Engine temperature prediction method and system, electronic equipment and storage medium

Abstract

The invention provides an engine temperature prediction method and system, electronic equipment and a storage medium, and relates to the technical field of engine safety, and the method comprises the steps: inputting engine temperature data into a pre-trained temperature prediction model, and obtaining predicted temperature data of an engine at the next moment; the training process of the temperature prediction model comprises the following steps: inputting historical temperature data into each hidden layer of an initial model in sequence, and performing coding and decoding reconstruction processing to obtain a reconstruction feature vector corresponding to each hidden layer; calculating a comparison error between the reconstructed feature vector of each hidden layer and the initial feature vector to obtain a comparison error of each hidden layer; and determining a target hidden layer number of the initial model based on the comparison error, and taking the target hidden layer number as a final temperature prediction model. In the model training process, the number of hidden layers is adaptively controlled based on the comparison error, dynamic determination of the optimal layer number is achieved, the model complexity is reduced, over-fitting and under-fitting are avoided, and the accuracy and reliability of engine temperature prediction are improved.

Description

Technical Field

[0001] This application relates to the field of engine safety technology, and in particular to an engine temperature prediction method, system, electronic device and storage medium. Background Technology

[0002] With the rapid development of the new energy vehicle industry, range-extended electric vehicles (REEVs) have gradually become the focus of market attention due to their combination of the high efficiency of electric drive and the convenience of refueling. During the operation of a REEV, the engine, as an auxiliary power source, is directly affected by temperature changes, impacting system safety, fuel efficiency, and overall vehicle reliability. Improper engine temperature control can lead to overheating damage, reduced power, and even safety hazards. Therefore, accurate prediction and early warning of engine temperature are crucial for improving the performance and safety of REEVs.

[0003] However, existing engine temperature prediction methods still have many shortcomings. On the one hand, traditional prediction models based on physical modeling or shallow learning are poorly adapted to complex operating conditions and struggle to handle the nonlinear and high-dimensional complex features in engine operating data. On the other hand, while deep neural networks have improved prediction accuracy, they generally rely on a fixed number of hidden layers, making it difficult to balance model complexity with generalization performance. Too many hidden layers can easily lead to overfitting and prediction distortion, while insufficient hidden layers can cause underfitting and fail to fully extract feature information.

[0004] Therefore, there is an urgent need for a new engine temperature prediction method that can adaptively control model complexity, improve prediction accuracy and generalization ability. Summary of the Invention

[0005] In view of the above problems, embodiments of this application provide a communication fault handling method, system, electronic device, and storage medium to overcome or at least partially solve the above problems.

[0006] A first aspect of this application provides an engine temperature prediction method, the method comprising: The engine's current temperature data is input into a pre-trained temperature prediction model, which includes multiple preset target hidden layers, each of which includes a set of encoders and decoders. The temperature data at the current moment is reconstructed by encoding and decoding each of the encoders and decoders of the multiple target hidden layers to obtain the predicted temperature data of the engine at the next moment. The training process of the temperature prediction model includes: Historical temperature data is sequentially input into each hidden layer of the initial model. The encoder of the hidden layer performs dimensionality reduction, and the decoder of the hidden layer and the dimensionality-upgrading layer of the graph convolutional network perform dimensionality-upgrading to obtain the reconstructed feature vector corresponding to each hidden layer. Calculate the comparison error between the reconstructed feature vector and the initial feature vector for each hidden layer to obtain the comparison error for each hidden layer. Based on the contrast error corresponding to each hidden layer, the target number of hidden layers in the initial model is determined, and the initial model with the determined target number of hidden layers is used as the temperature prediction model.

[0007] Optionally, the encoding, decoding, and reconstruction process of any round of the temperature prediction model includes: The feature vector output by the encoder in the (N-1)th round is input into the encoder in the Nth round for dimensionality reduction, where N is an integer greater than 1; The feature vector after dimensionality reduction of the encoder in the Nth round is input into the decoder in the Nth round to obtain the feature vector after dimensionality increase of the decoder in the Nth round. The feature vector after dimensionality increase of the decoder in the Nth round is input into the dimensionality increase layer of the graph convolutional network for dimensionality increase processing to obtain the reconstructed feature vector in the Nth round. The dimensionality increase layer of the graph convolutional network performs dimensionality increase processing according to the number of rounds corresponding to the number of layers in the hidden layer where the decoder is located. The step of determining the target number of hidden layers in the initial model based on the contrast error corresponding to each hidden layer includes: Based on the reconstructed feature vector of the Nth round and the initial feature vector, the comparison error of the Nth round is obtained; The number of target hidden layers in the initial model is determined based on the comparison error in the Nth round and the comparison error in the (N-1)th round.

[0008] Optionally, after determining the temperature prediction model, the training step of the temperature prediction model further includes: Based on each of the aforementioned comparison errors, the similarity between each pair of adjacent target hidden layers is calculated to obtain the similarity matrix of the temperature prediction model; the difference between the similarity matrix and the preset identity matrix is ​​calculated to obtain the comparison error fusion matrix. The contrast error fusion matrix is ​​used as an adjacency matrix and input into the graph convolutional network along with the reconstructed feature vectors corresponding to each of the target hidden layers to perform feature fusion and generate a fused feature vector. Based on the fused feature vector, the parameters of the temperature prediction model are optimized.

