Train wheelset axle temperature prediction and hierarchical early warning method and system based on recursive multivariate interaction network

Abstract

The application provides a train wheel set axle temperature prediction and hierarchical early warning method and system based on a recursive multivariate interaction network, belongs to the technical field of train state monitoring and early warning, takes the temperature data of each wheel set bearing in the same train bogie as input, mines the correlation characteristics between different monitoring quantities through multivariate interaction modeling, and combines recursive time series modeling to build the dynamic dependence relationship of bearing temperature evolution with time, realizes accurate prediction of future wheel set bearing temperature, and realizes effective discrimination of predicted temperature rise through setting of a hierarchical early warning mechanism. The application overcomes problems such as insufficient utilization of multivariate information and insufficient time dependence description, can maintain high prediction accuracy under different time spans, realizes accurate discrimination of hierarchical early warning, and has advantages such as high prediction accuracy, good early warning effect, strong robustness and wide application range compared with a traditional single network, and can provide reliable technical support for train bearing health management and operation safety.

Description

Technical Field

[0001] This invention relates to the field of train condition monitoring and early warning technology, specifically to a train wheelset axle temperature prediction and graded early warning method and system based on recursive multivariate interactive networks. Background Technology

[0002] High-speed rail operates under complex and diverse conditions. Over long periods of operation, trains are prone to wear and tear, as well as component aging, all of which pose varying degrees of safety risks. Bearings, as crucial components of the train's transmission system, including bogie axle boxes, gearboxes, and motors, play a critical role in ensuring safe train operation. However, bearings are also among the most easily damaged components during operation. Therefore, real-time monitoring of bearing status is essential for timely maintenance and repair decisions. Bearing temperature directly reflects its operating condition. Abnormal temperature rises in bearings during train operation can not only trigger emergency braking and cause economic losses, but in more serious cases, may even lead to train derailment, threatening passenger safety. Therefore, real-time prediction and tiered early warning of bearing temperature, and understanding its changing trends in advance, are of great significance for train operation maintenance and spare parts management, while also providing a reliable basis for optimizing and adjusting train operation strategies.

[0003] Train axle temperature prediction is a typical multivariate time series forecasting task. Traditional statistical modeling and empirical analysis methods have certain limitations in handling multivariate time series forecasting tasks. In contrast, deep learning technology shows a significant advantage due to its stronger feature representation capabilities. Recurrent neural networks improve the accuracy of temperature prediction by introducing time step information, and subsequently, prediction networks such as gated recurrent units, long short-term memory networks, and graph attention networks have gradually emerged. However, existing single networks still have problems that need to be solved in practical train axle temperature prediction applications: existing methods still lack an effective integration mechanism for the temperature states of different bearings within the same bogie, failing to fully construct the temperature correlation between bearings; at the same time, traditional time series forecasting models, when dealing with temperature series of different types and scales, only model from a single time dimension, failing to simultaneously capture the temporal dependence and spatial coupling characteristics of temperature changes. These shortcomings cause significant biases in the model's prediction of bearing temperature, making it difficult to predict key risk points such as sudden temperature rises and abnormal hot states in a timely manner, affecting the prediction accuracy and safety reliability in practical applications. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for predicting and classifying early warning of train wheelset axle temperature based on a recursive multivariate interactive network, so as to solve at least one of the technical problems existing in the background art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] In a first aspect, the present invention provides a method for predicting and classifying early warning of train wheelset axle temperature based on a recursive multivariate interaction network, comprising:

[0007] Obtain train speed data;

[0008] A pre-trained train wheelset axle temperature prediction grading model is used to process the acquired speed data to obtain the predicted bearing temperature at that operating speed. The train wheelset axle temperature prediction grading model includes: a multivariate dependency structure, which constructs a multivariate correlation weight matrix and obtains standardized interaction weights through normalization; a recursive multivariate interaction network, which comprehensively characterizes the complex multivariate dependencies within a time window while ensuring sufficient preservation of feature information; a spatiotemporal sequence encoding module, which simultaneously characterizes the dependencies between multivariates and their temporal evolution; a recursive time series prediction module, which progressively maps the encoded multivariate feature sequence to the predicted bearing temperature values ​​for future times; and a dual-branch fusion structure is introduced for prediction mapping to fully exploit the nonlinear relationship between time dependence and candidate features.

[0009] Based on the automatic graded early warning mechanism for predicted temperature rise, the bearing temperature prediction value is determined to exceed the corresponding threshold according to the early warning threshold set for different operating speeds. If the threshold is exceeded, an early warning is issued.

[0010] As a further limitation of the first aspect of the present invention, constructing a multivariate correlation weight matrix and obtaining standardized interaction weights through normalization processing includes: treating each time series in the dataset as a variable sequence, establishing connections between variables based on their correlations to form an interaction structure composed of a set of variables and their dependencies; and defining correlation weight elements based on the correlations between the time series. The exponentially decaying weights between time steps represent the variables. and At time step The weight of the effect.

[0011] As a further limitation of the first aspect of the present invention, the recursive multivariate interaction network, while ensuring sufficient preservation of feature information, more comprehensively characterizes the complex multivariate dependencies within a time window, including:

[0012] For the input multivariate features, calculate the i-th variable and its associated variables. The single-head attention coefficients are normalized, and the sum of all attention coefficients related to the i-th variable is normalized to ensure that the sum of the attention coefficients is 1, resulting in standardized attention weights. Through weighted summation and self-loop compensation, each temporal variable can absorb the correlation information of its neighbors while retaining its own features during updates, achieving stable and flexible temporal feature aggregation. The vectors transmitted by the single-head information in each layer are aggregated, and a multi-head attention mechanism is implemented using multiple sets of attention weights in parallel computation. This allows variables to learn neighborhood dependencies in different feature subspaces, achieving multi-scale feature aggregation. The network is enhanced from local representation to global modeling, enabling variables to aggregate features of far-distance related variables. A normalized interaction weight matrix is ​​also included. For the characteristics of the current variable Perform feature aggregation and then use the linear transformation matrix Matrix multiplication is used to map the feature space, and then features from the previous layer are fused together. ,use Adjust the dimensions to align the feature space.

