Train hot wheelset monitoring method based on SRU-Hyena space-time double-flow fusion

Abstract

This invention relates to a train hot axle monitoring method based on SRU-Hyena spatiotemporal dual-stream fusion, belonging to the field of train inspection technology. To address the technical challenges of limited accuracy and difficulty in uniformly modeling the spatiotemporal correlation of temperature sequences in traditional train hot axle monitoring methods under complex operating conditions, this invention obtains an enhanced axle temperature temporal tensor based on background temperature points, multiple axle temperature points, and statistical features of these points. An enhanced feature representation is obtained based on a composite position encoding layer and fed into two independent branches for processing. The temporal dynamic branch obtains temporal features through a multi-layer Simple Recurrent Unit (SRU), while the spatial correlation branch obtains spatial features through a multi-layer Hyena encoder based on long convolution and gating structures. A spatiotemporal fusion features are then obtained using a spatiotemporal feature interaction layer. Scaling and translation parameters are derived from a static conditional tensor, and affine transformations are performed on the spatiotemporal fusion features based on these parameters. Finally, a classifier is used to obtain the train hot axle monitoring results.

Description

Technical Field

[0001] This invention belongs to the field of train inspection technology and relates to a method for monitoring hot axles in trains. Background Technology

[0002] Traditional methods for monitoring hot axles in trains have limited accuracy under complex operating conditions and struggle to uniformly model the spatiotemporal correlation of temperature sequences. With the continuous expansion of railway transportation and the constant improvement of traffic safety standards, existing monitoring schemes based on fixed thresholds or shallow neural networks generally lack effective collaborative analysis and deep feature extraction capabilities for multi-dimensional, long-cycle, heterogeneous time-series monitoring data composed of multiple axle temperature, wheel temperature, and ambient temperature probes. Summary of the Invention

[0003] This invention aims to solve the technical problems of limited accuracy of traditional train hot axle monitoring methods under complex operating conditions and the difficulty in uniformly modeling the spatiotemporal correlation of temperature sequences.

[0004] A train hot axle monitoring method based on SRU-Hyena spatiotemporal dual-stream fusion includes:

[0005] The processing steps for the shaft temperature time series tensor include:

[0006] S1. Construct the axial temperature time series tensor based on the background temperature point and multiple axial temperature points. ;

[0007] S2, Based on Axial Temperature Time Series Tensor The statistical characteristics of multiple shaft temperature points are enhanced. Primary embedding features are obtained through linear transformation matrix. ;

[0008] S3, Primary Embedding Features The data is input to the composite position coding layer, which encodes the time step sequence and the spatial probe position separately and then fuses them to obtain the composite position code. Then and The enhanced feature representation is obtained by adding elements together and aligning them. ;

[0009] S4, Enhanced Features The data is fed into two independent branches in parallel for processing: a temporal dynamic branch and the enhanced features. After dimensionality reshaping, the input is a SRU network containing multiple layers of simple recurrent units (SRUs) to obtain temporal features. Spatial correlation branch: the enhanced feature After dimensionality reshaping, the input is a Hyena network containing multiple layers of Hyena encoders based on long convolutions and gating structures, to obtain spatial features. ;

[0010] S5, will and The common input is fed into the spatiotemporal feature interaction layer, which then uses... For querying, with Simultaneously serving as both key and value, it is obtained based on an attention mechanism. Using the same method, For querying, with Simultaneously serving as both key and value, it is obtained based on an attention mechanism. ,based on and The splicing result yields fusion features ,Will The input gated fusion layer performs deep interaction and adaptive fine-grained fusion through gate weights to generate unified spatiotemporal fusion features. ;

[0011] Static condition tensor processing includes:

[0012] S6. Construct a static conditional tensor based on background temperature, multiple wheel temperatures, and operational data. ; the static condition tensor Encode as a condition vector ;

[0013] S7, For condition vectors Features are generated through a characteristic linear modulation mechanism and spatiotemporal fusion. Scaling parameters corresponding to the dimensions Translation parameters ;based on and spatiotemporal fusion features Affine transformation is performed to obtain conditional feature representations. ;

[0014] and conditional feature representation The steps to obtain the target classification.

[0015] Furthermore, in step S3, based on the axis temperature time series tensor... The statistical characteristics of multiple shaft temperature points are enhanced. The process includes:

[0016] Calculate the statistical characteristics of multiple shaft temperature points:

[0017] ,

[0018] in, The rows represent the expected value of the shaft temperature, the standard deviation of the shaft temperature, the maximum value of the shaft temperature, the minimum value of the shaft temperature, the difference between the expected shaft temperature and the background temperature, and the difference between the maximum shaft temperature and the background temperature, respectively. For vectors The One element, Starting from 0, Corresponding shaft temperature point, Corresponding to the background temperature points, there are a total of One temperature point; Axial temperature time series tensor The middle represents the time step. One side of the shaft probe , No. Shaft temperature vectors measured at each axial position;

[0019] And thus obtain .

[0020] Furthermore, in step S3, the composite position coding layer encodes the time step sequence and the spatial probe position respectively, and then fuses them to obtain the composite position code. The process includes:

[0021] For the time step sequence, sine and cosine coding is used to obtain the time position coding matrix. For the spatial probe position, a time-position encoding matrix is ​​obtained using sine and cosine coding. Broadcast temporal and spatial codes to the full dimension. get and Where B is the axis index The number of times, T is the time step index. The quantity, N, is the probe number index corresponding to a temperature measurement point. The quantity, D is the number of pairs. The dimension is set during linear transformation; the composite positional code is obtained by fusing the temporal and spatial codes. .