[0009] Optionally, determining the number of the target hidden layer of the initial model based on the comparison error in the Nth round and the comparison error in the (N-1)th round includes: If the comparison error in the Nth round is greater than the comparison error in the (N-1)th round, the number of rounds in the Nth round is determined as the number of the target hidden layer. If the comparison error in the Nth round is less than or equal to the comparison error in the (N-1)th round, the feature vector of the encoder after dimensionality reduction in the Nth round is input into the (N+1)th hidden layer for encoding and decoding reconstruction processing in the (N+1)th round.

[0010] Optionally, the input feature vector is subjected to dimensionality reduction processing by an encoder, including: The encoder maps the input feature vector to a low-dimensional space, and performs initial dimensionality reduction on the feature vector in the low-dimensional space to obtain a low-dimensional space feature vector. The feature vectors in the low-dimensional space are randomly sampled to obtain the feature vectors after dimensionality reduction by the encoder.

[0011] Optionally, the feature vectors after dimensionality reduction by the encoder are subjected to dimensionality increase processing to obtain reconstructed feature vectors, including: The decoder performs initial up-dimensionality processing on the feature vector after dimensionality reduction by the encoder to obtain an initial reconstructed feature vector. The feature structure of the initial reconstructed feature vector is the same as the feature structure of the feature vector after dimensionality reduction by the encoder. Based on the local topological structure information of the initial reconstructed feature vector, the initial reconstructed feature vector is increased in dimensionality through multiple rounds of the dimensionality-increasing layer of the graph convolutional network to obtain the reconstructed feature vector. The number of rounds of dimensionality increase through the dimensionality-increasing layer is the same as the number of layers in the hidden layer where the decoder is located, and the feature structure of the reconstructed feature vector is the same as the feature structure of the initial feature vector.

[0012] Optionally, the step of performing encoding, decoding, and reconstruction processing on the temperature data at the current moment through the encoders and decoders of the multiple target hidden layers to obtain the predicted temperature data of the engine at the next moment includes: The temperature data at the current moment is subjected to feature extraction to obtain the initial feature vector of the temperature data at the current moment; The initial feature vector of the temperature data at the current moment is sequentially input into the encoder of each target hidden layer for dimensionality reduction, and the feature vector after dimensionality reduction by each encoder is input into its corresponding decoder to obtain multiple feature vectors after dimensionality increase by each decoder. The multiple feature vectors after being upscaled by each of the decoders are respectively input into the upscaled layer of the temperature prediction model. In the upscaled layer, the multiple feature vectors after being upscaled by each of the decoders are upscaled according to the layer value of their respective target hidden layers to obtain the reconstructed feature vectors corresponding to each target hidden layer. Based on the contrast error fusion matrix determined during the model training phase, feature fusion is performed on the reconstructed feature vector to obtain the fused feature vector; Based on the fused feature vector, the predicted temperature data of the engine at the next moment is determined.

[0013] A second aspect of this application provides an engine temperature prediction system, the system comprising: The input module is used to input the engine's current temperature data into a pre-trained temperature prediction model. The temperature prediction model includes multiple preset target hidden layers, and each target hidden layer includes a set of encoders and decoders. The encoding / decoding and reconstruction processing module is used to encode and decode the temperature data at the current moment through the encoders and decoders of the multiple target hidden layers to obtain the predicted temperature data of the engine at the next moment. The temperature prediction model includes: The encoding, decoding, and reconstruction processing submodule is used to sequentially input historical temperature data into each hidden layer of the initial model, perform dimensionality reduction processing through the encoder of the hidden layer, and perform dimensionality increase processing through the decoder of the hidden layer and the dimensionality increase layer of the graph convolutional network to obtain the reconstruction feature vector corresponding to each hidden layer. The first calculation submodule is used to calculate the comparison error between the reconstructed feature vector and the initial feature vector of each hidden layer, and obtain the comparison error of each hidden layer. The first determining submodule is used to determine the target number of hidden layers in the initial model based on the contrast error corresponding to each hidden layer, and to use the initial model with the determined target number of hidden layers as the temperature prediction model.

[0014] Optionally, the encoding / decoding reconstruction processing submodule includes: The first dimensionality reduction subunit is used to input the feature vector output by the encoder in the (N-1)th round into the encoder in the Nth round for dimensionality reduction processing, where N is an integer greater than 1; The first dimension-upgrading subunit is used to input the dimension-reduced feature vector of the encoder in the Nth round into the decoder in the Nth round to obtain the dimension-upgrading feature vector of the decoder in the Nth round. The second dimension-upgrading subunit is used to input the dimension-upgrading feature vector of the decoder in the Nth round into the dimension-upgrading layer of the graph convolutional network for dimension-upgrading processing to obtain the reconstructed feature vector in the Nth round. The dimension-upgrading layer of the graph convolutional network performs dimension-upgrading processing according to the number of rounds corresponding to the number of layers in the hidden layer where the decoder is located. The first determining subunit is used to determine the target number of hidden layers in the initial model based on the contrast error corresponding to each hidden layer, including: The second determining subunit is used to obtain the comparison error of the Nth round based on the reconstructed feature vector of the Nth round and the initial feature vector; The third determining subunit is used to determine the number of target hidden layers of the initial model based on the comparison error of the Nth round and the comparison error of the (N-1)th round.

