A hybrid deep neural network-based real-time prediction system and method for temperature field of a traction system

Abstract

The application discloses a kind of real-time prediction system and method of traction system temperature field based on hybrid deep neural network, and prediction system includes: data acquisition module, frequency domain conversion module, spatial branch module, time branch module, feature fusion module, fully connected layer module, prediction output module;Data acquisition module is used to record the original temperature data of different components of traction system at the same time by thermocouple, synchronously acquires the operating condition information of the traction system, and exports the original temperature data sequence and operating condition information data set formed;Frequency domain conversion module is used to perform Fourier transform on the original temperature data sequence exported by data acquisition module, extract the frequency domain features representing the nature law of temperature change and potential periodic pattern, and generate graph structure data with spatial structure association;The application solves the technical problems of existing traction system temperature field prediction lag, insufficient accuracy and poor adaptability.

Description

Technical Field

[0001] This invention relates to the field of traction system temperature field prediction technology, and in particular to a real-time prediction system and method for traction system temperature field based on a hybrid deep neural network. Background Technology

[0002] As the core power unit of mobile equipment such as trains and rail vehicles, the traction system's operating status directly determines the equipment's safety, stability, and service life. During high-load, long-term operation, key components of the traction system, such as the traction transformer, traction converter, and traction motor, generate a large amount of heat. Abnormal changes in the temperature field can easily lead to problems such as insulation aging, component wear, and performance degradation. In severe cases, this can cause system failure and shutdown, resulting in significant operational losses.

[0003] Current traction system temperature monitoring primarily relies on real-time data acquisition, using hardware such as thermocouples and temperature sensors to obtain current temperature data. This approach only enables "post-event monitoring" and cannot predict future temperature trends, failing to meet the needs of preventative maintenance and fault early warning in traction systems. To achieve temperature prediction, existing technologies have developed prediction methods based on traditional machine learning (such as support vector regression and linear regression) and single neural networks (such as recurrent neural networks, RNNs). However, these methods have significant limitations: traditional machine learning methods struggle to capture the complex temporal dependencies and spatial correlations of temperature data, resulting in low prediction accuracy; single neural network models can only focus on feature extraction in the time or spatial dimensions, failing to effectively integrate the spatiotemporal coupling characteristics of the temperature field. This leads to large medium- and long-term prediction errors, weak generalization ability, and difficulty in adapting to the actual scenarios of traction systems with multiple components, multiple measuring points, and dynamically changing operating conditions.

[0004] Therefore, there is an urgent need for a temperature field prediction technology that can simultaneously mine the spatiotemporal characteristics of the temperature field of the traction system, with high prediction accuracy, strong real-time performance, and excellent generalization ability, in order to solve the problems of prediction lag, insufficient accuracy, and poor adaptability in the existing technology, and provide reliable technical support for real-time monitoring, fault early warning, and preventive maintenance of the traction system. Summary of the Invention

[0005] The purpose of this invention is to provide a real-time prediction system and method for the temperature field of a traction system based on a hybrid deep neural network, so as to solve the technical problems of lag, insufficient accuracy and poor adaptability in the prediction of the temperature field of the traction system in the prior art.

[0006] To achieve the above objectives, the present invention provides a real-time temperature field prediction system for a traction system based on a hybrid deep neural network, comprising:

[0007] The data acquisition module is used to synchronously record the raw temperature data of different components of the traction system at the same time through thermocouples, synchronously collect the operating condition information of the traction system, and output the resulting raw temperature data sequence and operating condition information dataset.

[0008] The frequency domain transformation module is used to perform Fourier transform on the raw temperature data sequence output by the data acquisition module, extract frequency domain features that characterize the essential laws and potential periodic patterns of temperature changes, and generate graph structure data with spatial structure correlation.

[0009] The spatial branching module is used to process the graph structure data with spatial structure associations generated by the frequency domain transformation module, and to use a graph attention network to model the spatial dependencies between graph nodes, outputting a spatial feature vector that characterizes the spatial distribution of the temperature field of the traction system.

[0010] The time branch module is used to learn complex time patterns from the raw temperature data sequence acquired by the data acquisition module. It uses stacked gated cyclic units (GRUs) for time series modeling and outputs a time feature vector that characterizes the time evolution of the temperature field of the traction system.

[0011] The feature fusion module is used to receive and fuse the spatial feature vector output by the spatial branch module and the temporal feature vector output by the temporal branch module, and output a fused feature vector that characterizes the spatiotemporal evolution of the temperature field of the traction system.

[0012] The fully connected layer module is used to receive the fused feature vector output by the feature fusion module and perform nonlinear mapping to output a high-order, high-dimensional feature vector characterizing the spatiotemporal coupling characteristics of the temperature field of the traction system.

[0013] The prediction output module is used to receive the high-order, high-dimensional feature vectors output by the fully connected layer module and predict the temperature field of the traction system at future moments based on the high-order, high-dimensional feature vectors.

[0014] Preferably, the data acquisition module synchronously records the raw temperature data of each measuring point at the same time through thermocouples deployed at different component measuring points of the traction system, synchronously acquires the operating condition information of the traction system, and outputs the raw temperature data sequence and operating condition information dataset, specifically including:

[0015] Each thermocouple at the measuring point is... Temperature data is sampled at fixed time intervals of seconds to form a raw temperature data sequence. ;

[0016] For the original temperature data sequence Perform resampling processing and set a uniform sampling frequency to make the original temperature data comparable in the time dimension;

[0017] The original temperature data sequence after resampling Standardization is performed to eliminate the influence of dimensions. The standardized calculation formula is as follows:

[0018] ;

[0019] in, Indicates the first The original temperature values ​​of each measuring point after standardization. For the first Original temperature values ​​at each measuring point and They represent the first Mean and standard deviation of the original temperature data at each measuring point;

[0020] Output the normalized raw temperature data sequence and operating condition information dataset.

[0021] Preferably, the frequency domain transformation module performs a Fast Fourier Transform (FFT) on the standardized raw temperature data sequence output by the data acquisition module to extract frequency domain feature vectors that characterize the essential laws and potential periodic patterns of temperature changes, and generates graph-structured data with spatial structure correlation, specifically including:

[0022] Perform a Fast Fourier Transform on the standardized raw temperature data sequence for each measuring point to obtain the frequency domain feature vector. The expression is:

[0023] ;

[0024] in, This indicates the Fast Fourier Transform operation. Indicates frequency; For the first The original temperature values ​​of each measuring point after standardization;

[0025] The spatial location information of each measurement point is bound to its frequency domain feature vector to construct graph nodes, and an adjacency matrix is ​​constructed to describe the connection strength between graph nodes based on the physical proximity relationship between the measurement points.

[0026] Based on graph nodes and adjacency matrices, graph structure data with spatial structural relationships is generated.

[0027] Preferably, the spatial branching module processes the graph structure data with spatial structure associations generated by the frequency domain transformation module, and uses a graph attention network to model the spatial dependencies between graph nodes, outputting a spatial feature vector characterizing the spatial distribution characteristics of the traction system's temperature field, specifically including:

[0028] For any two graph nodes in graph structure data and Calculate the attention coefficient between their feature vectors The expression is:

[0029] ;

[0030] ;

[0031] in, , Graph nodes Graph Nodes The initial feature vector, For learnable parameter matrix, This is the attention vector; For graph nodes and The original attention scores between them This is the vector concatenation operator;

[0032] Based on the calculated attention coefficients, the weighted aggregation graph nodes... Update graph nodes based on the neighbor node feature information. The feature representation is expressed as:

[0033] ;

[0034] in, Represents graph nodes The set of neighboring nodes, where act is the activation function; The updated feature vectors of the graph nodes;

[0035] The graph structure data is extracted layer by layer by a multi-layer graph attention network, and a dynamic attention mechanism is used to adaptively adjust the feature weights of graph nodes in different time periods, and finally output a spatial feature vector that represents the spatial distribution characteristics of the system temperature field.