[0022] Furthermore, the specific processing steps for time-series dynamic branches include:

[0023] Will Reshaping the 3D matrix Form As the first-layer SRU in the SRU network, the input will be superscript. This represents the input of the first-level SRU;

[0024] Processing of a single-layer SRU: For the input Transform to obtain The data is processed using a dual gating mechanism that combines temperature adaptation and gradient enhancement.

[0025] ,

[0026] ,

[0027] in, Features obtained from a temperature-adaptive forgetting gate; Indicates the activation function; The SRU network can learn the forget gate weight matrix. The learnable forget gate bias vector for the SRU network; This is the temperature sensitivity coefficient; As a temperature sensitivity adjustment factor, Temperature rise factor; This is the matrix representing the maximum shaft temperature. This is the matrix representing the minimum shaft temperature.

[0028] ,

[0029] ,

[0030] in, Features obtained from the reset gate; For SRU networks, the learnable reset gate weight matrix, The SRU network can learn the reset gate bias vector; It is the gradient sensitivity coefficient; for Change in the matrix of maximum axis temperature at any given moment;

[0031] SRU output characteristics are obtained based on a dual gating mechanism of temperature adaptation and gradient enhancement. :

[0032] ,

[0033] ,

[0034] in, The hidden state at the current time step. This is the hidden state from the previous time step. , It is an all-zero matrix; This represents element-wise multiplication;

[0035] The features obtained after multi-layer SRU are shaped to obtain the dimension. Temporal characteristics .

[0036] Furthermore, the temperature rise factor is

[0037] .

[0038] Furthermore, the specific processing steps for spatially related branches include:

[0039] Will Reshaping the 3D matrix

[0040] The first layer of the Hyena network reshapes the input as follows:

[0041] ,

[0042] ,

[0043] in, , This represents the hidden dimension to which each probe feature of each axis is mapped within the model; For the reshaping operation, From a four-dimensional matrix Reshape the 3D matrix, then transform it along the axes and probe dimensions to obtain the 3D probe axis temperature feature matrix. The projection is a two-dimensional projection related to the time step. Probe shaft temperature characteristic matrix ; superscript This represents the input to the first layer of Hyena;

[0044] Single-layer Hyena processing: for input After projection and segmentation, the result is obtained ;

[0045] against Time dimension mean After projection and Function processing yields time kernel coefficients Then, the time kernel is calculated:

[0046] , ,

[0047] in, For the first Layer time kernel coefficient The One component; The time delay variable in the time kernel function. For the first Layer time kernel A learnable decay rate, For the first Layer time kernel One learnable oscillation frequency;

[0048] Calculate the spatial kernel matrix:

[0049] ,

[0050] in, Number the axes; It is a learnable controllable distance sensitivity coefficient;

[0051] After activation function get Iterative calculations are performed based on the spatiotemporal convolution operator:

[0052] ,

[0053] ,

[0054] in, To compute the spatiotemporal convolution operator; It is the inverse Fourier transform; This represents element-wise multiplication; yes 3D identity matrix Represents the tensor product;

[0055] according to Get Hyena's output: ;in, As an intermediate variable, It is a learnable weight matrix. It is a learnable bias vector;

[0056] The features obtained after multiple layers are shaped to obtain the dimension. Spatial features , It is the feature dimension of the Hyena space.

[0057] Furthermore, in step S5, with For querying, with Simultaneously serving as both key and value, it is obtained based on an attention mechanism. The process is as follows:

[0058] ,

[0059]

[0060] ,

[0061] ,

[0062] in, , , The projection weight matrix represents the spatial features. A query matrix with time features. The key matrix represents the spatial characteristics. The value matrix of spatial features, For layer normalization;

[0063] ,

[0064] in, , , For the variables used in the summation, The maximum number of elements; For learnable scaling parameters, For learnable translation parameters; To prevent division by zero of constants.

[0065] Furthermore, in step S5, deep interaction and adaptive fine-grained fusion are performed through gating weights to generate unified spatiotemporal fusion features. The process includes:

[0066] ,

[0067] ,

[0068] ,

[0069] in, It is a learnable weight matrix. It is a learnable bias vector; To adjust the parameter matrix.

[0070] Furthermore, the operational data described in S6 includes vehicle speed, month, and hours; a static conditional tensor is constructed based on background temperature and multiple wheel temperatures, along with the operational data. The process includes:

[0071] ,

[0072] in, Background temperature, For wheel temperature, For vehicle speed, For months, In hours.

[0073] Furthermore, in step S6, the static condition tensor... Encode as a condition vector The process includes:

[0074] ,

[0075] ,

[0076] ,

[0077] in, This is the initial input for encoding. It is a learnable weight matrix. It is a learnable bias vector; It is an activation function. This represents the total number of encoding layers.

[0078] Beneficial effects:

[0079] 1. Efficient Modeling of Long-Term Time-Series Features: This invention employs the SRU time-series modeling unit to achieve efficient parallel processing of long-series shaft temperature data while maintaining linear computational complexity. This design effectively overcomes the inherent defects of traditional recurrent neural networks, such as slow training, difficulty in capturing long-term dependencies, and susceptibility to gradient vanishing, significantly improving the model's ability to extract long-term dynamic trends and periodic patterns in the evolution of bearing temperature.