[0013] As a further limitation of the first aspect of the present invention, the spatiotemporal sequence encoding module simultaneously characterizes the dependencies between multiple variables and their temporal evolution, including: performing time series encoding on the obtained variable feature matrix; and encoding the matrix into a time series matrix of length [missing information]. The time windows are stacked into a three-dimensional tensor of "time, variables and features", introducing time series dependencies as input to the next time series prediction network; the change trajectory of each variable within the time window is extracted, and the three-dimensional tensor is treated as a multivariate time series for batch input.

[0014] As a further limitation of the first aspect of the present invention, the encoded multivariate feature sequence is gradually mapped to the predicted bearing temperature value at future time by the recursive time series prediction module, so as to realize the modeling and prediction of the temperature change trend of train wheelset bearings; for each time step, the state is updated by combining the layer normalized gated loop unit, and the feature dimension of the sample at the current time step is standardized.

[0015] As a further limitation of the first aspect of the invention, a dual-branch fusion structure is introduced for prediction mapping to fully explore the nonlinear relationship between time dependence and candidate features; this structure, on the one hand, affects the final hidden state Layer normalization and nonlinear function activation are performed to extract stationary features; on the other hand, candidate hidden states are utilized. go through The activation captures its short-term dynamic changes, and the two parts are weighted and fused by linear transformation to obtain the final prediction output.

[0016] Secondly, the present invention provides a train wheelset axle temperature prediction and hierarchical early warning system based on a recursive multivariate interaction network, comprising:

[0017] The acquisition module is used to acquire train speed data;

[0018] The prediction module processes the acquired speed data using a pre-trained train wheelset axle temperature prediction grading model to obtain the predicted bearing temperature value at that operating speed. The train wheelset axle temperature prediction grading model includes: a multivariate dependency structure, which constructs a multivariate correlation weight matrix and obtains standardized interaction weights through normalization; a recursive multivariate interaction network, which comprehensively characterizes the complex multivariate dependencies within a time window while ensuring sufficient preservation of feature information; a spatiotemporal sequence encoding module, which simultaneously characterizes the dependencies between multivariates and their temporal evolution; a recursive time series prediction module, which progressively maps the encoded multivariate feature sequence to the predicted bearing temperature value at future times; and a dual-branch fusion structure is introduced for prediction mapping to fully exploit the nonlinear relationship between time dependence and candidate features.

[0019] The early warning module is used for an automatic graded early warning mechanism based on predicted temperature rise. It determines whether the predicted bearing temperature exceeds the corresponding threshold according to the early warning threshold set for different operating speeds. If the threshold is exceeded, an early warning is issued.

[0020] Thirdly, the present invention provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the train wheelset axle temperature prediction and graded early warning method based on a recursive multivariate interactive network as described in the first aspect.

[0021] Fourthly, the present invention provides a computer device including a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the train wheelset axle temperature prediction and graded early warning method based on recursive multivariate interactive network as described in the first aspect.

[0022] Fifthly, the present invention provides an electronic device, comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the train wheelset axle temperature prediction and graded early warning method based on recursive multivariate interactive networks as described in the first aspect.

[0023] The beneficial effects of this invention are as follows: By combining multivariate interactive modeling with a time-series prediction network, it can simultaneously characterize the correlation between bearing temperature data and their time-dependent evolution, and provide tiered early warnings based on the prediction results. Experimental results show that, compared with a single prediction network, the prediction accuracy of this method is improved by more than 15% in temperature prediction. Comparative evaluation of the prediction effects across different time spans shows that this method exhibits good robustness and generalization ability in various scenarios. Furthermore, by establishing a tiered early warning mechanism, prediction results across different time spans can be graded and judged, enabling timely early warning of bearing overheating risks. This invention can more accurately characterize the temperature change trend of train wheelset bearings and provide timely early warnings, offering more reliable technical support for safe train operation.

[0024] The advantages of additional aspects of the invention will be set forth more clearly in the following description or will be learned by practice of the invention. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart of the train wheelset axle temperature prediction and graded early warning method based on recursive multivariate interactive network as described in an embodiment of the present invention.

[0027] Figure 2 This is a framework diagram of the train wheelset axle temperature prediction and graded early warning method based on recursive multivariate interactive network as described in an embodiment of the present invention.

[0028] Figure 3 This is a graph showing the train axle temperature data according to an embodiment of the present invention.

[0029] Figure 4 This is a graph showing the prediction results for a short time series dataset as described in an embodiment of the present invention.

[0030] Figure 5 This is a graph showing the prediction results for a short time series dataset as described in an embodiment of the present invention.

[0031] Figure 6 This is a comparison chart of the prediction results of this method and other methods on a long-term series dataset as described in this embodiment of the invention.

[0032] Figure 7 This is a schematic diagram of the graded early warning results under different time spans as described in the embodiments of the present invention. Detailed Implementation

[0033] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0034] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0035] It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as here.

[0036] To facilitate understanding of the present invention, the present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. However, the specific embodiments do not constitute a limitation on the embodiments of the present invention.

[0037] Those skilled in the art should understand that the accompanying drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily essential for implementing the present invention.

[0038] This invention proposes a method for predicting and classifying the temperature of train wheelset axles based on a recursive multivariate interaction network. For temperature prediction, multivariate interaction modeling is combined with a time-series prediction network. First, the recursive multivariate interaction network is used to weight and fuse different wheelset bearings and their related monitoring quantities within the same bogie, extracting multivariate features characterizing the state transmission and coupling relationships between bearings. Then, these features are input into a time-series encoding module, introducing the dependency relationship of temperature evolution over time, completing the joint modeling of multivariate state information and time-varying information. Finally, the encoded time-series features are input into the time-series prediction network to obtain the temperature prediction results of the train wheelset bearings. Based on this, a classifying early warning mechanism is constructed, using the difference between the predicted bearing temperature and the outdoor temperature as the predicted temperature rise, and combining the mean and standard deviation of the temperature rise under different speed conditions to set multi-level thresholds, thus achieving a classifying early warning of bearing overheating risk. This invention effectively mines the correlation and interaction information between multiple variables while fully preserving the long-term dependency relationship of the temperature sequence, and achieves bearing operation risk assessment based on the predicted temperature rise, providing maintenance personnel with operable safety early warning and decision-making basis.