[0036] Preferably, the time branching module learns the time pattern of the standardized raw temperature data sequence provided by the data acquisition module. It uses stacked gated recurrent units (GRUs) to capture the long-term dependencies of the standardized temperature data sequence and leverages an improved time attention mechanism (ITA) to mine the correlation between different standardized temperature data features at a specific time step, outputting a time feature vector that quantitatively represents the temporal evolution of the traction system's temperature field; specifically including:

[0037] The normalized raw temperature data sequence is input into the first-layer GRU unit to calculate the hidden state at the current time step:

[0038] ;

[0039] ;

[0040] ;

[0041] ;

[0042] in, For the first-layer GRU at time step The update gate, For the first-layer GRU at time step The reset door, For the first-layer GRU at time step The candidate hidden state, For the first-layer GRU at time step The final hidden state output, For the first-layer GRU at time step The hidden state, For time step The input temperature value, To update the gate weight matrix of the first-layer GRU, Reset the weight matrix of the gates in the first-layer GRU. This is the weight matrix of the candidate states in the first-layer GRU. Update the bias vector of the gate for the first-layer GRU. Set the bias vector for the gate of the first-level GRU reset. This is the bias vector for the candidate states of the first-layer GRU. For activation function, It is the hyperbolic tangent function;

[0043] By stacking multiple GRU units sequentially and extracting temporal features layer by layer, a preliminary time feature sequence is output. ;

[0044] Using one-dimensional convolutional blocks on the initial time feature sequence Perform local feature extraction and enhancement to uncover and strengthen local temporal dependencies in the data;

[0045] By utilizing an improved time attention mechanism, the influence of recent temperature changes is amplified by weighting based on time intervals, ultimately yielding a time feature vector. .

[0046] The preferred, improved time attention mechanism is formulated as follows:

[0047] ;

[0048] ;

[0049] in, For the hidden state of the last layer of the stacked GRU at time step t, For learnable parameter matrix, Let be the attention weight at time step t. This is the weighted representation of the time features, i.e., the final output time feature vector. .

[0050] Preferably, the feature fusion module receives and concatenates the spatial feature vector output by the spatial branch module with the temporal feature vector output by the temporal branch module, specifically including:

[0051] For spatial feature vectors With time feature vector Perform vector concatenation to obtain the initial fused feature vector. The expression is:

[0052] ;in, This is the vector concatenation operator;

[0053] For the initial fused feature vector Normalization is performed to eliminate dimensional and numerical distribution differences between different feature dimensions. The normalization expression is as follows:

[0054] ;

[0055] in, This is the normalized fused feature vector. and These are the initial fused feature vectors. The mean and standard deviation;

[0056] The feature fusion module also employs a multi-scale feature interaction mechanism to process the normalized fused feature vector. Perform cross-scale feature mining to enhance the interaction between spatiotemporal features at different scales and further improve the model's expressive power;

[0057] Output normalized fused feature vector It provides input for the fully connected layer module.

[0058] Preferably, the fully connected layer module is used to receive the normalized fused feature vector output by the feature fusion module. It then performs a nonlinear mapping to output a high-dimensional feature vector supporting temperature field prediction, specifically including:

[0059] Normalize and fuse feature vectors The input is the first fully connected layer, which performs linear transformations and activation function processing sequentially to generate intermediate feature vectors. The expression is:

[0060] ;

[0061] The intermediate feature vector output by the first fully connected layer The second fully connected layer is input to further perform nonlinear transformations and higher-order feature extraction, as expressed in the following expression:

[0062] ;

[0063] in, , These are the learnable weight matrices for two fully connected layers. , These are the learnable bias terms for two fully connected layers. For activation function, This is the intermediate feature vector after the first nonlinear transformation. This is a high-order abstract feature vector obtained through secondary feature extraction;

[0064] The fully connected layer module employs an Adam optimizer with an adaptive learning rate, dynamically adjusting the learning rate through an exponential decay mechanism. The final output is a high-order, high-dimensional feature vector representing the spatiotemporal coupling characteristics of the traction system's temperature field. .

[0065] Preferably, the prediction output module predicts the temperature field of the traction system at future moments based on the high-order, high-dimensional feature vector output by the fully connected layer module, specifically including:

[0066] The high-order, high-dimensional feature vector output by the fully connected layer module The input / output layer uses linear regression to obtain the predicted temperature value, expressed as follows:

[0067] ;

[0068] in, Predicting temperature values ​​for a single time step. The output layer can be learned weights. For learnable bias terms of the output layer;

[0069] After multi-time-step prediction, the system outputs a sequence of future temperature values ​​for the traction system. This sequence is then compared with the corresponding actual temperature values ​​to calculate the error. The mean square error is used as the error evaluation function, expressed as follows:

[0070] ;

[0071] in, To predict the sample size, For the first The true temperature value of each sample For the first The predicted temperature values ​​for each sample, where ∑ represents the summation operation. Mean squared error;

[0072] Based on the calculated mean squared error (MSE), the model parameters of the entire neural network are iteratively optimized through the backpropagation algorithm, and the prediction error is controlled within a preset range of 0.3℃ to 0.8℃.

[0073] In another aspect, the present invention provides a method for real-time prediction of the temperature field of a traction system based on a hybrid deep neural network, comprising the following steps:

[0074] S1. Data Acquisition and Preprocessing: The raw temperature data of key components of the traction system are collected synchronously through thermocouple sensors, and the operating condition data of the traction system are collected synchronously. The raw temperature data are resampled and standardized in sequence, and the standardized temperature data sequence and operating condition information dataset are output.

[0075] S2. Construction of Hybrid Deep Neural Network Model: Build a dual-branch hybrid deep neural network model. The dual branches include a spatial branch and a temporal branch, which extract the spatial and temporal features of the temperature field, respectively. After feature fusion, the prediction results are output through a fully connected layer.

[0076] S3. Model Training and Optimization: The mean squared error is used as the loss function, and the Adam optimizer with exponential decay mechanism is used to train the model. Data augmentation, multi-scale feature interaction, dynamic attention and early stopping mechanism are introduced to optimize the model performance and control the prediction error within the preset range.

[0077] S4. Temperature Field Prediction: Input the standardized temperature data sequence into the trained and optimized model, and output the future temperature field prediction sequence of the traction system.

[0078] The beneficial effects of this invention are:

[0079] This invention constructs a complete temperature field prediction process for a traction system by building a data acquisition module, a frequency domain transformation module, a spatial branching module, a time branching module, a feature fusion module, a fully connected layer module, and a prediction output module. The traction system first uses thermocouple sensors to collect raw temperature data from multiple key components and measuring points, simultaneously acquiring operating condition information such as speed, torque, and load. After standardization to eliminate dimensional differences and resampling to unify the time scale, the consistency and reliability of the input data are improved from the source. Subsequently, the frequency domain transformation module performs a Fast Fourier Transform on the standardized raw temperature data to accurately extract the essential laws and potential periodic characteristics of temperature changes, and combines this with the spatial location information of the measuring points to construct graph-structured data, intuitively mapping the spatial distribution relationship of the temperature field. The spatial branching module uses a Graph Attention Network (GAT) to adaptively model the complex spatial dependencies and heat conduction relationships between graph nodes, efficiently mining the spatial distribution characteristics of the temperature field. The time branching module captures data through stacked gated recurrent units (GRUs). By capturing long-term dependencies in time series data and employing an improved temporal attention mechanism (ITA) to precisely focus on key time step features, the system enhances its ability to extract temporal evolution patterns. The feature fusion module performs vector concatenation and normalization of spatial and temporal features, while introducing a multi-scale feature interaction mechanism to deepen cross-dimensional feature complementarity and collaborative expression, thereby improving feature representation accuracy. The fully connected layer module optimizes feature combinations through multi-layer nonlinear mapping and dynamically adjusts the learning rate using an Adam optimizer with exponential decay, significantly improving model training convergence efficiency and stability. The prediction output module performs linear regression operations based on the optimized fused features, ultimately achieving high-precision real-time prediction of the traction system temperature field with a prediction error of less than 0.5℃, providing reliable technical support for preventative maintenance and fault early warning of equipment. Attached Figure Description

[0080] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0081] Figure 1 This is a schematic diagram of a real-time temperature field prediction system for a traction system based on a hybrid deep neural network, as described in an embodiment of the present invention.