[0080] 2. Multi-scale global spatial correlation modeling: This invention introduces the Hyena long convolutional structure to construct a multi-level, large receptive field spatial correlation modeling capability. This mechanism breaks through the structural constraints of traditional convolutional neural networks that are limited to local neighborhoods, and can accurately characterize the spatial correlation between different probes (including internal and external probes) based on the coupling of physical heat conduction and mechanical vibration, realizing global perception and cross-scale modeling of the axle temperature field distribution characteristics of the entire vehicle.

[0081] 3. Deep Fusion and Synergy of Spatiotemporal Features: This invention constructs a deep interaction path between the SRU temporal dynamic feature stream and the Hyena spatial correlation feature stream through the collaborative design of a temporal and spatial feature extraction network architecture and an adaptive gating fusion mechanism. This fusion mechanism can dynamically adjust the information interaction weights according to the spatiotemporal characteristics of the input data, thereby achieving effective complementarity and synergistic enhancement of spatiotemporal features, avoiding information loss and modeling bias caused by feature separation or simple splicing in traditional methods.

[0082] 4. Strong generalization capability adapting to complex operating conditions: This invention addresses the multi-condition characteristics of train axle temperature data under the influence of factors such as train speed, ambient temperature, and seasonal conditions. It optimizes the positional encoding of physical information embedding and the hierarchical feature fusion strategy. This enables the model to adaptively learn the influence of different operating conditions on temperature distribution, significantly improving the robustness and generalization performance of accurate classification and diagnosis of hot axle states (normal, slightly hot, strongly hot, and intensely hot) in complex and variable scenarios. Attached Figure Description

[0083] Figure 1 Here is a flowchart of the train hot axle monitoring method;

[0084] Figure 2 This is a diagram of the SRU-Hyena spatiotemporal dual-stream fusion network architecture. Detailed Implementation

[0085] It should be noted that, where there is no conflict, the various embodiments disclosed in this application can be combined with each other.

[0086] To address the problems existing in the background technology, this invention proposes a train hot axle monitoring method based on SRU-Hyena spatiotemporal dual-stream fusion. The SRU stream focuses on efficiently learning the long-range dynamic dependencies and temporal evolution patterns in the temperature sequences of each measuring point; the Hyena stream, based on its long convolution and gating structure, specifically extracts multi-scale correlation features from the spatial topology and thermal conductivity physical properties of the sensor network. By constructing a temporal and spatial dual-stream collaborative architecture and designing an adaptive gating fusion mechanism, while combining the unique periodicity, trend, and spatial constraints of train axle temperature data, targeted optimizations are made to the model input representation and feature interaction. This achieves deep synergy and complementary enhancement of the two cutting-edge sequence modeling methods, SRU and Hyena, forming the core innovative system of this invention. This scheme significantly improves the accurate identification, fine classification, and system robustness of hot axle states (normal, slightly hot, strongly hot, and intensely hot) under complex and variable operating environments.

[0087] Specific implementation method one: Combining Figure 1 and Figure 2 This implementation method is described below.

[0088] This embodiment presents a train hot axle monitoring method based on SRU-Hyena spatiotemporal dual-stream fusion, which achieves train hot axle monitoring through an SRU-Hyena spatiotemporal dual-stream fusion neural network. This embodiment provides a structure including an SRU layer and a Hyena encoder layer. However, it should be noted that, based on the network structure of this embodiment, those skilled in the art can adjust the number of SRU layers and the number of Hyena encoder layers, and can also adjust the network structure according to actual needs.

[0089] Referring to Table 1, the network structure of this embodiment is described. The SRU-Hyena spatiotemporal dual-stream fusion neural network described in this embodiment includes: a feature enhancement embedding layer, a position encoding layer, a feature extraction network, a spatiotemporal feature interaction layer, a gated fusion layer, a pooling layer, and a classification layer.

[0090] Table 1

[0091]

[0092] For multi-source heterogeneous imaging data and state condition data, this implementation method adopts an innovative design of split-path asynchronous input. The specific input path is described as follows:

[0093] S1, Main path input of the shaft temperature time series tensor:

[0094] Axial temperature time series tensor The core multidimensional temporal signal characterizing the dynamic physical processes of the target object serves as the backbone input source for the neural network model. This tensor is received at the front-end input layer of the model and undergoes feature evolution through the following hierarchical progressive processing path.

[0095] S2, Primary Feature Embedding:

[0096] The shaft temperature time series tensor First, the data is input into the model's feature enhancement embedding layer (Layer 1 in Table 1). This layer adds statistical features (feature enhancement through adding statistical features) to the axial temperature time-series tensor and background temperature time-series data, thus enhancing the model. Then, through a learnable linear transformation matrix and bias vector This maps the original sensor point-dimensional space to a higher-dimensional latent feature space, yielding the primary embedding features. .

[0097] S3, Composite Location Information Injection:

[0098] Primary Embedded Features The data is input to the composite location coding layer, which generates a composite location code that includes the time step sequence and the spatial probe position. Then and Element-wise alignment and addition are performed to form an enhanced feature representation containing complete structural prior information. .

[0099] S4. Spatiotemporal Feature Separation and Extraction:

[0100] Enhanced features The features are fed in parallel into two independently optimized feature extraction branches for processing.