[0039] Example 1

[0040] In this embodiment 1, a train wheelset axle temperature prediction and grading early warning system based on a recursive multivariate interaction network is first provided, including: an acquisition module for acquiring train running speed data; a prediction module for processing the acquired speed data using a pre-trained train wheelset axle temperature prediction and grading model to obtain the predicted bearing temperature value at that running speed; wherein, the train wheelset axle temperature prediction and grading model includes: a multivariate dependency structure, constructing a multivariate correlation weight matrix and obtaining standardized interaction weights through normalization processing; a recursive multivariate interaction network, which more comprehensively characterizes the complex multivariate dependencies within a time window while ensuring sufficient preservation of feature information; a spatiotemporal sequence encoding module, which simultaneously characterizes the dependencies between multivariates and their evolution over time; a recursive time series prediction module, which gradually maps the encoded multivariate feature sequence to the predicted bearing temperature value at future times; and a dual-branch fusion structure is introduced for prediction mapping to fully explore the nonlinear relationship between time dependency and candidate features. An early warning module is used for an automatic grading early warning mechanism based on predicted temperature rise, which determines whether the predicted bearing temperature value exceeds the corresponding threshold according to the set early warning threshold at different running speeds, and if it exceeds the threshold, performs early warning processing.

[0041] In this embodiment, a train wheelset axle temperature prediction and graded early warning method is implemented based on the above system. The method includes: acquiring train operating speed data using an acquisition module; processing the acquired speed data using a pre-trained train wheelset axle temperature prediction and graded model using a prediction module to obtain the predicted bearing temperature value at that operating speed; wherein the train wheelset axle temperature prediction and graded model includes: a multivariate dependency structure, constructing a multivariate correlation weight matrix and obtaining standardized interaction weights through normalization; a recursive multivariate interaction network, which comprehensively characterizes the complex multivariate dependencies within a time window while ensuring sufficient preservation of feature information; a spatiotemporal sequence encoding module, which simultaneously characterizes the dependencies between multiple variables and their temporal evolution; a recursive time series prediction module, which gradually maps the encoded multivariate feature sequence to the predicted bearing temperature value at future times; and introducing a dual-branch fusion structure for prediction mapping to fully explore the nonlinear relationship between time dependency and candidate features. Finally, an early warning module, based on an automatic graded early warning mechanism for predicted temperature rise, determines whether the predicted bearing temperature value exceeds the corresponding threshold according to the set early warning thresholds for different operating speeds. If the threshold is exceeded, an early warning is issued.

[0042] The process involves constructing a multivariate association weight matrix and obtaining standardized interaction weights through normalization. This includes treating each time series in the dataset as a variable sequence, establishing connections between variables based on their correlations to form an interaction structure composed of the variable set and its dependencies; and defining association weight elements based on the correlations between the time series. The exponentially decaying weights between time steps represent the variables. and At time step The weight of the effect.

[0043] Specifically, each time series in the dataset is first treated as a sequence of variables. Connections are established between these variables based on their correlations, forming an interaction structure composed of the variable set and its dependencies. The structure is defined as follows: ,in Represents a set of variables. Indicates the connection relationship between variables. The feature sequence is the sequence of time points in the dataset, where each time step is represented as... , Indicates at time step The characteristic matrix of the time variable, For each time step Feature dimensions, variable set If two variables are correlated at a certain time step, it is denoted as... Dependencies are determined by the correlation between variables:

[0044] ;

[0045] Secondly, the multivariate correlation weight matrix is ​​adaptively constructed and normalized. The specific process includes three parts: correlation weight calculation, weighting matrix acquisition, and symmetric normalization. First, the correlation weight elements are defined based on the correlation between time series. The exponentially decaying weights between time steps represent the variables. and At time step Effect weight:

[0046]

[0047] in To adjust the coefficients and control the smoothness of the weight distribution among variables, a weighted metric matrix is ​​introduced to measure the "overall correlation" between each time step and its adjacent time steps. The correlation weight matrix is ​​then symmetrically normalized to obtain a standardized interaction weight matrix.

[0048]

[0049] ( )

[0050] in Representing variables Total connection strength, This indicates the generation of a diagonal matrix where only the elements on the main diagonal are non-zero, and all other elements are 0. Preserve matrix symmetry.

[0051] Recursive multivariate interaction networks, while ensuring sufficient preservation of feature information, more comprehensively characterize complex multivariate dependencies within a time window, including: for input multivariate features, calculating the i-th variable and its associated variables. The single-head attention coefficients are normalized, and the sum of all attention coefficients related to the i-th variable is normalized to ensure that the sum of the attention coefficients is 1, resulting in standardized attention weights. Through weighted summation and self-loop compensation, each temporal variable can absorb the correlation information of its neighbors while retaining its own features during updates, achieving stable and flexible temporal feature aggregation. The vectors transmitted by the single-head information in each layer are aggregated, and a multi-head attention mechanism is implemented using multiple sets of attention weights in parallel computation. This allows variables to learn neighborhood dependencies in different feature subspaces, achieving multi-scale feature aggregation. The network is enhanced from local representation to global modeling, enabling variables to aggregate features of far-distance related variables. A normalized interaction weight matrix is ​​also included. For the characteristics of the current variable Perform feature aggregation and then perform linear transformation matrix Matrix multiplication is used to map the feature space, and then features from the previous layer are fused together. ,use Adjust the dimensions to align the feature space.

[0052] Specifically, a recursive multivariate interaction network is constructed to more comprehensively characterize complex multivariate dependencies within a time window while ensuring sufficient preservation of feature information, including:

[0053] S2-1, Input Multivariate Features , ∈ , It is the number of variables. Indicates the dimension of the input features. This represents the dimension of the output feature, and the corresponding feature vector of its variable is... , ∈ . No. Variables and their related variables The single-head attention coefficient is calculated as follows:

[0054] ;

[0055] in It is a variable For variables Attention weight coefficients; by implementing a masking attention mechanism, they are calculated only between pre-determined pairs of variables that are correlated. Ignore irrelevant or weakly correlated variable combinations. Introduce learnable bias terms. Enhance the ability to distinguish between variables. This is a normalization term to prevent gradient explosion and make interactive feature modeling more efficient and stable.