[0082] Figure 2 This is a schematic diagram of the temperature field prediction model architecture for a dual-branch traction system in an embodiment of the present invention.

[0083] Figure 3 This is a structural diagram of the GRU stacking module in an embodiment of the present invention.

[0084] Figure 4This is a comparison chart of the training loss of various ablation schemes in a 5-minute prediction scenario. Figure 5 This is a comparison chart of the training loss of various ablation schemes in a 15-minute prediction scenario. Figure 6 This is a comparison chart of the training loss of various ablation schemes in a 25-minute prediction scenario. Detailed Implementation

[0085] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0086] One embodiment of the present invention provides a real-time temperature field prediction system for a traction system based on a hybrid deep neural network, with reference to... Figure 1 , Figure 2 , Figure 3 , Figure 4 It includes: a data acquisition module, a frequency domain transformation module, a spatial branching module, a temporal branching module, a feature fusion module, a fully connected layer module, and a prediction output module.

[0087] The data acquisition module synchronously records the raw temperature data of each measuring point at the same time through thermocouples deployed at different component measuring points of the traction system. It also synchronously acquires the operating condition information of the traction system and outputs the raw temperature data sequence and operating condition information dataset, specifically including:

[0088] Each thermocouple at the measuring point is... Temperature samples are taken at fixed time intervals of seconds to form a raw temperature data sequence;

[0089] The original temperature data sequence is resampled, and a uniform sampling frequency is set to make the original temperature data comparable in the time dimension.

[0090] The resampled original temperature data sequence is standardized to eliminate the influence of dimensions. The standardization calculation formula is as follows:

[0091] ;

[0092] in, Indicates the first The original temperature values ​​of each measuring point after standardization. For the first Original temperature values ​​at each measuring point and They represent the first Mean and standard deviation of the original temperature data at each measuring point;

[0093] Output the standardized raw temperature data sequence and operating condition information dataset.

[0094] The frequency domain transformation module performs a Fast Fourier Transform (FFT) on the standardized raw temperature data sequence output by the data acquisition module. This extracts frequency domain feature vectors representing the essential laws and potential periodic patterns of temperature changes, and generates graph-structured data with spatial structural correlations. Specifically, this includes:

[0095] Perform a Fast Fourier Transform (FFT) on the standardized raw temperature data sequence at each measurement point to obtain the frequency domain feature vector. The expression is:

[0096] ;

[0097] in, This indicates the Fast Fourier Transform operation. Indicates frequency; For the first The original temperature values ​​of each measuring point after standardization;

[0098] The spatial location information of each measurement point is bound to its frequency domain feature vector to construct graph nodes. Based on the physical proximity relationship between measurement points, an adjacency matrix is ​​constructed to describe the connection strength between nodes. Based on the graph nodes and the adjacency matrix, graph structure data with spatial structure association is generated.

[0099] The spatial branching module processes the graph structure data with spatial structure associations generated by the frequency domain transformation module, and uses a graph attention network to model the spatial dependencies between graph nodes, outputting a spatial feature vector characterizing the spatial distribution characteristics of the traction system's temperature field, specifically including:

[0100] For any two nodes in a graph structure data and Calculate the attention coefficient between their feature vectors The expression is:

[0101] ;

[0102] ;

[0103] in, , Graph nodes Graph Nodes The initial feature vector, For learnable parameter matrix, This is the attention vector; For graph nodes and The original attention scores between them This is the vector concatenation operator;

[0104] Based on the calculated attention coefficients, the weighted aggregation graph nodes... Update the node with the characteristic information of its neighboring nodes. The feature representation is expressed as:

[0105] ;

[0106] in, Represents graph nodes The set of neighboring nodes, where act is the activation function; The updated feature vectors of the graph nodes;

[0107] The spatial branching module extracts spatial features from graph structure data layer by layer through a multi-layer graph attention network, and adopts a dynamic attention mechanism to adaptively adjust the feature weight allocation of graph nodes in different time periods, and finally outputs a spatial feature vector that represents the spatial distribution characteristics of the system temperature field.

[0108] The time branching module learns complex time patterns from the standardized raw temperature data sequence provided by the data acquisition module, referencing... Figure 2 This method employs stacked gated recurrent units (GRUs) to capture long-term dependencies in standardized temperature data sequences and utilizes an improved time attention mechanism (ITA) to mine the correlations between different standardized temperature data features at specific time steps, outputting a time feature vector that quantitatively characterizes the temporal evolution of the traction system's temperature field; specifically including:

[0109] The normalized raw temperature data sequence is input into the first-layer GRU unit to calculate the hidden state at the current time step:

[0110] ;

[0111] ;

[0112] ;

[0113] ;

[0114] in, For the first-layer GRU at time step The update gate, For the first-layer GRU at time step The reset door, For the first-layer GRU at time step The candidate hidden state, For the first-layer GRU at time step The final hidden state output, For the first-layer GRU at time step The hidden state, For time step The input temperature value, To update the gate weight matrix of the first-layer GRU, Reset the weight matrix of the gates in the first-layer GRU. This is the weight matrix of the candidate states in the first-layer GRU. Update the bias vector of the gate for the first-layer GRU. Set the bias vector for the gate of the first-level GRU reset. This is the bias vector for the candidate states of the first-layer GRU. For activation function, It is the hyperbolic tangent function;

[0115] By stacking multiple GRU units sequentially and extracting temporal features layer by layer, a preliminary time feature sequence is output. Using one-dimensional convolutional blocks for temporal feature sequences Local feature extraction and enhancement are performed to uncover and strengthen local temporal dependencies in the data.

[0116] The time branching module utilizes an improved time attention mechanism (ITA) to weight the impact of recent temperature changes based on time intervals, ultimately obtaining a time feature vector. .

[0117] The feature fusion module receives and concatenates the spatial feature vector output by the spatial branch module with the temporal feature vector output by the temporal branch module, specifically including:

[0118] For spatial feature vectors With time feature vector Perform vector concatenation to obtain the initial fused feature vector. The expression is:

[0119] ;in, This is the vector concatenation operator;

[0120] For the initial fused feature vector Normalization is performed to eliminate dimensional and numerical distribution differences between different feature dimensions. The normalization expression is as follows:

[0121] ;

[0122] in, This is the normalized fused feature vector. and These are the initial fused feature vectors. The mean and standard deviation;

[0123] The feature fusion module also employs a multi-scale feature interaction mechanism to process the normalized fused feature vector. Perform cross-scale feature mining to enhance the interaction between spatiotemporal features at different scales and further improve the model's expressive power;

[0124] Output normalized fused feature vector It provides input for the fully connected layer module.

[0125] The fully connected layer module is used to receive the normalized fused feature vector output by the feature fusion module. It then performs a nonlinear mapping to output a high-dimensional feature vector supporting temperature field prediction, specifically including:

[0126] Normalize and fuse feature vectors The input is the first fully connected layer, which performs linear transformations and activation function processing sequentially to generate intermediate feature vectors. The expression is:

[0127] ;

[0128] The intermediate feature vector output by the first fully connected layer The second fully connected layer is input to further perform nonlinear transformations and higher-order feature extraction, as expressed in the following expression:

[0129] ;

[0130] in, , These are the learnable weight matrices for two fully connected layers. , These are the learnable bias terms for two fully connected layers. For activation function, This is the intermediate feature vector after the first nonlinear transformation. This is a high-order abstract feature vector obtained through secondary feature extraction;

[0131] The fully connected layer module employs an Adam optimizer with an adaptive learning rate, dynamically adjusting the learning rate through an exponential decay mechanism. The final output is a high-order, high-dimensional feature vector representing the spatiotemporal coupling characteristics of the temperature field in the traction system. .