[0015] Optionally, the encoding / decoding reconstruction processing submodule further includes: The first calculation subunit is used to calculate the similarity between two adjacent target hidden layers based on each of the comparison errors, and obtain the similarity matrix of the temperature prediction model; The second calculation subunit is used to perform a difference operation between the similarity matrix and the preset identity matrix to obtain a comparison error fusion matrix; The feature fusion subunit is used to take the contrast error fusion matrix as an adjacency matrix and input it into the graph convolutional network along with the reconstructed feature vectors corresponding to each of the target hidden layers to generate a fused feature vector. The parameter optimization subunit is used to optimize the parameters of the temperature prediction model based on the fused feature vector.

[0016] Optionally, the step of determining the number of target hidden layers of the initial model based on the comparison error in the Nth round and the comparison error in the (N-1)th round, the third determining subunit includes: The fourth determining subunit is used to determine the number of rounds in the Nth round as the number of layers of the target hidden layer when the comparison error in the Nth round is greater than the comparison error in the (N-1)th round. The encoding / decoding reconstruction subunit is used to input the feature vector after dimensionality reduction of the encoder in the Nth round into the N+1th hidden layer for encoding / decoding reconstruction processing when the comparison error in the Nth round is less than or equal to the comparison error in the (N-1)th round.

[0017] Optionally, the input feature vector is reduced in dimensionality by an encoder, and the encoding / decoding reconstruction processing submodule includes: The mapping subunit is used to map the input feature vector to a low-dimensional space through the encoder, and to perform initial dimensionality reduction processing on the feature vector in the low-dimensional space to obtain a low-dimensional space feature vector. The random sampling subunit is used to randomly sample the feature vectors in the low-dimensional space to obtain the feature vectors after dimensionality reduction by the encoder.

[0018] Optionally, the feature vector after dimensionality reduction by the encoder is subjected to dimensionality increase processing to obtain a reconstructed feature vector. The encoding / decoding reconstruction processing submodule includes: The initial dimensionality-up subunit is used to perform initial dimensionality-up processing on the feature vector after dimensionality reduction by the encoder through the decoder to obtain the initial reconstructed feature vector. The feature structure of the initial reconstructed feature vector is the same as the feature structure of the feature vector after dimensionality reduction by the encoder. The multi-round dimensionality-upgrading subunit is used to perform dimensionality-upgrading on the initial reconstructed feature vector through the dimensionality-upgrading layer of the graph convolutional network for multiple rounds based on the local topological structure information of the initial reconstructed feature vector, so as to obtain the reconstructed feature vector. The number of rounds of dimensionality-upgrading through the dimensionality-upgrading layer is the same as the number of layers of the hidden layer where the decoder is located, and the feature structure of the reconstructed feature vector is the same as the feature structure of the initial feature vector.

[0019] Optionally, the step of performing encoding and decoding reconstruction processing on the temperature data at the current moment through the encoders and decoders of the multiple target hidden layers to obtain the predicted temperature data of the engine at the next moment, the encoding and decoding reconstruction processing module includes: The feature extraction submodule is used to extract features from the temperature data at the current moment to obtain the initial feature vector of the temperature data at the current moment. The second determining submodule is used to sequentially input the initial feature vector of the temperature data at the current moment into the encoder of each target hidden layer for dimensionality reduction processing, and input the feature vector after dimensionality reduction processing by each encoder into its corresponding decoder to obtain multiple feature vectors after dimensionality increase processing by each decoder. The third determining submodule is used to input the multiple feature vectors after being upscaled by each of the decoders into the upscaled layer of the temperature prediction model, and to upscale the multiple feature vectors after being upscaled by each of the decoders according to the layer value of their respective target hidden layers in the upscaled layer, so as to obtain the reconstructed feature vectors corresponding to each target hidden layer. The feature fusion submodule is used to perform feature fusion on the reconstructed feature vector based on the contrast error fusion matrix determined during the model training phase, so as to obtain a fused feature vector. The fourth determination submodule is used to determine the predicted temperature data of the engine at the next moment based on the fused feature vector.

[0020] A third aspect of this application provides an electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the engine temperature prediction method as described in the first aspect of this application.

[0021] A fourth aspect of this application provides a readable storage medium storing a program or instructions that, when executed by a processor, implement the steps of the engine temperature prediction method as described in the first aspect of this application.

[0022] The beneficial effects of this application are: This application provides an engine temperature prediction method, the method comprising: inputting the engine's current temperature data into a pre-trained temperature prediction model, the temperature prediction model comprising multiple preset target hidden layers, each target hidden layer comprising a set of encoders and decoders; performing encoding, decoding, and reconstruction processing on the current temperature data through the encoders and decoders of the multiple target hidden layers to obtain the predicted temperature data of the engine at the next moment; the training process of the temperature prediction model comprising: sequentially inputting historical temperature data into each hidden layer of an initial model, performing dimensionality reduction processing through the encoder of the hidden layer, and performing dimensionality increase processing through the decoder of the hidden layer and the dimensionality increase layer of a graph convolutional network to obtain the reconstructed feature vector corresponding to each hidden layer; calculating the comparison error between the reconstructed feature vector corresponding to each hidden layer and the initial feature vector respectively to obtain the comparison error corresponding to each hidden layer; determining the number of target hidden layers of the initial model based on the comparison error corresponding to each hidden layer, and using the initial model with the determined number of target hidden layers as the temperature prediction model.