[0056] S2-2, Attention weight normalization, for the first... All attention coefficients related to each variable are normalized to ensure that the sum of the attention coefficients is 1, resulting in standardized attention weights:

[0057]

[0058] in The standard attention mechanism obtained through softmax is improved by introducing a regularization balance term. Smooth attention distribution to avoid gradients that are too large or too small, thereby increasing network stability and generalization ability.

[0059] S2-3, Attention-based Neighborhood Information Fusion. Through weighted summation and self-loop compensation, each time variable can absorb the correlation information of its neighbors while retaining its own characteristics during the update, achieving stable and flexible aggregation of temporal features.

[0060]

[0061] in It is an activation function that maintains linearity when the input is greater than 0 and sets it to zero when the input is less than 0. In order to be with the first A set of variables related to a variable. This is the weight matrix calculated in the previous step. Each is the current variable and related variables Input features and variables The aggregated output features and Each is the current variable and related variables The linear transformation matrix, These are self-loop weighting coefficients used to control the proportion of a variable's own features in the aggregation, so that it can retain its own feature information.

[0062] S2-4. Aggregate the vectors transmitted by the single-head information in each layer, and use multiple sets of attention weights to perform parallel computation to implement a multi-head attention mechanism. This enables variables to learn neighborhood dependencies in different feature subspaces, achieving multi-scale feature aggregation.

[0063]

[0064] in, The number of multi-headed attention, For the first Variables under the attention mechanism For variables Attention weights The first Feature transformation matrix and self-loop transformation matrix of each attention head Corresponding to the The self-loop fusion coefficient of each attention head This represents vector concatenation, which joins multiple results to form a high-dimensional representation. This is the activation function.

[0065] S2-5, Multi-layer Recursive Propagation and Residual Enhancement. This step enhances the network's features from local representations to global modeling, allowing variables to aggregate features from distantly related variables. Normalized interaction weight matrix. For the characteristics of the current variable Perform feature aggregation and then perform linear transformation matrix Matrix multiplication is used to map the feature space, and then features from the previous layer are fused together. ,use Adjusting the dimensions and aligning the feature space, the specific formula is as follows:

[0066]

[0067] in This indicates the intensity of residual injection, which ultimately integrates the current layer propagation features with historical layer residual information and outputs higher-order features after nonlinear activation.

[0068] The spatiotemporal sequence encoding module simultaneously characterizes the dependencies between multiple variables and their temporal evolution, including: performing time series encoding on the obtained variable feature matrix; and encoding a sequence of length... The time windows are stacked into a three-dimensional tensor of "time, variables and features", introducing time series dependencies as input to the next time series prediction network; the change trajectory of each variable within the time window is extracted, and the three-dimensional tensor is treated as a multivariate time series for batch input.

[0069] Specifically, time-series encoding is performed on the obtained variable feature matrix. The current feature dimension is... , For the number of variables, The feature dimension for each variable. (The length is...) The time windows are stacked into a three-dimensional tensor of "time, variable, and feature", introducing time series dependencies as input to the next step of the time series prediction network:

[0070]

[0071] in This represents a three-dimensional tensor of the current time window. This indicates a stacking operation, which stacks the features from each past moment in chronological order to form a temporal feature block. Then, the trajectory of each variable's change within the time window is extracted, treating the 3D tensor as a multivariate time series for batch input.

[0072]

[0073]

[0074] in Indicates the first The variables at time... eigenvectors, For time series samples, It is a collection of multivariate time series.

[0075] The recursive time series prediction module gradually maps the encoded multivariate feature sequence to the predicted bearing temperature value at future time, thereby realizing the modeling and prediction of the temperature change trend of train wheelset bearings. For each time step, the state is updated by combining the layer normalized gated loop unit, and the feature dimension of the sample at the current time step is standardized.

[0076] Specifically, a recursive time-series prediction module progressively maps the encoded multivariate feature sequence to predicted bearing temperatures for future times, enabling modeling and prediction of the temperature variation trend of train wheelset bearings. For each time step, a gated recurrent unit with layer normalization is used for state updates, and the feature dimensions of the samples at the current time step are standardized. Layer normalization , Calculate the mean of the input vector, and calculate the denominator. Standard deviation, scale normalization To use a smaller value to prevent the denominator from being 0, These are the learnable scaling and translation coefficients, used to recover the optimal distribution scale required by the network and adjust the mean. The specific formulas are as follows:

[0077]

[0078]

[0079]

[0080]

[0081] in Controlling the degree to which the network forgets the past, Controlling the ratio of historical information to candidate new information, This is a candidate hidden state. For variables At time step The hidden state; These are all learnable parameters of the network. Representation layer normalization, It is a non-linear activation function. The hyperbolic tangent function maps the input to the interval (−1, 1). Compared to the zero-centered sigmoid function, training is more stable. This indicates element-wise multiplication.

[0082] A dual-branch fusion structure is introduced for prediction mapping to fully explore the nonlinear relationship between time dependence and candidate features; this structure, on the one hand, affects the final hidden state. Layer normalization and nonlinear function activation are performed to extract stationary features; on the other hand, candidate hidden states are utilized. go through The activation captures its short-term dynamic changes, and the two parts are weighted and fused by linear transformation to obtain the final prediction output.

[0083] Specifically, a dual-branch fusion structure is introduced for prediction mapping to fully explore the nonlinear relationship between time dependence and candidate features. This structure, on the one hand, affects the final hidden state. Layer normalization and nonlinear function activation are performed to extract stationary features; on the other hand, candidate hidden states are utilized. go through The activation captures its short-term dynamic changes, and the two parts are then weighted and fused through a linear transformation to obtain the final prediction output:

[0084]

[0085] in These are the linear mapping matrices for the final hidden state and the candidate hidden states, respectively. These are the bias terms for the two branches. Describe the output layer weight matrix. This represents the bias term of the output layer. For output dimensions, This represents the network's predicted output.