[0132] The prediction output module predicts the future temperature field of the traction system based on the high-order, high-dimensional feature vectors output by the fully connected layer module, specifically including:

[0133] The high-order, high-dimensional feature vector output by the fully connected layer module The input / output layer uses linear regression to obtain the predicted temperature value, expressed as follows:

[0134] ;

[0135] in, Predicting temperature values ​​for a single time step. The output layer can be learned weights. For learnable bias terms of the output layer;

[0136] After multi-time-step prediction, the system outputs a sequence of future temperature values ​​for the traction system. This sequence is then compared with the corresponding actual temperature values ​​to calculate the error. The mean square error is used as the error evaluation function, expressed as follows:

[0137] ;

[0138] in, To predict the sample size, For the first The true temperature value of each sample For the first The predicted temperature values ​​for each sample, where ∑ represents the summation operation. Mean squared error;

[0139] The prediction output module uses the calculated mean squared error (MSE) to iteratively optimize the model parameters of the entire neural network through the backpropagation algorithm, thereby controlling the prediction error within a preset range of 0.3℃ to 0.8℃.

[0140] Another aspect of this invention provides a method for real-time prediction of the temperature field of a traction system based on a hybrid deep neural network, the specific implementation steps of which are as follows:

[0141] First, perform data acquisition and preprocessing step S1:

[0142] Step S1.1, Data Acquisition

[0143] This embodiment uses high-precision thermocouple sensors to monitor the temperature of key heat-generating components in the traction system, and simultaneously collects traction system operating condition data (including speed, torque, load, etc.) as model input features to construct a complete temperature field prediction dataset.

[0144] The specific arrangement, quantity, and corresponding data of the thermocouples are shown in Table 1, comprehensively covering the core heat-generating components of the traction system to ensure that the collected data can accurately represent the temperature field distribution of the traction system.

[0145] Table 1

[0146]

[0147] By using the thermocouples distributed in various key components, the raw temperature data of each measuring point in the traction system can be acquired in real time and synchronously. Combined with the operating condition data such as speed, torque, and load during the operation of the traction system, the raw operating condition dataset is formed, providing timely, comprehensive and critical data support for subsequent temperature field prediction.

[0148] Step S1.2: Data Preprocessing

[0149] The raw temperature data collected may have issues such as inconsistent sampling frequencies and large differences in the units of data from different measuring points, which can affect the training effect and prediction accuracy of the hybrid deep neural network model. Therefore, the raw temperature data is resampled and standardized sequentially. The specific steps are as follows:

[0150] Step S1.2.1, Resampling Processing: Set a uniform sampling frequency (sampling once every 60 seconds in this embodiment) and resample the original temperature data of different measuring points and different sampling frequencies to make all temperature data comparable in the time dimension and ensure the temporal consistency of the input data of the hybrid deep neural network model.

[0151] Step S1.2.2, Standardization Processing: A standardization formula is used to eliminate dimensional differences between different measurement points, improving the stability of the prediction model training. The standardization calculation formula is as follows:

[0152] ;

[0153] in, Indicates the first The original temperature values ​​of each measuring point after standardization. For the first Original temperature values ​​at each measuring point and They represent the first Mean and standard deviation of the original temperature data at each measuring point;

[0154] Through the above data acquisition and preprocessing steps, a standardized temperature data sequence and operating condition information dataset are obtained, ensuring high-quality input data and providing a reliable data foundation for the subsequent construction, training, and prediction of hybrid deep neural network models.

[0155] Next, perform step S2 of building the hybrid deep neural network model:

[0156] After completing the data acquisition and preprocessing in step S1, this embodiment constructs a hybrid deep neural network model. This model adopts a dual-branch architecture with parallel spatial and temporal branches to extract the spatial and temporal features of the traction system temperature field, respectively. After feature fusion, the prediction results are output through a fully connected layer. The specific implementation is as follows:

[0157] Step S2.1: Design of a dual-branch hybrid deep neural network model architecture

[0158] like Figure 2 As shown, the dual-branch hybrid deep neural network model built in this embodiment includes two parallel branches, a spatial branch and a temporal branch, as well as a feature fusion module and a fully connected layer module. All parts work collaboratively.

[0159] Spatial branch: mainly responsible for extracting the spatial distribution characteristics of the raw temperature data, modeling the spatial dependence between different thermocouple measuring points, and capturing the spatial coupling characteristics of the temperature field of the traction system.

[0160] Time branch: mainly responsible for capturing the time pattern of temperature changes, extracting the time evolution law of temperature data, and exploring the impact of raw temperature data on future temperature changes;

[0161] Feature fusion module: Receives feature vectors from the spatial and temporal branches, completes feature fusion, and outputs a fused feature vector characterizing the spatiotemporal evolution of the temperature field of the traction system;

[0162] Fully connected layer module: performs nonlinear mapping on the fused feature vectors and outputs high-order, high-dimensional feature vectors that support temperature field prediction, providing feature support for subsequent prediction outputs.

[0163] Step S2.2, Specific Implementation of Spatial Branching

[0164] The spatial branch uses a graph attention network (GAT) to model the spatial dependencies between thermocouple measuring points. Combined with the frequency domain features extracted by Fourier transform, it effectively extracts the spatial distribution characteristics of the traction system's temperature field. The specific implementation steps are as follows:

[0165] Step S2.2.1: Extract frequency domain features using Fourier transform:

[0166] Perform a Fourier transform on the standardized temperature data sequence of each thermocouple measuring point obtained in step S1 to extract the essential laws and potential periodic patterns of temperature changes, and obtain the frequency domain feature representation of that measuring point, as shown in the following formula: ;

[0167] in, Indicates the first Frequency domain characteristics of individual thermocouple measurement points This indicates the Fast Fourier Transform operation. Indicates frequency; For the first The original temperature values ​​of each measuring point after standardization;

[0168] Step S2.2.2: Construct graph data:

[0169] The spatial location information of each thermocouple measuring point (i.e., the coordinate information of the thermocouple arrangement) is bound to the frequency domain feature vector extracted in step S2.2.1 to construct graph nodes, as shown in the following formula: ;

[0170] in, Indicates the first Each graph node The frequency domain characteristics of the nodes in this graph are... This refers to the spatial location information of the nodes in the graph.

[0171] Meanwhile, based on the physical proximity between each thermocouple measuring point, an adjacency matrix is ​​constructed to describe the connection strength between different graph nodes. The closer the measuring points are, the higher the connection strength of the corresponding graph nodes, ultimately forming graph structure data with spatial structure association.

[0172] Step S2.2.3, Graph Attention Network (GAT) Calculation:

[0173] By utilizing a Graph Attention Network (GAT) to propagate and aggregate information from each graph node layer by layer along the constructed graph structure, the complex spatial dependencies between different thermocouple graph nodes are automatically captured, achieving deep extraction of spatial features. The calculation formula for the Graph Attention Network is as follows:

[0174] ;

[0175] ;

[0176] in, Represents graph nodes Graph Nodes Attention coefficient between them and Representing graph nodes Graph Nodes The initial feature vector (i.e., graph node) Graph Nodes (feature representation) For learnable parameter matrix, For attention vectors, Represents graph nodes The set of neighboring graph nodes, The activation function is (in this embodiment, the Sigmoid function is used). For graph nodes The feature vector updated after feature aggregation.

[0177] Step S2.2.4, Spatial Feature Output:

[0178] The spatial features are extracted layer by layer by a multi-layer graph attention network (GAT) (two-layer GAT is used in this embodiment), and finally the spatial feature vector S representing the spatial distribution characteristics of the temperature field of the traction system is output and transmitted to the feature fusion module.