[0023] The engine temperature prediction method proposed in this application can dynamically determine the optimal number of hidden layers by adaptively controlling the number of hidden layers based on the contrast error during model training, thereby effectively reducing model complexity while avoiding overfitting and underfitting. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic flowchart of the steps of an engine temperature prediction method provided in an embodiment of this application; Figure 2 This is a schematic diagram of encoding and decoding feature reconstruction with adaptive control of the number of hidden layers provided in an embodiment of this application; Figure 3 This is a flowchart illustrating an adaptive control method for a hidden layer provided in an embodiment of this application; Figure 4 This is a schematic diagram illustrating the training process of a temperature prediction model provided in an embodiment of this application; Figure 5 This is a schematic diagram of an engine temperature prediction system provided in an embodiment of this application; Figure 6 This is a schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0026] Exemplary embodiments of this application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art.

[0027] In a first aspect, this application provides a method for predicting engine temperature, such as... Figure 1 As shown, the method includes: Step S101: Input the engine temperature data at the current moment into a pre-trained temperature prediction model. The temperature prediction model includes multiple preset target hidden layers, and each target hidden layer includes a set of encoders and decoders.

[0028] In this step, the engine's current temperature data is input into a pre-trained temperature prediction model. The temperature prediction model contains multiple preset target hidden layers, each comprising a set of cooperating encoders and decoders. The encoder performs dimensionality reduction on the input features to extract low-dimensional latent features and reduce redundant information; the decoder performs dimensionality upscaling on the reduced features, restoring them to the same structure and dimension as the original features of the target hidden layer corresponding to the input, thus ensuring information integrity.

[0029] Step S102: The temperature data at the current moment is reconstructed by encoding and decoding the encoders and decoders of the multiple target hidden layers to obtain the predicted temperature data of the engine at the next moment.

[0030] In this step, the engine temperature data at the current moment is processed through a layer-by-layer encoding and decoding reconstruction using encoders and decoders with multiple target hidden layers. During this process, the current temperature data undergoes dimensionality reduction by each encoder layer and dimensionality increase by the decoder, gradually forming multiple reconstructed feature vectors at different levels. Based on these feature vectors, the predicted engine temperature data for the next moment is generated. By extracting and reconstructing features through multiple hidden layers, the nonlinear relationship between engine temperature and time can be captured more accurately, thereby improving the reliability and accuracy of the prediction results.

[0031] The training process of the temperature prediction model includes: Historical temperature data is sequentially input into each hidden layer of the initial model. The encoder of the hidden layer performs dimensionality reduction, and the decoder of the hidden layer and the dimensionality-upgrading layer of the graph convolutional network perform dimensionality-upgrading to obtain the reconstructed feature vector corresponding to each hidden layer. Calculate the comparison error between the reconstructed feature vector and the initial feature vector for each hidden layer to obtain the comparison error for each hidden layer. Based on the contrast error corresponding to each hidden layer, the target number of hidden layers in the initial model is determined, and the initial model with the determined target number of hidden layers is used as the temperature prediction model.

[0032] In this application, the training process of the temperature prediction model specifically includes: First, historical temperature data is sequentially input into each hidden layer of the initial model. The encoder of each hidden layer performs dimensionality reduction on the input features. Then, the dimensionality reduction features are increased through the corresponding decoder and the dimensionality increase layer of the graph convolutional network to obtain the reconstructed feature vectors corresponding to each hidden layer. Next, the comparison error between the reconstructed feature vectors of each hidden layer and the input initial feature vectors is calculated, and this is used as an indicator to measure the modeling ability of different hidden layers. Finally, based on the comparison error values ​​of each hidden layer, the optimal number of hidden layers is dynamically determined to obtain the target number of hidden layers. The determined target number of hidden layers is then applied to the initial model to obtain the final temperature prediction model.

[0033] The engine temperature prediction method proposed in this application can dynamically determine the optimal number of hidden layers by adaptively controlling the number of hidden layers based on the contrast error during model training, thereby effectively reducing model complexity while avoiding overfitting and underfitting.

[0034] In one embodiment, the encoding, decoding, and reconstruction process of the temperature prediction model in any round includes: The feature vector output by the encoder in the (N-1)th round is input into the encoder in the Nth round for dimensionality reduction, where N is an integer greater than 1; The feature vector after dimensionality reduction of the encoder in the Nth round is input into the decoder in the Nth round to obtain the feature vector after dimensionality increase of the decoder in the Nth round. The feature vector after dimensionality increase of the decoder in the Nth round is input into the dimensionality increase layer of the graph convolutional network for dimensionality increase processing to obtain the reconstructed feature vector in the Nth round. The dimensionality increase layer of the graph convolutional network performs dimensionality increase processing according to the number of rounds corresponding to the number of layers in the hidden layer where the decoder is located. The step of determining the target number of hidden layers in the initial model based on the contrast error corresponding to each hidden layer includes: Based on the reconstructed feature vector of the Nth round and the initial feature vector, the comparison error of the Nth round is obtained; The number of target hidden layers in the initial model is determined based on the comparison error in the Nth round and the comparison error in the (N-1)th round.