[0101] Temporal dynamic branches: the enhanced features After dimensional reshaping, the input is fed into the SRU network, which consists of three layers of simple recurrent units (SRUs), namely the simple recurrent units (SRUs) corresponding to layers 3-5 in Table 1, which are passed through in sequence to capture and abstract short-term fluctuations, medium-term trends and long-term dependency patterns in the time series layer by layer.

[0102] Spatial correlation branches: the enhanced features After being reshaped in another dimension, the data is input into the Hyena network, which consists of three layers of Hyena encoders based on long convolution and gating structures. These layers are sequentially input into the Hyena encoder layers corresponding to layers 6-8 in Table 1. This branch extracts the global spatial correlation features of the temperature field at the same time from different sensor measurement points across the entire train through multi-scale, large receptive field convolutional kernel operations.

[0103] S5, Spatiotemporal Feature Interactive Fusion:

[0104] Advanced timing features output by timing dynamic branches High-level spatial features output by spatially related branches The features are jointly input into the spatiotemporal feature interaction layer to generate fused features. (Layer 9 in Table 1), then... The input gated fusion layer (layer 10 in Table 1) performs deep interaction and adaptive fine-grained fusion through learnable gate weights to generate unified spatiotemporal fusion features. .

[0105] S6. Asynchronous conditionalization of static conditional tensors:

[0106] Static condition tensor To characterize auxiliary conditions related to environmental background and operating conditions, an asynchronous delayed input strategy is adopted, with conditional injection performed during the feature fusion stage. The input and processing path is as follows:

[0107] Delayed input point: the static condition tensor Instead of being input at the model front end, it is in the spatiotemporal fusion features. After generation, it is formally introduced into the processing flow at the gating fusion layer.

[0108] Static feature encoding: Within the gated fusion layer, a separate static feature encoding submodule is set up. This submodule includes one or more fully connected layers. The static conditional tensor... The result is encoded into a low-dimensional dense conditional vector through nonlinear transformation and compression via a static feature encoding submodule. .

[0109] S7, Feature Modulation Fusion:

[0110] Condition vector The fusion process involves a feature-linear modulation mechanism. Specifically, it utilizes learnable parameter matrices to generate spatiotemporal fusion features. Scaling parameters corresponding to the dimensions Translation parameters :

[0111] ,

[0112] S8. Feature Transformation and Prediction:

[0113] Affine transformation of the spatiotemporal fusion features is performed using the following formula:

[0114] ,

[0115] in, This represents element-wise multiplication. Therefore, static conditional information, in the form of a global conditional field, adaptively modulates and corrects the spatiotemporal fusion feature representation, forming the final conditional feature representation. This is used for subsequent probabilistic decision-making and classification, i.e., fed into the classifier for classification.

[0116] Through the above-described branched design, the shaft temperature time-series data and the static conditional data follow different input paths and processing sequences: the core dynamic signal is input from the model front end along the main path and forms the basic representation after deep spatiotemporal feature extraction; while the static auxiliary signal is injected in the form of modulation parameters in the high-level feature abstraction stage through an asynchronous conditional path. This method achieves optimal adaptation to the characteristics of heterogeneous data, avoids information interference caused by early feature mixing, and enhances the model's ability to discriminate complex situations through later fine modulation.

[0117] Example

[0118] Step 1: Processing flow of shaft temperature timing tensor:

[0119] Each axis has two probes on one side (left and right): an inner probe and an outer probe. Each probe measures 32 axis temperature sampling points along the axis, for a total of 64 sampling points on one side of the axis.

[0120] Axis Index: , This represents the total number of axles.

[0121] Time step index: , This represents the number of time steps.

[0122] Lateral Identifier Index: (0 represents the left side, 1 represents the right side).

[0123] Probe type index: (0 indicates inward exploration, 1 indicates outward exploration).

[0124] Temperature point index: ( The time is the probe's shaft temperature point. (This is the background temperature point of the probe).

[0125] Shaft temperature uses variables Indicates the time step One side of the shaft probe , No. Shaft temperature vector measured at each axial position It is actually a vector composed of a 33-bit axial temperature sequence and background temperature. (Corresponding temperature data).

[0126] make And then according to Constructing the axial temperature spatiotemporal distribution tensor ;

[0127] Based on the shaft temperature sequence and background temperature, statistical features including the expected shaft temperature, standard deviation of shaft temperature, maximum shaft temperature, and minimum shaft temperature are calculated and extracted. Furthermore, the difference between the expected shaft temperature and the background temperature (to characterize the overall temperature rise level), and the difference between the maximum shaft temperature and the background temperature (to characterize the maximum instantaneous temperature rise) are constructed, thus forming an enhanced feature set that comprehensively characterizes the bearing's thermal state and stability. These features are combined into a statistical feature vector.

[0128] ,

[0129] in, The rows represent the expected value of the shaft temperature, the standard deviation of the shaft temperature, the maximum value of the shaft temperature, the minimum value of the shaft temperature, the difference between the expected shaft temperature and the background temperature, and the difference between the maximum shaft temperature and the background temperature, respectively.

[0130] For vectors The One element, Starting from 0, The time is the probe's shaft temperature point. The time is the background temperature point of the probe.

[0131] This leads to an enhanced representation of the vector. And then according to Build .

[0132] (i) Perform independent linear projection on each spatiotemporal location using linear transformation parameters:

[0133] ,

[0134] in, , These are the parameters for linear transformation; To hide the dimension, in this embodiment .

[0135] right Activate: And then according to get .