[0179] Step S2.3, Specific Implementation of Time Branching

[0180] The time branch employs a stacked gated recurrent unit (GRU) combined with an improved temporal attention mechanism (ITA) to learn the complex temporal patterns of the raw temperature data, mine the most critical temporal dynamics for temperature prediction at a specific time step, and complete the extraction of temporal features. The specific implementation steps are as follows:

[0181] Step S2.3.1, Stacking GRU module operations:

[0182] The standardized raw temperature data sequence obtained in step S1 is input into the first-layer GRU unit to calculate the hidden state at the current time step and capture the short-term temporal dependencies of the raw temperature data. In this embodiment, three layers of GRU units are stacked sequentially, and after layer-by-layer temporal feature extraction, a preliminary temporal feature sequence is output. The calculation formula for the first-layer GRU cell is as follows:

[0183] ;

[0184] ;

[0185] ;

[0186] ;

[0187] in, For the first-layer GRU at time step The update gate, For the first-layer GRU at time step The reset door, For the first-layer GRU at time step The candidate hidden state, For the first-layer GRU at time step The final hidden state output, For the first-layer GRU at time step The hidden state, For time step The input temperature value, To update the gate weight matrix of the first-layer GRU, Reset the weight matrix of the gates in the first-layer GRU. This is the weight matrix of the candidate states in the first-layer GRU. Update the bias vector of the gate for the first-layer GRU. Set the bias vector for the gate of the first-level GRU reset. This is the bias vector for the candidate states of the first-layer GRU. For activation function, It is the hyperbolic tangent function;

[0188] Step S2.3.1, Processing one-dimensional convolutional blocks:

[0189] Preliminary temporal feature sequences of the output of stacked GRU units using one-dimensional convolutional blocks. Local feature extraction and enhancement are performed to uncover and strengthen the local temporal dependencies in the original temperature data, resulting in the convolutional feature matrix, as shown in the following formula: ;

[0190] in, The feature matrix after convolution. Preliminary time feature sequence;

[0191] Step S2.3.2, Improved Temporal Attention (ITA) Mechanism Calculation:

[0192] By using an improved time attention mechanism, the correlation between different temperature field data features at a specific time step is explored. The influence of recent temperature changes is strengthened by weighting the data according to the time interval, as shown in the following formula:

[0193] ;

[0194] ;

[0195] in, For the hidden state of the last layer of the stacked GRU at time step t, For learnable parameter matrix, Let be the attention weight at time step t. This is the weighted representation of the time features, i.e., the final output time feature vector. .

[0196] Step S2.4: Feature fusion and fully connected layer mapping

[0197] After obtaining the spatial and temporal feature vectors, feature fusion is performed through a feature fusion module, followed by nonlinear mapping through a fully connected layer module. The specific implementation steps are as follows:

[0198] Step S2.4.1, Feature splicing:

[0199] The spatial feature vector output by the spatial branch With the time feature vector output by the time branch Perform vector concatenation to obtain the initial fused feature vector. The formula is:

[0200] ;in, This is the vector concatenation operator;

[0201] Step S2.4.2, Normalization:

[0202] For the initial fused feature vector Normalization is performed to eliminate differences in dimensionality and numerical distribution between different feature dimensions, ensuring stable operation of subsequent fully connected layer modules. The normalization expression is:

[0203] ;

[0204] in, This is the normalized fused feature vector. and These are the initial fused feature vectors. The mean and standard deviation;

[0205] Step S2.4.3, Multi-scale feature interaction:

[0206] A multi-scale feature interaction mechanism is introduced to process the normalized fused feature vector. Cross-scale feature mining is conducted to enhance the interaction between spatiotemporal features at different scales and further improve the model's feature representation capabilities.

[0207] Step S2.4.4, Fully Connected Layer Mapping:

[0208] The normalized fused feature vector The input fully connected layer module (two fully connected layers are used in this embodiment) is subjected to nonlinear mapping to extract high-order, high-dimensional features, providing support for subsequent temperature prediction. The specific formula is as follows:

[0209] ;

[0210] ;

[0211] in, , These are the learnable weight matrices for two fully connected layers. , These are the learnable bias terms for two fully connected layers. For activation function, This is the intermediate feature vector after the first nonlinear transformation. This refers to the high-order abstract feature vector obtained through secondary feature extraction, i.e., the high-order, high-dimensional feature vector output by the fully connected layer module. .

[0212] Through the above steps, the hybrid deep neural network model constructed in this embodiment can effectively integrate the spatiotemporal characteristics of the temperature field of the traction system, comprehensively characterize the spatiotemporal coupling characteristics of the temperature field, lay the foundation for achieving high-precision temperature prediction, and provide reliable technical support for real-time monitoring and fault early warning of the train traction system.

[0213] Step S3: Training and Optimization of Hybrid Deep Neural Network Model

[0214] After completing the construction of the hybrid deep neural network model in step S2, the hybrid deep neural network model is trained and optimized to ensure that it can accurately capture the dynamic changes of the traction system's temperature field over time and operating conditions, effectively reduce prediction errors, accelerate the temperature prediction process, and ensure the stable operation of the traction system. The specific implementation is as follows:

[0215] Step S3.1: Loss Function and Optimizer Selection

[0216] This embodiment uses mean squared error (MSE) as the loss function to measure the error between the predicted value and the actual temperature value of the hybrid deep neural network model. MSE directly affects the training effect and prediction accuracy of the hybrid deep neural network model. The calculation formula is as follows:

[0217] ;

[0218] in, To predict the sample size, For the first The true temperature value of each sample For the first The predicted temperature values ​​for each sample, where ∑ represents the summation operation. Mean squared error;

[0219] Mean squared error (MSE) can effectively quantify how well a prediction model approximates the actual temperature. It is suitable for continuous value regression tasks such as temperature prediction and has good applicability.

[0220] This embodiment employs the Adam optimizer with an adaptive learning rate, combined with an exponential decay mechanism to dynamically adjust the learning rate, addressing the instability issue during the training of hybrid deep neural network models. This results in rapid model convergence and strong generalization ability. The Adam optimizer combines the advantages of momentum and RMSProp, dynamically adjusting the learning rate of each parameter based on the first and second moment estimates of the gradient. The exponential decay mechanism continuously refines the learning rate, ensuring the model reaches convergence quickly and stably. Its calculation formula is as follows:

[0221] ;

[0222] in, This represents the initial learning rate (set to 0.001 in this embodiment). The attenuation rate (set to 0.0001 in this embodiment) This represents the number of iterations.

[0223] The exponential decay mechanism can effectively prevent the learning rate from becoming too large in the later stages of training, thereby avoiding oscillations of the hybrid deep neural network model around the optimal solution and improving the stability of training.

[0224] Step S3.2, Training Process and Parameter Settings

[0225] This embodiment uses actual collected traction system operation data as the experimental dataset. The specific training process and parameter settings are as follows:

[0226] The experimental dataset in this embodiment contains 98,000 data samples, spanning 3 months. The data from the last 5 days is used as the test set (for subsequent performance verification of the hybrid deep neural network model), and the remaining data is used as the training set (for training the hybrid deep neural network model). The ratio of the training set to the test set is approximately 9:1.

[0227] During training, the hybrid deep neural network model employs a mini-batch training strategy. In each iteration, 32 data samples are randomly selected for forward and backward propagation to reduce computational overhead and accelerate convergence. After each iteration, the loss value of the hybrid deep neural network model is calculated on both the training and validation sets (10% of the training set is used as the validation set). Evaluation metrics such as MAE (mean absolute error), MSE (mean squared error), and R² (goodness of fit) are recorded to assess the predictive performance of the hybrid deep neural network model in real time. If the loss value of the hybrid deep neural network model on the validation set does not decrease significantly (the decrease is less than 0.0001) for 10 consecutive iterations, training is terminated early to prevent overfitting and improve training efficiency and model generalization ability.

[0228] Step S3.3, Model Optimization Strategy

[0229] To further improve the prediction performance, generalization ability, and stability of the hybrid deep neural network model, this embodiment adopts the following optimization strategy:

[0230] Data augmentation: During training, random noise and time-series perturbations are injected into the original temperature data to enhance the generalization ability of the hybrid deep neural network model and improve its adaptability to temperature changes under different operating conditions.

[0231] Multi-scale Feature Interaction: In the feature fusion module, a multi-scale feature interaction mechanism is introduced. By extracting fused features through convolutional kernels of different sizes, the interaction between features of different scales is enhanced, further improving the feature representation capability of the hybrid deep neural network model.

[0232] Dynamic Attention Mechanism: In Graph Attention Network (GAT) and Improved Temporal Attention Mechanism (ITA), a dynamic attention mechanism is adopted. Based on the temperature change characteristics of different operating stages of the traction system (start-up stage, stable operation stage, shutdown stage), the weight allocation in different time periods is adaptively adjusted, enabling the hybrid deep neural network model to more flexibly focus on the features that are key to the prediction results.