[0086] Construct an automatic graded early warning mechanism for predicted temperature rise, and set early warning thresholds for different operating speeds, specifically including:

[0087] S6-1. Set filtering conditions. Since the absolute temperature of the bearing is greatly affected by factors such as ambient temperature and seasonal changes, to improve the stability of the early warning threshold, the difference between the bearing temperature and the outdoor temperature is used as the temperature rise index. ,calculate At this moment, the first Axle temperature rise of vehicle number 1 relative to outdoor temperature , For actual measured temperature, The outdoor temperature is used as the reference point; then, the operating conditions are divided according to the train's operating speed, i.e. At the same time, a uniform velocity condition is set, and at a speed of... The length of the center is time window Within, calculate the difference between the maximum and minimum running speeds. ,determination Less than the threshold At that time, the velocity is constant. The specific formula is as follows:

[0088]

[0089]

[0090] S6-2. Establish train temperature rise data at different operating speeds. Select time sets for different train sets that meet the speed range and have small fluctuations, and then analyze each time set... Calculate its maximum temperature rise across all vehicles. If the maximum temperature rise does not exceed the upper limit This means that the temperature rise at that moment is considered a healthy sample, as shown in the following formula:

[0091]

[0092] For each moment of the healthy sample Calculate the median temperature rise for each axle. This serves as a representative value for temperature rise at that moment, while also requiring that the absolute deviation between the actual temperature rise of each train set and the median does not exceed [a certain value]. Otherwise, a second round of elimination is performed. Taking into account both speed and temperature rise constraints, the final set of time periods considered to be under healthy operating conditions is obtained. This is used for subsequent grading threshold calculations. Simultaneously, the time set is divided according to different speeds, i.e.

[0093]

[0094] S6-3, Multi-level early warning threshold calculation and predicted temperature rise determination. (Regarding operating speed...) ,calculate Mean temperature rise in the next time series and standard deviation :

[0095]

[0096]

[0097] in The total number of samples under these operating conditions. For working conditions Next Temperature rise data at any given time. (Using...) Criteria, in operating conditions Built-in three-level early warning thresholds:

[0098]

[0099] Calculate the temperature rise of the S5 prediction results , ,in That is, time. Based on the prediction results, the warning level at the current moment is defined as follows: 0 represents normal situation; 1 represents Level 1 warning; 2 represents Level 2 warning; and 3 represents Level 3 warning. The classification is determined as follows:

[0100] .

[0101] Example 2

[0102] like Figure 1 , Figure 2 As shown in this embodiment, a graded early warning method for predicting train wheelset axle temperature based on a recursive multivariate interaction network is provided, which specifically includes the following steps:

[0103] S1. Perform time series reconstruction and sampling consistency on the original data, and divide the dataset. This includes the following steps:

[0104] S1-1. Time Series Reconstruction and Sampling Consistency. The first step is to fill in the blank data. The original data contained multiple blank data points, causing breaks in the data chain and incomplete information. Furthermore, the data collection time intervals were inconsistent, and timestamps could not be synchronized, affecting the accuracy of the data and the effectiveness of subsequent analysis. Therefore, it is necessary to remove abnormal time intervals and adjust the time intervals to ensure that data is collected at 30-second intervals. Simultaneously, duplicate data must be removed. For duplicate timestamps and duplicate axis temperature data columns, the data structure needs to be reorganized to ensure data accuracy.

[0105] 1) Missing values ​​are filled using linear interpolation. By performing linear interpolation between adjacent observation points, the broken time-numerical chain is restored, ensuring that subsequent resampling and statistical summarization are performed on the "complete trajectory," avoiding bias caused by gaps. The specific formula is shown in Example 1. Based on this, time interval adjustment and standardization are performed to place the inconsistent timestamps on the same grid, eliminating the discretization error introduced by unequal sampling. Consistent with linear interpolation, the measured values ​​are time-aligned without changing the meaning of the physical quantities. For the original time series By adjusting the resampling function This fixes the sampling interval to be [value]. . Aggregate resampling can also be performed, recording the Kth time window. its internal observation set The following are the formulas for mean aggregation and median aggregation, respectively:

[0106]

[0107] 2) Abnormal time interval filtering, defining timestamp differences:

[0108]

[0109]

[0110] in , For adjacent time intervals, if their time intervals are... Then the index This is recorded as an anomaly, among which This refers to the tolerance for recording jitter. Abnormal acquisition is identified by characterizing excessively dense, sparse, and abrupt time segments within the recording time interval. This is achieved through... By comparing the period-doubling relationships, both single-step deviations and multi-cycle jumps can be detected. Specifically, if If the sampling is too dense, it is mostly repetitive or jittery, and needs to be deduplicated or merged; if If the sampling is too sparse, it is considered a missing segment and requires interpolation during the resampling stage. Interpolation is prohibited for excessively long gaps that exceed the threshold, and gaps are retained.

[0111] 3) Duplicate value removal and consistency reconstruction. For the same timestamp There are multiple records below By compressing multiple readings at the same time into a unique representative value, "multi-value conflicts" caused by repeated writing at the acquisition layer are eliminated, thus ensuring time alignment and index uniqueness. The specific formula is as follows:

[0112]

[0113] S1-2, Prepare the dataset. Process the dataset in batches. Set the feature values ​​of 2 rows as a group. Set as The 20th line is used for validation, forming the first set of data. The vehicle group data is then processed in batches and used as the dataset for subsequent network training. The specific calculation formula is as follows:

[0114]

[0115]

[0116] Let the original sequence be , in length Using a sliding window of value 20 as input, the corresponding output... Take the corresponding Then for the first The data from the rows is used to obtain the training sample set.

[0117]

[0118] in The total number of sample groups, These represent the first group, the second group, and so on, extracted from the original sequence, up to the [number missing]th group. Group data.

[0119] S1-3. Split the dataset. Divide the preprocessed data into a training set and a test set. The test set is further divided into short-time datasets and long-time datasets based on different time spans, which are used to evaluate the model's prediction accuracy and generalization ability in different prediction scenarios.

[0120] S2. The definition of a multivariate dependency structure includes the following steps:

[0121] S2-1. Treat each time series in the dataset as a variable sequence, and establish connections between variables based on their correlations to form an interactive structure composed of the variable set and its dependencies. See Example 1 for details.

[0122] S2-2, Adaptive Construction and Normalization of Multivariate Association Weight Matrix. The specific process includes three parts: association weight calculation, weighting matrix acquisition, and symmetric normalization. First, based on the correlation between time series, the association weight elements are defined. The exponentially decaying weights between time steps represent the variables. and At time step The role weights are determined by introducing a weighted metric matrix to measure the "overall correlation" between each time step and its adjacent time steps, and the correlation weight matrix is ​​symmetrically normalized to obtain a standardized interaction weight matrix.