[0136] (ii) Composite positional coding:

[0137] (1) Time location coding:

[0138] This embodiment uses sine and cosine encoding:

[0139] ,

[0140] ,

[0141] in, yes Time index, Used to represent The dimension of the index;

[0142] This leads to the matrix of time position encoding. , is represented as:

[0143] ,

[0144] (2) Spatial location coding

[0145] This embodiment uses sine and cosine encoding:

[0146] ,

[0147] ,

[0148] in, yes The probe number index is as follows: in this embodiment, 0 represents the left inner probe, 1 represents the left outer probe, 2 represents the right inner probe, and 3 represents the right outer probe. Used to represent The dimension of the index.

[0149] This leads to the spatial location encoding matrix. , is represented as:

[0150] ,

[0151] in, For each temperature measurement point, a probe number index is assigned. Quantity ( ).

[0152] Broadcast the temporal and spatial codes to the full dimension:

[0153] ,

[0154] ,

[0155] broadcast The specific operation is as follows:

[0156] ,

[0157] ,

[0158] in, As axis index, For time step index, Index the probe number. express The Middle The complete encoded vector for each time step. express The Middle The complete encoded vector of each probe location;

[0159] , The broadcast operation expands the dimension of the complete encoded vector to the same level as the feature tensor. Matching dimensions.

[0160] Composite positional coding is a weighted sum of temporal and spatial coding:

[0161] , ,

[0162] in, For weight fusion.

[0163] Add the primary embedding features and the composite positional encoding element-wise aligned:

[0164] , ,

[0165] Element-wise expansion is represented as:

[0166] ,

[0167] (iii) Enhanced features This will be used as input to the feature extraction network:

[0168] SRU stream input: will The reshaped input is fed into the SRU network to capture temporal dynamics.

[0169] Hyena stream input: will The reshaped data is then fed into the Hyena network to capture spatiotemporal dependencies.

[0170] (A) The input reshaping process of the first-layer SRU in the SRU network is as follows:

[0171] ,

[0172] , ,

[0173] Among them, superscript This represents the input to the first layer of the SRU; This indicates a reshaping operation. This indicates a flattening operation. Its function is to From a four-dimensional matrix Reshaping the 3D matrix .

[0174] During forward propagation of a single-layer SRU, the input , , , Output , For the number of floors, .

[0175] The processing procedure for a single-layer SRU is as follows:

[0176] First, the input Perform the transformation:

[0177]

[0178] in, This is the learnable candidate state weight matrix for the SRU network.

[0179] The processing of other SRU layers is the same as that of the first SRU layer, except that the input is the output of the previous layer.

[0180] The three-layer SRU architecture innovatively adopts a layered enhancement strategy, introducing dual gating mechanisms of temperature adaptation and gradient enhancement:

[0181] Temperature-adaptive forget gate:

[0182] ,

[0183] ,

[0184] in, Features obtained from a temperature-adaptive forgetting gate; This represents the activation function. ; The SRU network can learn the forget gate weight matrix. The learnable forget gate bias vector for the SRU network; ( () represents the temperature sensitivity coefficient; As a temperature sensitivity adjustment factor, Temperature rise factor; This is the matrix representing the maximum shaft temperature. This is the matrix representing the minimum shaft temperature.

[0185] ,

[0186] ,

[0187] ,

[0188] in, As axis index, For time step index, For lateral indexing, This is an index for the probe type. for 3D identity matrix.

[0189] Temperature sensitivity adjustment factor The numerator approximately represents the temperature difference between the probes. If all axle temperatures are generally high, the temperature difference between the probes will be small; if only one probe has an abnormally high temperature, while the other probes have normal or very low temperatures, the temperature difference between the probes will be large. The denominator approximately represents how much the hottest axle is hotter than the background temperature, with a default minimum temperature difference of 30. In hot weather or during long-distance braking, the axle temperature across the entire vehicle will generally rise, increasing the temperature difference in the denominator, but the temperature difference between the probes in the numerator will not change drastically. The overall increase will not be significant. However, if a genuine thermal shaft failure exists (usually only one shaft is faulty), the denominator temperature rise difference will increase, and the shaft temperature difference between the probes in the numerator will change drastically. The overall size will increase. Increasing the value increases the probability that the characteristic data is identified as hot axle fault data. A general increase in axle temperature across the entire vehicle does not increase the probability of identifying characteristic data as hot axle fault data; this only increases the probability when a genuine hot axle fault exists.

[0190] Gradient-enhanced reset gate:

[0191] ,

[0192] ,

[0193] in, Features obtained from the reset gate; For SRU networks, the learnable reset gate weight matrix, The SRU network can learn the reset gate bias vector; ( ) is the gradient sensitivity coefficient (manually tuned and set, generally set to...). The value range is , Large values ​​may lead to gradient explosion, while small values ​​may lead to gradient vanishing. Generally, values ​​between 1 and 2 are suitable. (Value is 1.5). , Consistent with the above; The change in the matrix of the maximum shaft temperature at time t.

[0194] SRU output characteristics are obtained based on a dual gating mechanism of temperature adaptation and gradient enhancement. :

[0195] ,

[0196] ,

[0197] in, The hidden state at the current time step. This is the hidden state from the previous time step. , It is an all-zero matrix; , This indicates element-wise multiplication.

[0198] The formula for stacking three-layer SRUs is as follows:

[0199] ,

[0200] ,

[0201] ,

[0202] ,

[0203] in, It is the final output after plastic surgery. It is the time feature dimension of SRU.