[0233] Early Stopping Mechanism: During training, if the loss value of the hybrid deep neural network model on the validation set does not decrease significantly for 10 consecutive rounds, training is terminated early to prevent overfitting of the hybrid deep neural network model and improve training efficiency.

[0234] Through the aforementioned training and optimization strategies, the hybrid deep neural network model constructed in this embodiment can converge quickly and exhibit excellent performance in the traction system temperature field prediction task. Experimental results show that the training loss of this hybrid deep neural network model in the 5-minute prediction scenario is only about 0.003, and the MAE, MSE, and R² indices are 0.761, 1.342, and 0.965, respectively, demonstrating good prediction performance and meeting the requirements for high-precision prediction of traction system temperature fields.

[0235] Step S4: Temperature Field Prediction

[0236] The standardized raw temperature data sequence and operating condition information from the test set are input into the trained and optimized hybrid deep neural network model. The hybrid deep neural network model extracts spatial and temporal features through spatial and temporal branches, respectively. After feature fusion and nonlinear mapping of the fully connected layer, it outputs the temperature field prediction temperature data sequence of the traction system for future multiple time steps (5 minutes, 15 minutes, and 25 minutes in this embodiment), specifically the future temperature prediction values ​​of each key component and each thermocouple measuring point.

[0237] To verify the effectiveness and superiority of the embodiments of the present invention, this embodiment comprehensively verifies the prediction performance of the hybrid deep neural network model through three types of experiments: ablation experiments, comparative experiments, and multi-monitoring point prediction comparisons, as detailed below:

[0238] Step S4.1, Ablation Experiment

[0239] This embodiment designs five ablation experimental schemes to analyze the impact of different model structures on prediction performance. The five schemes are as follows:

[0240] Option 1: (Method in this embodiment): Dual-branch hybrid deep neural network model (spatial branch GAT + temporal branch stacked GRU + ITA + multi-scale fusion).

[0241] Option 2: Single spatial branch model (only the spatial branch GAT is retained, and the temporal branch is removed).

[0242] Option 3: Single temporal branch model (only temporal branches are stacked in GRU+ITA, and spatial branches are removed).

[0243] Option 4: Single spatial branch model without attention (removing the GAT attention mechanism from Option 2);

[0244] Option 5: Single-time branch non-attention model (removing the ITA attention mechanism based on Option 3).

[0245] The training loss differences of five ablation schemes were compared and analyzed under three prediction scenarios: 5 minutes, 15 minutes, and 25 minutes. The experimental results are as follows: Figure 4 As shown.

[0246] right Figure 4 , Figure 5 , Figure 6 Analysis shows that as the model training process continues, the training loss curve exhibits a significant and stable downward trend. In the ablation experiment comparison, Schemes 4 and 5, due to their single-branch network architecture and lack of integrated attention mechanism, suffer from limited feature extraction capabilities and persistently high training loss values ​​(both greater than 0.04). Although Schemes 2 and 3 introduce basic attention modules, effectively enhancing the model's ability to capture key features, their convergence is still unsatisfactory due to their feature modeling methods that only consider a single spatial or temporal dimension, resulting in a high level of training loss. In contrast, Scheme 1, by constructing a spatiotemporal dual-dimensional feature fusion mechanism, simultaneously captures the coupling characteristics of the traction system's temperature field in both spatial distribution and temporal evolution, significantly reducing the training loss of the hybrid deep neural network model. Specifically, for the task of predicting the temperature field in the next 5 minutes, the training loss value is as low as 0.003. The experimental data fully validate the technical advantages of the dual-branch hybrid deep neural network model of this invention in the dynamic prediction of the traction system's temperature field.

[0247] Step S4.2, Comparative Experiment

[0248] To further verify the effectiveness of the method of this invention, two existing mainstream temperature prediction methods were used as comparative methods and tested together with the method of this invention. The differences in the values ​​of mean squared error (MSE), mean absolute error (MAE), and goodness of fit (R²) of the three prediction methods were compared to further verify the superiority of the method of this invention.

[0249] Existing Method 1: Temperature prediction method based on Support Vector Regression (SVR) predictor;

[0250] Existing Method 2: Temperature prediction method based on recurrent neural networks (RNN);

[0251] The method of this invention: a prediction method based on hybrid deep neural networks.

[0252] Table 2 shows the performance comparison of the three methods in prediction scenarios at 5 minutes, 15 minutes, and 25 minutes (the smaller the MSE value, the stronger the model stability; the smaller the MAE value, the higher the prediction accuracy; the closer R² is to 1, the better the model fit).

[0253] Table 2

[0254]

[0255] The quantitative analysis based on Table 2 shows that, under the three prediction scenarios, the method of this invention significantly outperforms the comparative methods in terms of MSE, MAE, and R², fully verifying the stability advantage and accuracy improvement of the method of this invention in the prediction of the temperature field of the traction system. Specifically, in the 5-minute short-term prediction scenario, the method of this invention has an MSE as low as 1.342, an MAE of only 0.761, and an R² of 0.965, which is an order of magnitude lower than the error index of the medium- and long-term prediction scenarios (15 / 25 minutes). It has excellent short-term prediction accuracy, can meet the needs of real-time status monitoring of the traction system, and has significant engineering application value.

[0256] Step S4.3, 3: Comparison of temperature field prediction at multiple monitoring points

[0257] Using test samples as input to the prediction model, the temperatures of key components and monitoring points of the traction system were predicted in a 5-minute prediction scenario. The comparison between the prediction results and the actual temperatures is shown in Table 3.

[0258] Table 3

[0259]

[0260] As shown in Table 3, the method of the present invention can accurately predict the temperature field of each key component and monitoring point of the traction system. The error between the predicted and actual temperature values ​​at each monitoring point is ≤0.5℃, and the prediction error is controlled within the preset range. This indicates that the method of the present invention has extremely high accuracy in predicting the temperature field of the traction system. It can capture the temperature change trend of different key components on a 5-minute time scale, providing reliable data for real-time monitoring of the temperature field of the traction system, early warning of faults, and preventive maintenance.

[0261] This invention also provides a real-time temperature field prediction system for traction systems based on hybrid deep neural networks for engineering applications. Focusing on the actual deployment needs of engineering projects, it achieves high reliability, strong robustness, and good maintainability of real-time temperature field prediction through a layered modular architecture design, engineering adaptation optimization, and full-process operation and maintenance support. It can be widely used in intelligent monitoring and thermal management scenarios of complex traction systems such as rail transit and electric vehicles, providing technical support for the safe operation, condition assessment, and preventive maintenance of key components of traction systems.

[0262] The real-time temperature field prediction system for the traction system provided in this embodiment adopts a layered modular architecture. Each layer is independently encapsulated and can be maintained and upgraded separately. The functional boundaries and data flow paths of each layer are clearly defined, ensuring the stable operation and flexible adaptation of the real-time temperature field prediction system for the traction system in complex engineering environments. It includes a data perception layer, a feature processing layer, a model calculation layer, and an application service layer. Each layer works together to complete the entire process of real-time temperature field prediction for the traction system. The specific functions and operating logic are as follows:

[0263] As the core data input for the real-time temperature field prediction system of the traction system, the data sensing layer collects raw temperature data from various measuring points in real time through a multi-channel thermocouple sensor network distributed on key components of the traction system. Simultaneously, it acquires key operating condition information such as speed, load, and current during the traction system's operation. All collected multi-source data is uniformly connected to the edge acquisition unit via an industrial communication bus, achieving time synchronization and preliminary verification of the data. To adapt to the complex communication environment of engineering sites, the data sensing layer also integrates data caching and breakpoint resume functions. It can temporarily store collected data when communication is interrupted or abnormal, and automatically upload the data after communication is restored, ensuring the integrity, timeliness, and continuity of the input data, providing a high-quality data foundation for subsequent feature processing and model calculations.

