is used to mine the spatial relationship between measurement points through multi-scale spatial attention and correct the attention score with relative distance physical prior, constrained by the physical topology of train formation, generate a dynamic adjacency matrix based on the measurement point position index, and aggregate spatial features through graph convolution guided by the dynamic adjacency matrix to output the spatiotemporal fusion features of known measurement points. The generation unit 207 is used to generate global coupler force by adaptively allocating contribution weights based on the known spatiotemporal fusion features of the measuring points, their position coordinates, and the target position coordinates, using a spatial attention mechanism.

[0091] The system provided in this application embodiment has the same implementation principle and technical effects as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the system embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0092] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0093] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0094] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0095] This application provides a non-transitory computer-readable storage medium that stores computer instructions, which cause the computer to execute embodiments of this application. Figure 1 The method provided in the illustrated embodiment.

[0096] The aforementioned computer-readable storage medium may be any combination of one or more computer-readable media. A computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), or flash memory, optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium may be any tangible medium that contains or stores a program that may be used by or in connection with an instruction execution system, apparatus, or device.

[0097] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including—but not limited to—electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0098] The program code contained on a computer-readable medium may be transmitted using any suitable medium, including—but not limited to—wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0099] Computer program code for performing the operations of the embodiments of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0100] The foregoing has described specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0101] In the description of the embodiments of this application, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the embodiments of this application. In the embodiments of this application, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in the embodiments of this application, as well as the features of different embodiments or examples.

[0102] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of embodiments of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0103] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order according to the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0104] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0105] It should be noted that the terminals involved in the embodiments of this application may include, but are not limited to, personal computers (PCs), personal digital assistants (PDAs), wireless handheld devices, tablet computers, mobile phones, MP3 players, MP4 players, etc.

[0106] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0107] Furthermore, the functional units in the various embodiments of this application can be integrated into a single processor, or each unit can exist physically separately, or two or more units can be integrated into a single unit. The integrated units described above can be implemented in hardware or in a combination of hardware and software functional units.

[0108] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0109] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present application should be included within the scope of protection of the present application.

Claims

1. A method for identifying the coupler force across the entire range of heavy-haul trains based on deep learning, characterized in that, The construction steps are as follows: Multi-scale one-dimensional convolutional feature extraction is performed on the time-series data of coupler force collected by the sensor, and the multi-scale local time-series features are output after weighted fusion by learnable weight coefficients. A bidirectional LSTM network is used to model the forward and backward long-range dependencies of the temporal features. Multi-subspace attention is calculated by combining a multi-head self-attention mechanism with relative position encoding. The bidirectional LSTM output and the multi-head self-attention output are fused through residual connection and layer normalization, and then concatenated with the multi-scale local temporal features and mapped to the target dimension. Constrained by the physical topology of train formation, the spatial correlation between measurement points is mined through multi-scale spatial attention and the attention score is corrected by physical prior with relative distance. A dynamic adjacency matrix is ​​generated based on the measurement point position index. The spatial features are aggregated through graph convolution guided by the dynamic adjacency matrix, and the spatiotemporal fusion features of known measurement points are output. Based on the spatial attention mechanism, contribution weights are adaptively assigned according to the known spatiotemporal fusion features of the measuring points, their position coordinates, and the target position coordinates, and the global coupler force is generated by interpolation.

2. The method for identifying the full-domain coupler force of heavy-haul trains based on deep learning according to claim 1, characterized in that, The multi-scale one-dimensional convolution employs three parallel convolution branches to capture the instantaneous change characteristics, short-term fluctuation characteristics, and medium-term trend characteristics of the time-series signal, respectively.

3. The method for identifying the full-domain coupler force of heavy-haul trains based on deep learning according to claim 1, characterized in that, The bidirectional LSTM network has a two-layer structure, with the hidden layer dimension set to half the input feature dimension, and the forward output and backward output concatenated to restore the input feature dimension.

4. The method for identifying the full-domain coupler force of heavy-haul trains based on deep learning according to claim 1, characterized in that, The multi-head self-attention mechanism maps features to H subspaces and computes attention weights in parallel, where H is a preset positive integer.

5. The method for identifying the full-domain coupler force of heavy-haul trains based on deep learning according to claim 1, characterized in that, The multi-scale spatial attention uses depthwise convolution branches with different kernel sizes to extract measurement point features at different spatial scales in parallel, and globally models the correlation between measurement points through the attention mechanism branches. The outputs of each branch are then fused to obtain multi-scale spatial attention weighted features.

6. The method for identifying the full-domain coupler force of heavy-haul trains based on deep learning according to claim 1, characterized in that, The graph convolution adopts a two-level structure. The first level performs matrix multiplication with the spatial attention output using the dynamic adjacency matrix as the left multiplication factor to obtain preliminary fused features. The second stage concatenates the preliminary fusion features with the spatial attention output along the feature dimension. After linear layer mapping and activation function processing, the output dimension is the spatiotemporal fusion feature of the known measurement points that is consistent with the spatial attention output dimension.

7. The method for identifying the full-domain coupler force of heavy-haul trains based on deep learning according to claim 1, characterized in that, The contribution weight is generated by concatenating the difference and variance between the target position and each measurement point position into a two-dimensional feature, mapping it into a high-dimensional feature through MLP, and taking the mean along the feature dimension to convert it into a scalar position influence factor. The scalar position influence factor is then generated by adding the basic attention score element by element and normalizing it.

8. The method for identifying the full-domain coupler force of heavy-haul trains based on deep learning according to claim 1, characterized in that, The multi-scale spatial attention uses numerical clamping to constrain the relative distance difference between measurement points within a preset range to adapt to different train formation lengths.