[0204] This invention innovatively introduces a dual gating mechanism of temperature adaptation and gradient enhancement into each layer of the three-layer SRU architecture: a temperature adaptation mechanism is embedded in the forget gate, enabling the network to intelligently adjust the strength of historical memory based on temperature anomalies, achieving "continuous high-temperature memory"; a gradient enhancement mechanism is embedded in the reset gate, enabling the system to dynamically switch strategies based on the rate of temperature change, achieving "temperature mutation-priority response". This dual-driven adaptive mechanism of temperature and gradient endows the model with physical perception capabilities, significantly improving the sensitivity and anti-interference ability of early thermal axis warning, and realizing a monitoring improvement from passive response to active perception.

[0205] (B) The input reshaping process of the first layer of the Hyena network is as follows:

[0206] ,

[0207] ,

[0208] in, , This represents the hidden dimension to which each probe feature of each axis is mapped within the model. For the reshaping operation, From a four-dimensional matrix Reshaping the 3D matrix ; It involves transforming along the axis and probe dimensions to obtain the three-dimensional probe axis temperature feature matrix. The projection is a two-dimensional projection related to the time step. The probe shaft temperature characteristic matrix.

[0209] The input to the Hyena layer is , For the number of floors, ;

[0210] Projection and block processing are performed on the input:

[0211] ,

[0212] ,

[0213] in, is a learnable parameter, representing the bias of the weights; This indicates block processing; , , As an intermediate variable;

[0214] against Perform the following gating process:

[0215] , ,

[0216] Calculate the mean of the input. And calculate the time kernel coefficients. :

[0217] , ,

[0218] , ,

[0219] in, This indicates the number of fundamental functions used to construct the temporal convolution kernel, in this embodiment. The value is 16; ; is a learnable parameter, representing the weights and biases.

[0220] Calculation time kernel:

[0221] , ,

[0222] in, For the first Layer time kernel coefficient The One component; This is the time delay variable in the time kernel function (the time interval between the current calculation time and the historical data time). , It is the maximum value of the time interval between the current and historical periods. The larger it is, the smaller the impact. For the first Layer time kernel A learnable decay rate, For the first Layer time kernel A learnable oscillation frequency.

[0223] Calculate the spatial kernel matrix:

[0224] , ,

[0225] in, Number the axis. The larger it is, the smaller the impact. It is a learnable controllable distance sensitivity coefficient.

[0226] Computing the spatiotemporal convolution operator:

[0227] ,

[0228] in, for The input matrix; It is the inverse Fourier transform; This represents element-wise multiplication; yes 3D identity matrix This represents the tensor product (Kronecker product).

[0229] Iterative computation based on spatiotemporal convolution operators:

[0230] ,

[0231] Ultimately based on Get Hyena's output: ;in, As an intermediate variable, It is a learnable weight matrix. It is a learnable bias vector.

[0232] This invention proposes an improved Hyena neural network for shaft temperature monitoring, with its core innovation lying in the design of a spatiotemporally separated convolutional kernel architecture. This architecture decouples the unified feature mixing process in the original Hyena operator into two parallel and collaborative independent computational branches: a temperature-adaptive temporal kernel and a temperature-driven spatial kernel. The temporal kernel is responsible for extracting multi-scale temporal features from long-sequence temperature signals. Its convolutional parameters are dynamically adjusted by a multilayer perceptron based on real-time temperature, thus adapting to the temporal dynamics under different thermal states. The spatial kernel models the topological associations and thermal coupling effects between multiple probes on the same bearing. Its interaction weights are adaptively generated by the real-time temperature difference between probes to dynamically characterize the heat diffusion process and local anomalous hotspots. The outputs of the two kernel functions are fused under a gating mechanism to form a feature representation that combines spatiotemporal collaborative sensing with physical interpretability, effectively improving the monitoring sensitivity and location accuracy of early bearing faults.

[0233] Three-layer Hyena stacking formula:

[0234] ,

[0235] ,

[0236] ,

[0237] ,

[0238] in, It is the final output after plastic surgery. It is the feature dimension of the Hyena space.

[0239] (iv) Two-way cross-attention interaction:

[0240] Based on time series characteristics For querying spatial features Simultaneously serving as both key and value ( Based on the attention mechanism, the temporal information at each time step is used to "query" and aggregate relevant spatial information:

[0241] ,

[0242] ,

[0243] ,

[0244] ,

[0245] Based on spatial characteristics For querying time series features Simultaneously serving as both key and value ( Based on an attention mechanism, the spatial information from each probe is used to "interview" and aggregate relevant temporal information:

[0246] ,

[0247] ,

[0248] ,

[0249] ,

[0250] in, , , The projection weight matrix represents the spatial features. , , The projection weight matrix represents the time features; and For the dimensions used to calculate queries, keys, and values, Output dimensions for spatiotemporal features. The query matrix for spatial features. The key matrix represents the spatial characteristics. The value matrix of spatial features, A query matrix with time features. The key matrix is ​​a time-dependent feature. It is a value matrix representing the time characteristics.

[0251] ,

[0252] in, , , For the variables used in the summation, The maximum number of elements; For learnable scaling parameters, For learnable translation parameters; It is a small constant, usually (To prevent division by zero).

[0253] ,

[0254] in, , For the variables used in the summation, This represents the maximum number of elements.

[0255] The spatiotemporal interaction layer output is obtained based on the bidirectional cross-attention concatenation result:

[0256]

[0257] in, yes and splicing, ; It is a dimension-reduced projection weight matrix (which will...) Dimensional reduction ), .