[0264] The feature processing layer receives the raw temperature data sequence from the data sensing layer and is responsible for data preprocessing and feature extraction, providing input features for the model calculation layer. First, the feature processing layer resamples and aligns the raw temperature data sequence to unify the data sampling frequency and time reference, eliminating temporal misalignment caused by differences in sensor sampling. Then, it performs data standardization to eliminate dimensional differences between different measuring points and data types, improving the convergence efficiency of subsequent model training and computation. Based on this, the system initiates two parallel processing paths: one path sends the raw temperature data sequence to the frequency domain analysis unit, extracting periodic and fluctuating features through spectral transformation and constructing physically meaningful graph structure data by combining the spatial location information of the measuring points; the other path retains the raw temperature data sequence for extracting dynamic temporal features. Simultaneously, the feature processing layer incorporates a data quality monitoring mechanism that automatically identifies outliers and missing values ​​in the data and completes outlier processing and missing data completion through preset algorithms, ensuring that the data quality input to the model calculation layer meets engineering operation requirements.

[0265] The model computation layer, as the core intelligent computing unit of the real-time temperature field prediction system for the traction system, adopts a dual-branch hybrid deep neural network architecture. It is specifically designed to process the spatial distribution and temporal dynamics of the traction system's temperature field, achieving high-precision temperature field prediction. The spatial branch takes the graph-structured data output from the feature processing layer as input, utilizes a graph attention network to model the spatial correlations between measurement points, adaptively learns the intensity of heat conduction influence between key components of the traction system, and characterizes the spatial distribution pattern of the temperature field. The temporal branch takes the original temperature data sequence retained by the feature processing layer as input, employing a combination of stacked gated recurrent units and an attention mechanism to effectively extract long-term dependencies from the original temperature data sequence, focusing on capturing temperature change patterns under non-stationary scenarios such as sudden changes in operating conditions and start-up / shutdown processes. The features extracted from the spatial and temporal branches are fed into a dedicated fusion module, where normalization processing and a multi-scale feature interaction mechanism complete feature integration. Finally, the features are input into a fully connected network to complete nonlinear mapping and generate high-order, high-dimensional feature vectors, providing core computational support for subsequent traction system temperature field prediction.

[0266] The application service layer is responsible for transforming the prediction results of the dual-branch hybrid deep neural network model computation layer into engineering-usable operation and maintenance support information, realizing the visualization of prediction results, real-time early warning, and decision support. Based on the computation results of the dual-branch hybrid deep neural network model, the application service layer automatically generates temperature field prediction data for multiple time steps of the traction system, and visualizes it in intuitive forms such as temperature distribution maps, trend curves, and safety threshold comparison charts, making it easy for operation and maintenance personnel to quickly grasp the temperature status of key components of the traction system. At the same time, it has a built-in safety threshold judgment mechanism. When the predicted temperature approaches or exceeds the preset safety threshold, it automatically triggers real-time alarms and pushes alarm information to the operation and maintenance terminal, realizing real-time monitoring and risk warning of the temperature field. In addition, the application service layer also has prediction result analysis and decision support functions, which can perform in-depth analysis of prediction data and push targeted early warning prompts and maintenance suggestions to operation and maintenance personnel, providing direct technical support for intelligent monitoring, status assessment, and preventive maintenance of the traction system.

[0267] To further enhance the engineering adaptability and maintainability of the real-time temperature field prediction system for the traction system, this embodiment also integrates system-level engineering deployment and maintenance design, specifically including two parts: deployment mode optimization and maintainability design.

[0268] 1. Deployment Mode:

[0269] This embodiment supports dual-mode deployment of local edge computing and cloud-based collaborative operation, and can be flexibly adapted to the needs of different engineering scenarios.

[0270] In local edge computing mode, the dual-branch hybrid deep neural network model computing unit runs directly on vehicle-mounted or ground-based edge devices, effectively reducing data transmission latency and meeting the low-latency requirement for real-time prediction of traction system temperature field. In cloud collaboration mode, the real-time prediction system for traction system temperature field can synchronously upload collected data and prediction results to the cloud data center, building an integrated intelligent operation and maintenance system of "end-edge-cloud" to realize data archiving, global operating condition analysis and model iterative optimization, meeting the operation and maintenance management needs of large-scale traction system clusters.

[0271] 2. System maintainability design:

[0272] This embodiment incorporates a professional model management module that supports online updates and version control of dual-branch hybrid deep neural network models. It can replace, adjust parameters, and roll back versions of dual-branch hybrid deep neural network models without interrupting the system's real-time prediction service, adapting to model iteration needs in engineering scenarios. Simultaneously, it provides an open application programming interface (API) for rapid integration with existing traction system monitoring platforms, fault diagnosis systems, or digital twin systems, without requiring large-scale modifications to existing systems. This effectively enhances the engineering integration capabilities and compatibility of the traction system's real-time temperature field prediction system, reducing engineering deployment costs.

[0273] In summary, the real-time temperature field prediction system for traction systems provided in this embodiment, through a layered modular architecture design, achieves end-to-end engineering adaptation of data acquisition, feature processing, model calculation, and application services, balancing prediction accuracy, operational stability, and ease of maintenance. This embodiment can be directly deployed in complex engineering scenarios, accurately predicting and warning of the temperature field of traction systems in real time, effectively solving the monitoring challenges in the thermal management of complex traction systems, providing strong technical support for the safe, reliable, and efficient operation of traction systems, and possessing extremely high engineering application value and promising prospects for widespread adoption.

[0274] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the embodiments of the present invention have been described in detail, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of protection of the claims of the present invention.

Claims

1. A real-time temperature field prediction system for a traction system based on a hybrid deep neural network, characterized in that, include: The data acquisition module is used to synchronously record the raw temperature data of different components of the traction system at the same time through thermocouples, synchronously acquire the operating condition information of the traction system, and output the formed raw temperature data sequence and operating condition information dataset. The frequency domain transformation module is used to perform Fourier transform on the original temperature data sequence output by the data acquisition module, extract frequency domain features that characterize the essential laws and potential periodic patterns of temperature changes, and generate graph structure data with spatial structure correlation. The spatial branching module is used to process the graph structure data with spatial structure association generated by the frequency domain conversion module, and to use a graph attention network to model the spatial dependencies between graph nodes, and output a spatial feature vector characterizing the spatial distribution characteristics of the temperature field of the traction system. The time branch module is used to learn the time pattern of the raw temperature data sequence acquired by the data acquisition module. It uses a stacked gated cyclic unit (GRU) for time series modeling and outputs a time feature vector that characterizes the time evolution of the temperature field of the traction system. The feature fusion module is used to receive and fuse the spatial feature vector output by the spatial branch module and the temporal feature vector output by the temporal branch module, and output a fused feature vector that characterizes the spatiotemporal evolution of the temperature field of the traction system. The fully connected layer module is used to receive the fused feature vector output by the feature fusion module and perform nonlinear mapping to output a high-order, high-dimensional feature vector characterizing the spatiotemporal coupling characteristics of the temperature field of the traction system. The prediction output module is used to receive the high-order, high-dimensional feature vector output by the fully connected layer module, and predict the temperature field of the traction system at future times based on the high-order, high-dimensional feature vector.

2. The real-time temperature field prediction system for a traction system based on a hybrid deep neural network according to claim 1, characterized in that: The data acquisition module synchronously records the raw temperature data of each measuring point at the same time through thermocouples deployed at different component measuring points of the traction system, synchronously acquires the operating condition information of the traction system, and outputs the raw temperature data sequence and operating condition information dataset, specifically including: Each thermocouple at the measuring point is... Temperature data is sampled at fixed time intervals of seconds to form a raw temperature data sequence. ; For the original temperature data sequence Perform resampling processing and set a uniform sampling frequency to make the original temperature data comparable in the time dimension; The original temperature data sequence after resampling Standardization is performed to eliminate the influence of dimensions. The standardized calculation formula is as follows: ； in, Indicates the first The original temperature values ​​of each measuring point after standardization. For the first Original temperature values ​​at each measuring point and They represent the first Mean and standard deviation of the original temperature data at each measuring point; Output the standardized raw temperature data sequence and operating condition information dataset.