[0123] S3. Construct a recursive multivariate interaction network, with the following specific steps:

[0124] S3-1. Input multivariate features. The i-th variable and its associated variables. The method for calculating the single-head attention coefficient is shown in Example 1.

[0125] S3-2, Attention weight normalization, for the first... All attention coefficients related to each variable are normalized to ensure that the sum of the attention coefficients is 1, thus obtaining the standardized attention weights.

[0126] S3-3, Attention-based Neighborhood Information Fusion. Through weighted summation and self-loop compensation, each time variable can absorb the correlation information of its neighbors while retaining its own characteristics during the update, achieving stable and flexible aggregation of temporal features.

[0127] S3-4. Aggregate the vectors transmitted by the single-head information of each layer, and use multiple sets of attention weights to calculate in parallel to realize the multi-head attention mechanism, so that variables can learn neighborhood dependencies in different feature subspaces and realize multi-scale feature aggregation.

[0128] S3-5, Multi-layer Recursive Propagation and Residual Augmentation. This step enhances the network's features from local representations to global modeling, allowing variables to aggregate features from distantly related variables. Normalized interaction weight matrix. For the characteristics of the current variable Perform feature aggregation and then use the linear transformation matrix Matrix multiplication is used to map the feature space, and then features from the previous layer are fused together. ,use Adjust the dimensions and align the feature space, as shown in Example 1.

[0129] S4. Spatiotemporal sequence coding module: After multiple layers of recursive multivariate interaction network, the resulting variable feature matrix can only reflect the information at the current moment and fails to reflect the dynamic characteristics of the variables evolving over time. Therefore, time series coding is required. The specific steps are shown in Example 1.

[0130] S5. The encoded multivariate feature sequence is gradually mapped to the predicted bearing temperature value at future time through the recursive time series prediction module, so as to realize the modeling and prediction of the temperature change trend of train wheelset bearings. The specific steps are shown in Example 1.

[0131] S6. The variable features obtained through the recursive temporal prediction module have fully integrated spatial dependency and temporal evolution information. However, the hidden state still needs to be further mapped to a specific prediction target space. Therefore, in order to fully explore the nonlinear relationship between temporal dependency and candidate features, a dual-branch fusion structure is introduced for prediction mapping. This structure, on the one hand, maps the final hidden state... Layer normalization and nonlinear function activation are performed to extract stationary features; on the other hand, candidate hidden states are utilized. go through The activation captures its short-term dynamic changes, and the two parts are weighted and fused by linear transformation to obtain the final prediction output.

[0132] In this embodiment, an automatic graded early warning mechanism for predicted temperature rise is constructed. The specific steps are as shown in Embodiment 1 and will not be repeated here.

[0133] like Figures 3 to 7 As shown, in this embodiment, the data used is the bogie bearing data of a certain EMU train G1-16 car from July 1, 2023 to August 29, 2023, a total of 60 days. The temporal data of 12 features of each train were measured by temperature sensors installed on each axle box of the EMU, as shown in Table 1 below.

[0134] Table 1. Temporal Characteristics of Experimental Data

[0135]

[0136] In the specific data analysis and network verification, 8 sets of bogie bearing temperature data were mainly used, i.e., the feature number is 8. Examples of train axle temperature data are shown in Table 2 below. Figure 3 As shown.

[0137] Table 2. Examples of Experimental Data

[0138]

[0139] In the specific experimental process, train G16 was divided into a training set and a test set. In order to fully demonstrate the effectiveness of the invention method, a long-term state prediction task and a short-term state prediction task were constructed, which are described in detail below: the axle temperature data of cars 1-3 is used as the training set, and the test set is divided into long-term and short-term parts: the short-term test set selects the axle temperature data of cars 4, 5, and 6 for 3 days, and the long-term test set selects the data of cars 7, 8, and 9 for 7 days. The specific time range and data scale are shown in Table 3 below.

[0140] Table 3. Training and testing dataset sizes

[0141]

[0142] To quantify the performance of different networks on the test and training sets, and to measure the network's generalization ability and prediction accuracy, commonly used prediction evaluation metrics, namely Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Root Mean Square Error (RMSE), are used to evaluate the network's prediction performance.

[0143] a) MAE measures the absolute difference between predicted and actual values ​​across each dimension, and then averages the results. This metric provides a clear picture of the prediction error level and is highly robust to outliers.

[0144]

[0145] in For the first line, number The predicted value of the column, For the first line, number The true value of the column, and These represent the number of rows and columns, respectively.

[0146] b) MAPE measures the proportion of prediction error to the true value, providing more comparable results across data of different dimensions and scales:

[0147]

[0148] c) RMSE squares the error term during calculation, thus amplifying the impact of large biased samples. It can more sensitively reflect larger errors in prediction and is an important indicator when emphasizing network robustness.

[0149]

[0150] The experiment used the method proposed in this invention to predict the temperature of rolling bearings, and compared it with three commonly used time-series prediction methods: Gated Recurrent Unit (GRU), Bidirectional Long Short-Term Memory Network (BiLSTM), and Graph Attention Network (GAT). Parameter settings are shown in Table 4.

[0151] Table 4. Parameter Settings

[0152]

[0153] 1) Shaft temperature prediction results: Substitute the experimental data into the network and comparison network of this embodiment, and compare the average value of three indicators of eight feature values. The experimental results are shown in Tables 5 and 6.

[0154] Table 5. Short-term forecast comparison results

[0155]

[0156] Table 6. Comparison Results of Long-Term Predictions

[0157]

[0158] Experimental results show that our proposed method significantly outperforms single-contrast networks in shaft temperature prediction tasks across different time spans, effectively characterizing complex patterns and multiple dependencies in sequences. Furthermore, the network advantage is more pronounced under long-term sequence conditions, likely due to the richer temporal dependencies and latent structural features contained in long sequences, which facilitates the network's accumulation and utilization of cross-scale information. By jointly modeling multivariate interaction information and temporal features, our method can more fully capture long-term dependencies, significantly improving the accuracy and reliability of shaft temperature prediction under various prediction step sizes and data distributions.