[0035] In this embodiment, firstly, the feature vector output by the encoder in the (N-1)th round is input into the encoder in the Nth round to perform dimensionality reduction on the feature vector, where N is an integer greater than 1. This dimensionality reduction operation removes redundant data while preserving the main feature information, thereby improving the efficiency and effectiveness of subsequent processing.

[0036] Then, the feature vector after dimensionality reduction by the encoder in the Nth round is input into the decoder in the Nth round. The decoder in the Nth round performs dimensionality upscaling on the dimensionality-reduced features, resulting in the dimensionality-upscaling feature vector output by the decoder in the Nth round. This dimensionality upscaling operation can recover the compressed information during dimensionality reduction to a certain extent, making the features structurally and dimensionally close to the original features when input to the hidden layer in the Nth round, thus maintaining the integrity required for the prediction task.

[0037] Next, the dimensionality-upgraded feature vector from the Nth round of the decoder is further input into the dimensionality-upgrading layer of the graph convolutional network for processing to obtain the reconstructed feature vector from the Nth round. The dimensionality-upgrading layer of the graph convolutional network performs the dimensionality-upgrading operation according to the number of rounds corresponding to the number of the hidden layers containing the decoder, thus ensuring that the dimensionality-upgrading process remains consistent with the propagation depth of the hidden layers. Through this process, not only is the feature dimension recovered in the numerical space, but the topological information of the graph structure is also used to enhance the expressive power of the features.

[0038] Furthermore, during training, the target hidden layer number of the initial model is determined based on the contrast error corresponding to each hidden layer. Specifically, firstly, the contrast error of the Nth round is obtained based on the difference between the reconstructed feature vector and the initial feature vector; then, the contrast error of the Nth round is compared with the contrast error of the (N-1)th round. When the comparison result of the contrast errors meets a preset condition, the target hidden layer number of the initial model can be determined. This embodiment, by comparing the errors round by round, can dynamically and accurately determine the optimal hidden layer depth of the model, thereby ensuring prediction accuracy while avoiding overfitting or underfitting problems during model training.

[0039] For example, this application provides as follows Figure 2 The diagram shown illustrates the encoding / decoding feature reconstruction using adaptive control of the number of hidden layers, as follows: Figure 2 As shown: First, temperature data of the engine under different operating conditions is collected, and this temperature data is used as input feature vectors and fed into the initial model. The temperature features at this stage are denoted as the initial feature vectors. and the corresponding adjacent moments Together they constitute the input graph structure, i.e., the initial feature map structure. .

[0040] Next, the encoding, decoding, and reconstruction process proceeds through multiple hidden layers. Each hidden layer contains an encoder, a decoder, and an error calculation unit, and its processing is as follows: The initial feature vector is used as the input to the encoder of the first hidden layer. The encoder's MLP (Multilayer Perceptron) performs dimensionality reduction on the initial feature vector, mapping the high-dimensional features to a low-dimensional latent space. Then, the dimensionality-reduced feature vector is input to the decoder of the first hidden layer. The MLP of the first hidden layer performs initial dimensionality upscaling to obtain the initial reconstructed features. Subsequently, these features are further input to the upscaling layer of a graph convolutional network. Based on the adjacency matrix, the features are propagated and topologically enhanced, ensuring that the features not only recover the same dimension as the input features in the numerical space but also fully utilize the relational information of the graph structure to obtain a reconstructed feature structure that is closer to the initial feature vector. Then the graph structure and By comparison, the comparison error b1 is obtained.

[0041] As the hidden layers propagate layer by layer, a series of contrast error values ​​will be obtained sequentially. During the training phase, the model determines whether to continue adding hidden layers by comparing the contrast error between adjacent layers: if the contrast error of the current layer is reduced compared to the previous layer, it continues to propagate to deeper layers; if the contrast error of the current layer is greater than or equal to the error of the previous layer, it indicates that continuing to add layers cannot effectively improve the reconstruction quality, at which point propagation is terminated, and the number of rounds corresponding to that layer is determined as the final target number of hidden layers.

[0042] In one embodiment, after determining the temperature prediction model, the training step of the temperature prediction model further includes: Based on each of the aforementioned comparison errors, the similarity between each pair of adjacent target hidden layers is calculated to obtain the similarity matrix of the temperature prediction model; the difference between the similarity matrix and the preset identity matrix is ​​calculated to obtain the comparison error fusion matrix. The contrast error fusion matrix is ​​used as an adjacency matrix and input into the graph convolutional network along with the reconstructed feature vectors corresponding to each of the target hidden layers to perform feature fusion and generate a fused feature vector. Based on the fused feature vector, the parameters of the temperature prediction model are optimized.

[0043] In this embodiment, based on the contrast error corresponding to each target hidden layer, the similarity between each pair of adjacent target hidden layers is calculated to obtain the similarity matrix of the temperature prediction model. This similarity matrix can comprehensively characterize the differences and correlations in the reconstructed features of different target hidden layers, enabling the model to intuitively capture the connections between multi-level features.