[0258] By fusing through a gated fusion layer, a unified spatiotemporal fusion feature is generated. :

[0259] ,

[0260] ,

[0261] ,

[0262] in, It is a learnable weight matrix. It is a learnable bias vector; To adjust the parameter matrix.

[0263] Step 2, Static Condition Tensor Processing Flow:

[0264] Axis Index: , Total number of axles; Time step index: , For time steps; 0 represents the left side, 1 represents the right side, and 2 represents both sides.

[0265] The static condition tensor is described as follows:

[0266] Background temperature: 1D (actual background temperature value and probe type index) It's irrelevant; the background temperature value is only differentiated between the left and right sides, corresponding to the inner and outer probes. The background temperature value for both the inner and outer probes is actually the same.

[0267] ;

[0268] Wheel temperature: 32 dimensions Wheel temperature is the temperature of the wheel tread surface. There are 32 points on one side, and all 32 temperature points are wheel temperature points.

[0269] Vehicle speed: 1 dimension (one for each vehicle). ;

[0270] Month: 2D (one per train, using sine and cosine encoding). ;

[0271] Hours: 2D (one per train, using sine and cosine encoding). ;

[0272] , ,

[0273] The month and hour are obtained by recording the timestamp of the train passing the detection equipment. The system determines whether the current operation is seasonal based on the month and whether it is daytime or nighttime based on the hour.

[0274] static condition tensor Encode as a low-dimensional dense condition vector :

[0275] ,

[0276] ,

[0277] ,

[0278] in, It is a learnable weight matrix. It is a learnable bias vector; It is an activation function. . This represents the total number of encoding layers.

[0279] based on Features are generated and fused spatiotemporally through learnable parameter matrices. Scaling parameters corresponding to the dimensions Translation parameters :

[0280] ,

[0281] in, and These are learnable parameters, learned through backpropagation.

[0282] Affine transformation is applied to the spatiotemporal fusion features to generate the final conditional feature representation. :

[0283] ,

[0284] Step 3: Based on conditional feature representation Target classification obtained:

[0285] , ,

[0286] in, and These are learnable parameters, learned through backpropagation. Output detection results: Level 0 (normal), Level 1 (minor heat warning), Level 2 (severe heat alarm), Level 3 (heating fault).

[0287] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for monitoring hot axles of trains based on SRU-Hyena spatiotemporal dual-stream fusion, characterized in that, include: The processing steps for the shaft temperature time series tensor include: S1. Construct the axial temperature time series tensor based on the background temperature point and multiple axial temperature points. ; S2, Based on Axial Temperature Time Series Tensor The statistical characteristics of multiple shaft temperature points are enhanced. Primary embedding features are obtained through linear transformation matrix. ; S3, Primary Embedding Features The data is input to the composite position coding layer, which encodes the time step sequence and the spatial probe position separately and then fuses them to obtain the composite position code. Then and The enhanced feature representation is obtained by adding elements together and aligning them. ; S4, Enhanced Features The data is fed into two independent branches in parallel for processing: a temporal dynamic branch and the enhanced features. After dimensionality reshaping, the input is a SRU network containing multiple layers of simple recurrent units (SRUs) to obtain temporal features. Spatial correlation branch: the enhanced feature After dimensionality reshaping, the input is a Hyena network containing multiple layers of Hyena encoders based on long convolutions and gating structures, to obtain spatial features. ; S5, will and The common input is fed into the spatiotemporal feature interaction layer, which then uses... For querying, with Simultaneously serving as both key and value, it is obtained based on an attention mechanism. Using the same method, For querying, with Simultaneously serving as both key and value, it is obtained based on an attention mechanism. ,based on and The splicing result yields fusion features ,Will The input gated fusion layer performs deep interaction and adaptive fine-grained fusion through gate weights to generate unified spatiotemporal fusion features. ; Static condition tensor processing includes: S6. Construct a static conditional tensor based on background temperature, multiple wheel temperatures, and operational data. ; the static condition tensor Encode as a condition vector ; S7, For condition vectors Features are generated through a characteristic linear modulation mechanism and spatiotemporal fusion. Scaling parameters corresponding to the dimensions Translation parameters ;based on and spatiotemporal fusion features Affine transformation is performed to obtain conditional feature representations. ; and conditional feature representation The steps to obtain the target classification.

2. The train hot axle monitoring method based on SRU-Hyena spatiotemporal dual-stream fusion according to claim 1, characterized in that, In step S3, based on the shaft temperature time series tensor The statistical characteristics of multiple shaft temperature points are enhanced. The process includes: Calculate the statistical characteristics of multiple shaft temperature points: ， in, The rows represent the expected value of the shaft temperature, the standard deviation of the shaft temperature, the maximum value of the shaft temperature, the minimum value of the shaft temperature, the difference between the expected shaft temperature and the background temperature, and the difference between the maximum shaft temperature and the background temperature, respectively. For vectors The One element, Starting from 0, Corresponding shaft temperature point, Corresponding to the background temperature points, there are a total of One temperature point; Axial temperature time series tensor The middle represents the time step. One side of the shaft probe , No. Shaft temperature vectors measured at each axial position; And thus obtain .