[0159] 2) Automatic graded early warning for predicted temperature rise

[0160] First, a temperature rise dataset under healthy conditions was constructed based on train operating speed and temperature rise constraints. Within a load range of 480-550 kN, a period of uniform speed operation with minimal speed fluctuations was selected within a certain time window, and abnormal data with axle temperature rise exceeding the upper limit were removed. Simultaneously, a horizontal comparison was made using the axle temperature rise of cars 1–16 within the same time period. Using the median temperature rise of multiple train sets as a reference, data deviating more than 10°C from the median were considered suspicious anomalies and removed a second time, thus constructing the temperature rise dataset. Based on this, the mean and standard deviation at different operating speeds were calculated to construct a tiered early warning threshold. The calculation results are shown in Table 7.

[0161] Table 7. Calculation results of warning thresholds at different operating speeds

[0162]

[0163] The short-term prediction results for car 5 and the long-term prediction results for car 7 were substituted into the constructed temperature rise graded early warning threshold discrimination mechanism. Based on the train's operating speed, the predicted temperature rise data for car 5 under high-speed conditions was selected for early warning discrimination, while the predicted temperature rise data for car 7 under low-speed conditions was selected. By classifying and judging the above two types of prediction results under the early warning threshold system corresponding to different speed conditions, the applicability and effectiveness of this method under different operating conditions were verified. The discrimination results show that as the train runs, the predicted axle temperature rise of both cars 5 and 7 triggers a level-one early warning.

[0164] Example 3

[0165] This embodiment 3 provides a non-transitory computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, they implement the train wheelset axle temperature prediction and graded early warning method based on a recursive multivariate interaction network as described above. The method includes:

[0166] Obtain train speed data;

[0167] A pre-trained train wheelset axle temperature prediction grading model is used to process the acquired speed data to obtain the predicted bearing temperature at that operating speed. The train wheelset axle temperature prediction grading model includes: a multivariate dependency structure, which constructs a multivariate correlation weight matrix and obtains standardized interaction weights through normalization; a recursive multivariate interaction network, which comprehensively characterizes the complex multivariate dependencies within a time window while ensuring sufficient preservation of feature information; a spatiotemporal sequence encoding module, which simultaneously characterizes the dependencies between multivariates and their temporal evolution; a recursive time series prediction module, which progressively maps the encoded multivariate feature sequence to the predicted bearing temperature values ​​for future times; and a dual-branch fusion structure is introduced for prediction mapping to fully exploit the nonlinear relationship between time dependence and candidate features.

[0168] Based on the automatic graded early warning mechanism for predicted temperature rise, the bearing temperature prediction value is determined to exceed the corresponding threshold according to the early warning threshold set for different operating speeds. If the threshold is exceeded, an early warning is issued.

[0169] Example 4

[0170] This embodiment 4 provides a computer device, including a memory and a processor, wherein the processor and the memory communicate with each other, and the memory stores program instructions that can be executed by the processor. The processor calls the program instructions to execute the train wheelset axle temperature prediction and graded early warning method based on a recursive multivariate interaction network as described above. The method includes:

[0171] Obtain train speed data;

[0172] A pre-trained train wheelset axle temperature prediction grading model is used to process the acquired speed data to obtain the predicted bearing temperature at that operating speed. The train wheelset axle temperature prediction grading model includes: a multivariate dependency structure, which constructs a multivariate correlation weight matrix and obtains standardized interaction weights through normalization; a recursive multivariate interaction network, which comprehensively characterizes the complex multivariate dependencies within a time window while ensuring sufficient preservation of feature information; a spatiotemporal sequence encoding module, which simultaneously characterizes the dependencies between multivariates and their temporal evolution; a recursive time series prediction module, which progressively maps the encoded multivariate feature sequence to the predicted bearing temperature values ​​for future times; and a dual-branch fusion structure is introduced for prediction mapping to fully exploit the nonlinear relationship between time dependence and candidate features.

[0173] Based on the automatic graded early warning mechanism for predicted temperature rise, the bearing temperature prediction value is determined to exceed the corresponding threshold according to the early warning threshold set for different operating speeds. If the threshold is exceeded, an early warning is issued.

[0174] Example 5

[0175] This embodiment 5 provides an electronic device, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes instructions to implement the train wheelset axle temperature prediction and graded early warning method based on recursive multivariate interactive networks as described above. The method includes:

[0176] Obtain train speed data;

[0177] A pre-trained train wheelset axle temperature prediction grading model is used to process the acquired speed data to obtain the predicted bearing temperature at that operating speed. The train wheelset axle temperature prediction grading model includes: a multivariate dependency structure, which constructs a multivariate correlation weight matrix and obtains standardized interaction weights through normalization; a recursive multivariate interaction network, which comprehensively characterizes the complex multivariate dependencies within a time window while ensuring sufficient preservation of feature information; a spatiotemporal sequence encoding module, which simultaneously characterizes the dependencies between multivariates and their temporal evolution; a recursive time series prediction module, which progressively maps the encoded multivariate feature sequence to the predicted bearing temperature values ​​for future times; and a dual-branch fusion structure is introduced for prediction mapping to fully exploit the nonlinear relationship between time dependence and candidate features.

[0178] Based on the automatic graded early warning mechanism for predicted temperature rise, the bearing temperature prediction value is determined to exceed the corresponding threshold according to the early warning threshold set for different operating speeds. If the threshold is exceeded, an early warning is issued.

[0179] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, 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.

[0180] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), 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.

[0181] 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.

[0182] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed 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.

[0183] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solutions disclosed in the present invention, various modifications or variations that can be made by those skilled in the art without creative effort should be included within the scope of protection of the present invention.