[0044] Subsequently, the similarity matrix is ​​subtracted from the preset identity matrix to obtain the comparison error fusion matrix. This fusion matrix, by introducing the subtraction operation, can reduce the correlation between redundant information while preserving the original error information, avoiding the subjectivity and limitations of manually setting weights, thus achieving a more objective basis for feature fusion.

[0045] Next, the contrast error fusion matrix is ​​used as the adjacency matrix, and combined with the reconstructed feature vectors of each target hidden layer, it is input into the graph convolutional network for feature fusion processing. The graph convolutional network can utilize the inter-layer similarity described by the fusion matrix to mine the complementarity and differences of features in different target hidden layers, thereby generating a more representative fused feature vector.

[0046] Finally, based on the generated fusion feature vector, the parameters of the temperature prediction model are further optimized.

[0047] This embodiment avoids the manual setting of weight coefficients in traditional feature fusion by constructing a similarity matrix and introducing difference calculation, which significantly improves the adaptability and objectivity of the fusion process. In addition, by using the contrast error fusion matrix as the adjacency matrix input graph convolutional network, topological constraints can be introduced in the feature fusion stage, so that features at different levels maintain a high degree of consistency in both numerical and structural spaces, thereby improving the robustness of feature representation.

[0048] In one embodiment, determining the number of the target hidden layer of the initial model based on the comparison error in the Nth round and the comparison error in the (N-1)th round includes: If the comparison error in the Nth round is greater than the comparison error in the (N-1)th round, the number of rounds in the Nth round is determined as the number of the target hidden layer. If the comparison error in the Nth round is less than or equal to the comparison error in the (N-1)th round, the feature vector of the encoder after dimensionality reduction in the Nth round is input into the (N+1)th hidden layer for encoding and decoding reconstruction processing in the (N+1)th round.

[0049] In this embodiment, it is determined whether the comparison error in the Nth round is greater than the comparison error in the (N-1)th round. The comparison error is calculated as the difference between the reconstructed feature vector in the current round and the initial feature vector, and is used to characterize the reconstruction capability of the current hidden layer. If the comparison error of a new layer increases as the number of hidden layers in the model gradually increases, it indicates that continuing to add hidden layers cannot effectively improve the feature representation capability, and may even lead to a decrease in model performance.

[0050] If the comparison error in the Nth round is greater than the comparison error in the (N-1)th round, the round number corresponding to the Nth round is determined as the layer number of the target hidden layer. That is, the Nth layer is taken as the final effective modeling depth, and a temperature prediction model is constructed accordingly. This can avoid overfitting or redundant calculation problems that may occur when the model is too deep.

[0051] If the comparison error in the Nth round is less than or equal to the comparison error in the (N-1)th round, it indicates that adding a hidden layer can further improve the reconstruction effect. Therefore, the dimensionality-reduced feature vector output by the encoder in the Nth round is input into the (N+1)th hidden layer, and the encoding and decoding reconstruction process continues in the (N+1)th round. In this embodiment, the above process is executed cyclically until the comparison error no longer decreases, ultimately obtaining the optimal number of target hidden layers.

[0052] This embodiment can dynamically determine the model depth by comparing the error change trends of two adjacent rounds, without needing to pre-fix the number of layers during the model design stage. This avoids the problems of overfitting and wasted computational resources caused by too many hidden layers.

[0053] For example, refer to Figure 3 The flowchart of the hidden layer adaptive control shown is as follows: Figure 3 As shown: First, the output features of the first hidden layer are input to the encoder of the first hidden layer for dimensionality reduction, resulting in a low-dimensional feature representation. Then, the decoder of the first hidden layer performs initial dimensionality upscaling reconstruction on the low-dimensional features, yielding the reconstructed feature vector of the first hidden layer. and with the initial feature vector The calculation difference between them is used to calculate the contrast error value corresponding to the first hidden layer. The comparison error can be calculated using the norm of the vector difference, for example, the formula: (1) in, Represents the initial feature vector. This represents the feature vector obtained from the reconstruction of the i-th hidden layer.

[0054] Furthermore, the calculated comparison error value Error value compared with the previous round Perform a comparison. If the judgment result is... > If the error continues to increase in this hidden layer, it indicates that the model's reconstruction performance deteriorates after this layer. In this case, the propagation process terminates, and the current layer number is taken as the target hidden layer number. If the judgment result is... ≤ If the model reconstruction effect is maintained or optimized, then the output features of the first hidden layer are passed to the second hidden layer for the next round of processing.

[0055] In the second hidden layer, the feature vector obtained by the encoder in the first hidden layer is also reconstructed by the encoder in the second hidden layer and the decoder in the second hidden layer, resulting in the reconstructed feature vector of the second hidden layer. Subsequently, the reconstructed feature is compared with the initial feature vector. The difference was calculated to obtain the comparison error value for the second round. Then and Compare: If > If the error starts to increase, propagation is terminated and the number of hidden layers is determined; if ≤ If the result is positive, it continues to be passed to the third hidden layer, and so on up to the i-th hidden layer.

[0056] In one embodiment, the input feature vector is dimensionality reduced by an encoder, including: The encoder maps the input feature vector to a low-dimensional space, and performs initial dimensionality reduction on the feature vector in the low-dimensional space to obtain a low-dimensional space feature vector. The feature vectors in the low-dimensional space are randomly sampled to obtain the feature vectors after dimensionality reduction by the encoder.