3. The train hot axle monitoring method based on SRU-Hyena spatiotemporal dual-stream fusion according to claim 1, characterized in that, In step S3, the composite position coding layer encodes the time step sequence and the spatial probe position respectively, and then fuses them to obtain the composite position code. The process includes: For the time step sequence, sine and cosine coding is used to obtain the time position coding matrix. For the spatial probe position, a time-position encoding matrix is ​​obtained using sine and cosine coding. Broadcast temporal and spatial codes to the full dimension. get and Where B is the axis index The number of times, T is the time step index. The quantity, N, is the probe number index corresponding to a temperature measurement point. The quantity, D is the number of pairs. The dimension is set during linear transformation; the composite positional code is obtained by fusing the temporal and spatial codes. .

4. The train hot axle monitoring method based on SRU-Hyena spatiotemporal dual-stream fusion according to claim 3, characterized in that, The specific processing steps for time-series dynamic branches include: Will Reshaping the 3D matrix Form As the first-layer SRU in the SRU network, the input will be superscript. This represents the input of the first-level SRU; Processing of a single-layer SRU: For the input Transform to obtain The data is processed using a dual gating mechanism that combines temperature adaptation and gradient enhancement. ， ， in, Features obtained from a temperature-adaptive forgetting gate; Indicates the activation function; The SRU network can learn the forget gate weight matrix. The learnable forget gate bias vector for the SRU network; This is the temperature sensitivity coefficient; As a temperature sensitivity adjustment factor, Temperature rise factor; This is the matrix representing the maximum shaft temperature. This is the matrix representing the minimum shaft temperature. ， ， in, Features obtained from the reset gate; For SRU networks, the learnable reset gate weight matrix, The SRU network can learn the reset gate bias vector; It is the gradient sensitivity coefficient; The change in the matrix of the maximum shaft temperature at time t; SRU output characteristics are obtained based on a dual gating mechanism of temperature adaptation and gradient enhancement. : ， ， in, The hidden state at the current time step. This is the hidden state from the previous time step. , It is an all-zero matrix; This represents element-wise multiplication; The features obtained after multi-layer SRU are shaped to obtain the dimension. Temporal characteristics .

5. The train hot axle monitoring method based on SRU-Hyena spatiotemporal dual-stream fusion according to claim 4, characterized in that, The temperature rise factor is: 。 6. The train hot axle monitoring method based on SRU-Hyena spatiotemporal dual-stream fusion according to claim 5, characterized in that, The specific processing steps for spatially related branches include: Will Reshaping the 3D matrix The first layer of the Hyena network reshapes the input as follows: ， ， in, , This represents the hidden dimension to which each probe feature of each axis is mapped within the model; For the reshaping operation, From a four-dimensional matrix Reshape the 3D matrix, then transform it along the axes and probe dimensions to obtain the 3D probe axis temperature feature matrix. The projection is a two-dimensional projection related to the time step. Probe shaft temperature characteristic matrix ; superscript This represents the input to the first layer of Hyena; Single-layer Hyena processing: for input After projection and segmentation, the result is obtained ; against Time dimension mean After projection and Function processing yields time kernel coefficients Then, the time kernel is calculated: ， ， in, For the first Layer time kernel coefficient The One component; This refers to the time delay variable in the time kernel function; For the first Layer time kernel A learnable decay rate, For the first Layer time kernel One learnable oscillation frequency; Calculate the spatial kernel matrix: ， in, Number the axes; It is a learnable controllable distance sensitivity coefficient; After activation function get Iterative calculations are performed based on the spatiotemporal convolution operator: ， ， in, To compute the spatiotemporal convolution operator; It is the inverse Fourier transform; This represents element-wise multiplication; yes 3D identity matrix Represents the tensor product; according to Get Hyena's output: ;in, As an intermediate variable, It is a learnable weight matrix. It is a learnable bias vector; The features obtained after multiple layers are shaped to obtain the dimension. Spatial features , It is the feature dimension of the Hyena space.

7. The train hot axle monitoring method based on SRU-Hyena spatiotemporal dual-stream fusion according to claim 6, characterized in that, In step S5, with For querying, with Simultaneously serving as both key and value, it is obtained based on an attention mechanism. The process is as follows: ， ， ， in, , , The projection weight matrix represents the spatial features. A query matrix with time features. The key matrix represents the spatial characteristics. The value matrix of spatial features, For layer normalization; ， in, , , For the variables used in the summation, The maximum number of elements; For learnable scaling parameters, For learnable translation parameters; To prevent division by zero of constants.

8. The train hot axle monitoring method based on SRU-Hyena spatiotemporal dual-stream fusion according to claim 6, characterized in that, In step S5, deep interaction and adaptive fine-grained fusion are performed through gating weights to generate unified spatiotemporal fusion features. The process includes: ， ， ， in, It is a learnable weight matrix. It is a learnable bias vector; To adjust the parameter matrix.

9. A train hot axle monitoring method based on SRU-Hyena spatiotemporal dual-stream fusion according to any one of claims 1 to 8, characterized in that, The operational data described in S6 includes vehicle speed, month, and hours; a static conditional tensor is constructed based on background temperature and multiple wheel temperatures, along with the operational data. The process includes: ， in, Background temperature, For wheel temperature, For vehicle speed, For months, In hours.

10. A train hot axle monitoring method based on SRU-Hyena spatiotemporal dual-stream fusion according to claim 9, characterized in that, In step S6, the static condition tensor is... Encode as a condition vector The process includes: ， ， ， in, This is the initial input for encoding. It is a learnable weight matrix. It is a learnable bias vector; It is an activation function. This represents the total number of encoding layers.

Images