Claims

1. A graded early warning method for predicting train wheelset axle temperature based on a recursive multivariate interactive network, characterized in that, include: Obtain train speed data; A pre-trained train wheelset axle temperature prediction grading model is used to process the acquired speed data to obtain the predicted bearing temperature value at the operating speed. The train wheelset axle temperature prediction grading model includes: a multivariate dependency structure, which constructs a multivariate correlation weight matrix and obtains standardized interaction weights through normalization; a recursive multivariate interaction network, which comprehensively characterizes the complex multivariate dependencies within a time window while ensuring sufficient preservation of feature information; a spatiotemporal sequence encoding module, which simultaneously characterizes the dependencies between multivariates and their temporal evolution; a recursive time series prediction module, which progressively maps the encoded multivariate feature sequence to the predicted bearing temperature value at future times; and a dual-branch fusion structure is introduced for prediction mapping to fully explore the nonlinear relationship between time dependency and candidate features. The recursive multivariate interaction network, which comprehensively characterizes the complex multivariate dependencies within a time window while ensuring sufficient preservation of feature information, includes: for the input multivariate features, calculating the i-th variable and its associated variables. The single-head attention coefficients are normalized, and the sum of all attention coefficients related to the i-th variable is normalized to ensure that the sum of the attention coefficients is 1, resulting in standardized attention weights. Through weighted summation and self-loop compensation, each temporal variable can absorb the correlation information of its neighbors while retaining its own features during updates, achieving stable and flexible temporal feature aggregation. The vectors transmitted by the single-head information in each layer are aggregated, and a multi-head attention mechanism is implemented using multiple sets of attention weights in parallel computation. This allows variables to learn neighborhood dependencies in different feature subspaces, achieving multi-scale feature aggregation. The network is enhanced from local representation to global modeling, enabling variables to aggregate features of far-distance related variables. A normalized interaction weight matrix is ​​also included. For the characteristics of the current variable Perform feature aggregation and then use the linear transformation matrix Matrix multiplication is used to map the feature space, and then features from the previous layer are fused together. ,use Adjust dimensions to align the feature space; Based on the automatic graded early warning mechanism for predicted temperature rise, the bearing temperature prediction value is determined to exceed the corresponding threshold according to the early warning threshold set for different operating speeds. If the threshold is exceeded, an early warning is issued.

2. The method for predicting and classifying early warning of train wheelset axle temperature based on recursive multivariate interactive networks according to claim 1, characterized in that, Constructing a multivariate association weight matrix and obtaining standardized interaction weights through normalization includes: treating each time series in the dataset as a variable sequence, establishing connections between variables based on their correlations to form an interaction structure composed of the variable set and its dependencies; and defining association weight elements based on the correlations between time series. The exponentially decaying weights between time steps represent the variables. and At time step The weight of the effect.

3. The method for predicting and classifying early warning of train wheelset axle temperature based on recursive multivariate interactive networks according to claim 1, characterized in that, The spatiotemporal sequence encoding module simultaneously characterizes the dependencies between multiple variables and their temporal evolution, including: performing time series encoding on the obtained variable feature matrix; and encoding a sequence of length... The time windows are stacked into a three-dimensional tensor of "time, variables and features", introducing time series dependencies as input to the next time series prediction network; the change trajectory of each variable within the time window is extracted, and the three-dimensional tensor is treated as a multivariate time series for batch input.

4. The method for predicting and classifying early warning of train wheelset axle temperature based on recursive multivariate interactive networks according to claim 1, characterized in that, The recursive time series prediction module gradually maps the encoded multivariate feature sequence to the predicted bearing temperature value at future time, thereby realizing the modeling and prediction of the temperature change trend of train wheelset bearings. For each time step, the state is updated by combining the layer normalized gated loop unit, and the feature dimension of the sample at the current time step is standardized.

5. The method for predicting and classifying early warning of train wheelset axle temperature based on recursive multivariate interactive networks according to claim 1, characterized in that, A dual-branch fusion structure is introduced for prediction mapping to fully explore the nonlinear relationship between time dependence and candidate features. This structure performs layer normalization and nonlinear function activation on the final hidden state to extract stationary features, utilizing the candidate hidden state after... The activation captures its short-term dynamic changes, and the two parts are weighted and fused by linear transformation to obtain the final prediction output.

6. A train wheelset axle temperature prediction and early warning system based on a recursive multivariate interactive network, characterized in that, include: The acquisition module is used to acquire train speed data; The prediction module processes the acquired speed data using a pre-trained train wheelset axle temperature prediction grading model to obtain the predicted bearing temperature value at that operating speed. The train wheelset axle temperature prediction grading model includes: a multivariate dependency structure, constructing a multivariate correlation weight matrix and obtaining standardized interaction weights through normalization; a recursive multivariate interaction network, which comprehensively characterizes the complex multivariate dependencies within a time window while ensuring sufficient preservation of feature information; a spatiotemporal sequence encoding module, which simultaneously characterizes the dependencies between multivariates and their temporal evolution; a recursive time series prediction module, which progressively maps the encoded multivariate feature sequence to the predicted bearing temperature value at future times; and a dual-branch fusion structure is introduced for prediction mapping to fully exploit the nonlinear relationship between time dependencies and candidate features. The recursive multivariate interaction network, which comprehensively characterizes the complex multivariate dependencies within a time window while ensuring sufficient preservation of feature information, includes: for the input multivariate features, calculating the i-th variable and its associated variables. The single-head attention coefficients are normalized, and the sum of all attention coefficients related to the i-th variable is normalized to ensure that the sum of the attention coefficients is 1, resulting in standardized attention weights. Through weighted summation and self-loop compensation, each temporal variable can absorb the correlation information of its neighbors while retaining its own features during updates, achieving stable and flexible temporal feature aggregation. The vectors transmitted by the single-head information in each layer are aggregated, and a multi-head attention mechanism is implemented using multiple sets of attention weights in parallel computation. This allows variables to learn neighborhood dependencies in different feature subspaces, achieving multi-scale feature aggregation. The network is enhanced from local representation to global modeling, enabling variables to aggregate features of far-distance related variables. A normalized interaction weight matrix is ​​also included. For the characteristics of the current variable Perform feature aggregation and then use the linear transformation matrix Matrix multiplication is used to map the feature space, and then features from the previous layer are fused together. ,use Adjust dimensions to align the feature space; The early warning module is used for an automatic graded early warning mechanism based on predicted temperature rise. It determines whether the predicted bearing temperature exceeds the corresponding threshold according to the early warning threshold set for different operating speeds. If the threshold is exceeded, an early warning is issued.

7. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the train wheelset axle temperature prediction and graded early warning method based on a recursive multivariate interactive network as described in any one of claims 1-5.

8. A computer device, characterized in that, The system includes a memory and a processor, which communicate with each other. The memory stores program instructions that can be executed by the processor. The processor calls the program instructions to execute the train wheelset axle temperature prediction and graded early warning method based on a recursive multivariate interactive network as described in any one of claims 1-5.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions that implement the train wheelset axle temperature prediction and graded early warning method based on a recursive multivariate interactive network as described in any one of claims 1-5.

Images