[0057] In this embodiment, the input feature vector is dimensionality reduced by an encoder in any hidden layer, specifically including: The input feature vector, which has a relatively high feature dimension, is fed into the encoder of the current hidden layer. The encoder maps this feature vector to a low-dimensional space and performs initial dimensionality reduction processing within this low-dimensional space to obtain a low-dimensional feature vector. In this embodiment, the initial dimensionality reduction process can be performed using a multilayer perceptron located within the encoder. By mapping the feature vector to a low-dimensional space, the feature dimension can be compressed, and some redundant information can be removed.

[0058] Subsequently, the feature vectors after the initial dimensionality reduction are randomly sampled in the low-dimensional space to obtain the final feature vectors after dimensionality reduction by the encoder. Random sampling enhances the diversity of feature representations while preserving key information, avoids the model getting trapped in local optima, and improves the model's ability to capture complex nonlinear features.

[0059] In one embodiment, the feature vector after dimensionality reduction by the encoder is subjected to dimensionality upscaling to obtain a reconstructed feature vector, including: The decoder performs initial up-dimensionality processing on the feature vector after dimensionality reduction by the encoder to obtain an initial reconstructed feature vector. The feature structure of the initial reconstructed feature vector is the same as the feature structure of the feature vector after dimensionality reduction by the encoder. Based on the local topological structure information of the initial reconstructed feature vector, the initial reconstructed feature vector is increased in dimensionality through multiple rounds of the dimensionality-increasing layer of the graph convolutional network to obtain the reconstructed feature vector. The number of rounds of dimensionality increase through the dimensionality-increasing layer is the same as the number of layers in the hidden layer where the decoder is located, and the feature structure of the reconstructed feature vector is the same as the feature structure of the initial feature vector.

[0060] In this embodiment, the dimensionality-reduced feature vector with low-dimensional features output by the encoder is input into the decoder. The decoder performs initial dimensionality upscaling on the vector using its internal multilayer perceptron to recover some of the feature information compressed during dimensionality reduction, thereby obtaining the initial reconstructed feature vector. It should be noted that the initial reconstructed feature vector maintains the same feature structure as the dimensionality-reduced feature vector, i.e., the dimensionality relationship between the two is a direct mapping from low-dimensional to high-dimensional. However, the initial reconstructed feature vector has not yet been completely restored to the same structure as the original input feature vector.

[0061] Subsequently, based on the local topological information of the initial reconstructed feature vector, the initial reconstructed feature vector is subjected to multiple rounds of dimensionality increase processing through the dimensionality increase layer of the graph convolutional network. In this process, the dimensionality increase layer of the graph convolutional network can utilize the local adjacency relationships and interdependencies between node features contained in the initial reconstructed feature vector to iteratively update and expand the features.

[0062] Specifically, each dimensionality-upgrading layer expands features based on the output of the previous round, thereby gradually recovering the compressed high-dimensional feature representation. Through multiple rounds of dimensionality-upgrading iterations, the complete representation of the input features under the global topological structure can be gradually reconstructed. It should be noted that the number of rounds of dimensionality-upgrading through the dimensionality-upgrading layer is the same as the number of hidden layers in the decoder, that is, each hidden layer corresponds to a fixed number of dimensionality-upgrading rounds to ensure the hierarchy and consistency of feature reconstruction.

[0063] Ultimately, after multiple rounds of dimensionality upscaling, the reconstructed feature vector maintains the same feature structure as the initial input feature vector, meaning it is restored to the same high-dimensional feature representation as the original features, thus completing the feature reconstruction process. It's important to note that the feature structure of the reconstructed feature vector is identical to that of the initial feature vector, indicating the same dimension and / or semantics. Specifically, at the dimensionality level: the initial feature vector generally originates from the engine temperature data after feature extraction, such as a d-dimensional vector; after undergoing the processes of "encoder dimensionality reduction, decoder dimensionality upscaling, and graph convolutional network dimensionality upscaling," the resulting reconstructed feature vector still returns to d-dimensional space. In other words, the dimensions of the input and output features are consistent. At the semantic level: although dimensionality reduction, dimensionality upscaling, and feature reconstruction are performed in the intermediate processes, the information content expressed by the reconstructed feature vector corresponds to that of the initial feature vector; that is, their feature components (e.g., temperature change trends, correlation patterns) maintain semantic consistency rather than being transformed into another feature type. In this embodiment, by first performing initial dimensionality increase and then combining it with the layer-by-layer dimensionality increase mechanism of graph convolutional networks, the compressed feature information can be better recovered, avoiding the information loss caused by simple linear dimensionality increase. In addition, by corresponding the number of dimensionality increase rounds with the number of hidden layers, the reconstruction depth and complexity of features at different levels are matched, avoiding the problems of over-reconstruction or under-reconstruction.