3. The real-time temperature field prediction system for a traction system based on a hybrid deep neural network according to claim 2, characterized in that: The frequency domain transformation module performs a Fourier transform on the standardized raw temperature data sequence output by the data acquisition module, extracts frequency domain feature vectors representing the essential laws and potential periodic patterns of temperature changes, and generates graph-structured data with spatial structure correlation, specifically including: Perform a Fast Fourier Transform on the standardized raw temperature data sequence for each measuring point to obtain the frequency domain feature vector. The expression is: ； in, This indicates the Fast Fourier Transform operation. Indicates frequency; For the first The original temperature values ​​of each measuring point after standardization; The spatial location information of each measurement point is bound to its frequency domain feature vector to construct graph nodes, and an adjacency matrix is ​​constructed to describe the connection strength between graph nodes based on the physical proximity relationship between the measurement points. Based on the graph nodes and the adjacency matrix, graph structure data with spatial structure association is generated.

4. The real-time temperature field prediction system for a traction system based on a hybrid deep neural network according to claim 3, characterized in that: The spatial branching module processes the graph structure data with spatial structure associations generated by the frequency domain transformation module, and uses a graph attention network to model the spatial dependencies between the graph nodes, outputting a spatial feature vector characterizing the spatial distribution characteristics of the traction system's temperature field, specifically including: For any two graph nodes in graph structure data Graph Nodes Calculate the attention coefficient between their feature vectors The expression is: ； ； in, , Graph nodes Graph Nodes The initial feature vector, For learnable parameter matrix, This is the attention vector; For graph nodes and The original attention scores between them This is the vector concatenation operator; Based on the calculated attention coefficients, the weighted aggregation graph nodes... Update graph nodes based on the neighbor node feature information. The feature representation is expressed as: ； in, Represents graph nodes The set of neighboring nodes, where act is the activation function; The updated feature vectors of the graph nodes; The graph structure data is extracted layer by layer by a multi-layer graph attention network, and a dynamic attention mechanism is adopted to adaptively adjust the feature weight allocation of graph nodes in different time periods, and finally output a spatial feature vector S representing the spatial distribution characteristics of the system temperature field.

5. The real-time temperature field prediction system for a traction system based on a hybrid deep neural network according to claim 4, characterized in that: The time branching module learns the time pattern of the standardized raw temperature data sequence provided by the data acquisition module, uses stacked gated loop units to capture the long-term dependencies of the standardized temperature data sequence, and utilizes an improved time attention mechanism to mine the correlation between different standardized temperature data features at a specific time step, outputting a time feature vector that quantitatively represents the temporal evolution of the traction system's temperature field; specifically including: The standardized raw temperature data sequence is input into the first-layer GRU unit to calculate the hidden state at the current time step: ; ; ; ; in, For the first-layer GRU at time step The update gate, For the first-layer GRU at time step The reset door, For the first-layer GRU at time step The candidate hidden state, For the first-layer GRU at time step The final hidden state output, For the first-layer GRU at time step The hidden state, For time steps The input temperature value, To update the gate weight matrix of the first-layer GRU, Reset the weight matrix of the gates in the first-layer GRU. This is the weight matrix of the candidate states in the first-layer GRU. Update the bias vector of the gate for the first-layer GRU. Set the bias vector for the gate of the first-level GRU reset. This is the bias vector for the candidate states of the first-layer GRU. For activation function, It is the hyperbolic tangent function; By stacking multiple GRU units sequentially and extracting temporal features layer by layer, a preliminary time feature sequence is output. ; The preliminary time feature sequence is processed using one-dimensional convolutional blocks. Perform local feature extraction and enhancement to uncover and strengthen local temporal dependencies in the data; By utilizing an improved time attention mechanism, the influence of recent temperature changes is amplified by weighting based on time intervals, ultimately yielding a time feature vector. .

6. The real-time temperature field prediction system for a traction system based on a hybrid deep neural network according to claim 5, characterized in that: The improved time attention mechanism is defined by the following formula: ； ； in, For the hidden state of the last layer of the stacked GRU at time step t, For learnable parameter matrix, Let be the attention weight at time step t. This is the weighted representation of the time features, i.e., the final output time feature vector. .

7. The real-time temperature field prediction system for a traction system based on a hybrid deep neural network according to claim 6, characterized in that: The feature fusion module receives the concatenation and fusion of the spatial feature vector output by the spatial branching module and the temporal feature vector output by the temporal branching module, specifically including: For spatial feature vectors With time feature vector Perform vector concatenation to obtain the initial fused feature vector. The expression is: ;in, This is the vector concatenation operator; For the initial fused feature vector Normalization is performed to eliminate dimensional and numerical distribution differences between different feature dimensions. The normalization expression is as follows: ； in, This is the normalized fused feature vector. and These are the initial fused feature vectors. The mean and standard deviation; The feature fusion module also employs a multi-scale feature interaction mechanism to process the normalized fused feature vector. Perform cross-scale feature mining to enhance the interaction between spatiotemporal features at different scales and further improve the model's expressive power; Output normalized fused feature vector It provides input for the fully connected layer module.

8. The real-time temperature field prediction system for a traction system based on a hybrid deep neural network according to claim 7, characterized in that: The fully connected layer module is used to receive the normalized fused feature vector output by the feature fusion module. It then performs a nonlinear mapping to output a high-dimensional feature vector supporting temperature field prediction, specifically including: Normalize and fuse feature vectors The input is the first fully connected layer, which performs linear transformations and activation function processing sequentially to generate intermediate feature vectors. The expression is: ； The intermediate feature vector output by the first fully connected layer The second fully connected layer is input to further perform nonlinear transformations and higher-order feature extraction, as expressed in the following expression: ； in, , These are the learnable weight matrices for two fully connected layers. , These are the learnable bias terms for two fully connected layers. For activation function, This is the intermediate feature vector after the first nonlinear transformation. This is a high-order abstract feature vector obtained through secondary feature extraction; The fully connected layer module employs an Adam optimizer with an adaptive learning rate, dynamically adjusting the learning rate through an exponential decay mechanism. The final output is a high-order, high-dimensional feature vector representing the spatiotemporal coupling characteristics of the traction system's temperature field. .

9. The real-time temperature field prediction system for a traction system based on a hybrid deep neural network according to claim 8, characterized in that: The prediction output module predicts the temperature field of the traction system at future moments based on the high-order, high-dimensional feature vector output by the fully connected layer module, specifically including: The high-order, high-dimensional feature vector output by the fully connected layer module The input / output layer uses linear regression to obtain the predicted temperature value, expressed as follows: ； in, Predicting temperature values ​​for a single time step. The output layer can be learned weights. For learnable bias terms of the output layer; After multi-time-step prediction, the system outputs a sequence of future temperature values ​​for the traction system. This sequence is then compared with the corresponding actual temperature values ​​to calculate the error. The mean square error is used as the error evaluation function, expressed as follows: ； in, To predict the sample size, For the first The true temperature value of each sample For the first The predicted temperature values ​​for each sample, where ∑ represents the summation operation. Mean square error; Based on the calculated mean square error, the model parameters of the entire neural network are iteratively optimized through the backpropagation algorithm, and the prediction error is controlled within a preset range of 0.3℃ to 0.8℃.

10. A method for real-time prediction of temperature field in a traction system based on a hybrid deep neural network, characterized in that, Includes the following steps: S1. Data Acquisition and Preprocessing: The raw temperature data of key components of the traction system are collected synchronously through thermocouple sensors, and the operating condition data of the traction system are collected synchronously. The raw temperature data are resampled and standardized in sequence, and the standardized temperature data sequence and operating condition information dataset are output. S2. Construction of Hybrid Deep Neural Network Model: A dual-branch hybrid deep neural network model is constructed, wherein the dual branches include a spatial branch and a temporal branch, which respectively extract the spatial and temporal features of the temperature field. After feature fusion, the prediction results are output through a fully connected layer. S3. Model Training and Optimization: The mean squared error is used as the loss function. The model is trained by combining the Adam optimizer with exponential decay mechanism. Data augmentation, multi-scale feature interaction, dynamic attention and early stopping mechanism are introduced to optimize the performance of the dual-branch hybrid deep neural network model and control the prediction error within the preset range. S4. Temperature Field Prediction: Input the standardized temperature data sequence into the trained and optimized dual-branch hybrid deep neural network model, and output the temperature data sequence for predicting the future temperature field of the traction system.

Images