[0064] In one embodiment, the step of encoding, decoding, and reconstructing the temperature data at the current moment using the encoders and decoders of the plurality of target hidden layers to obtain the predicted temperature data of the engine at the next moment includes: The temperature data at the current moment is subjected to feature extraction to obtain the initial feature vector of the temperature data at the current moment; The initial feature vector of the temperature data at the current moment is sequentially input into the encoder of each target hidden layer for dimensionality reduction, and the feature vector after dimensionality reduction by each encoder is input into its corresponding decoder to obtain multiple feature vectors after dimensionality increase by each decoder. The multiple feature vectors after being upscaled by each of the decoders are respectively input into the upscaled layer of the temperature prediction model. In the upscaled layer, the multiple feature vectors after being upscaled by each of the decoders are upscaled according to the layer value of their respective target hidden layers to obtain the reconstructed feature vectors corresponding to each target hidden layer. Based on the contrast error fusion matrix determined during the model training phase, feature fusion is performed on the reconstructed feature vector to obtain the fused feature vector; Based on the fused feature vector, the predicted temperature data of the engine at the next moment is determined.

[0065] In this embodiment, the engine temperature data at the current moment is encoded, decoded, and reconstructed using encoders and decoders in multiple target hidden layers to obtain the predicted engine temperature data for the next moment. Specifically: First, features are extracted from the engine temperature data at the current moment to obtain an initial feature vector for the temperature data at that moment. This initial feature vector can comprehensively characterize the temperature state under the current operating conditions and its potential changing trends, providing a basic input for subsequent deep feature processing.

[0066] Subsequently, the initial feature vector is sequentially input into encoders of multiple target hidden layers, where it undergoes dimensionality reduction. Each encoder compresses the input features into a low-dimensional space, thereby removing redundant features and highlighting the core information related to temperature evolution. Then, the dimensionality-reduced feature vectors from each encoder are input into their respective decoders for initial dimensionality upscaling, generating multiple dimensionality-upsized feature vectors. These feature vectors gradually reconstruct a high-dimensional space structurally, preserving some of the semantic information of the original input features.

[0067] Next, the feature vectors, after being upscaled by the decoder, are input into the upscaling layer of the temperature prediction model. Within this upscaling layer, multiple rounds of upscaling operations, matching the layer number of the corresponding target hidden layer, are performed on each feature vector to obtain the reconstructed feature vectors for each target hidden layer. This process ensures that the feature reconstruction depth and complexity are consistent across different levels, making the reconstruction results closer to the original input features in both numerical value and topological structure.

[0068] Then, based on the contrast error fusion matrix determined during the model training phase, feature fusion is performed on the reconstructed feature vectors output by each hidden layer. This contrast error fusion matrix reflects the similarity and differences between different layers, and can reasonably allocate the contribution of each hidden layer's reconstructed features during the fusion process, thereby generating a fused feature vector containing global information.

[0069] Finally, based on the generated fused feature vector, the engine temperature data for the next moment is predicted. This prediction result can accurately depict the short-term evolution trend of engine temperature, providing a basis for subsequent temperature warning and control.

[0070] For example, refer to Figure 4 The diagram shown illustrates the training process of the temperature prediction model. Figure 4 As shown: Collect historical temperature data of the engine and construct it as an initial feature vector. The initial feature vector The input is processed layer by layer in multiple hidden layers of the temperature prediction model. Taking the first hidden layer as an example, the input feature vector is reduced in dimensionality by the encoder to obtain a low-dimensional representation, and then reconstructed in dimensionality by the decoder to obtain the node feature vector output by the first hidden layer (the reconstructed feature vector). Subsequently, the feature vector of this node is... With the initial feature vector The difference was calculated to obtain the comparison error value for the first round. The calculation formula is as shown in Formula 1 above.

[0071] Similarly, the input to the second hidden layer is the output feature of the first hidden layer. After the same encoding, decoding and reconstruction process, the node feature vector of the second hidden layer (the reconstructed feature vector) is obtained. Then, the feature vector of this node With the initial feature vector The difference was calculated to obtain the comparison error value for the first round. This process continues until the (n-1)th hidden layer, at which point the corresponding contrast error value is obtained. .

[0072] In the above process, the node feature vectors (reconstructed feature vectors) of all hidden layers and their corresponding contrast error values ​​are saved. Finally, the set of node features output from n-1 hidden layers is input into the graph convolutional network, and feature fusion is performed using the "contrast error fusion matrix" as the adjacency matrix.

[0073] The comparison error fusion matrix is ​​generated as follows: First, based on the contrast error between adjacent hidden layers, pairwise similarity is calculated to form a similarity matrix; then, the difference between this similarity matrix and the n×n identity matrix is ​​performed to obtain the contrast error fusion matrix. Furthermore, the contrast error fusion matrix and the node feature sets output from each hidden layer are input into a graph convolutional network. After convolution and aggregation operations, a fused feature vector is obtained. This fused feature vector integrates multi-layer reconstruction information and cross-layer error information to a certain extent, and can better reflect the global and local relationships of engine temperature characteristics.

[0074] Finally, based on the fused feature vectors, the parameters of the entire temperature prediction model are optimized and updated. By continuously iterating through the above training process until the comparison error converges or reaches a preset threshold, the training of the temperature prediction model is complete.

[0075] In one embodiment, after the engine temperature prediction is completed, a temperature warning operation can be performed based on the prediction result. The process includes the following steps: First, the predicted engine temperature at the next moment is obtained from the temperature prediction model. This predicted temperature data reflects the engine's temperature change trend in the near future, and compared to relying solely on current temperature data collected by sensors, it can anticipate potential overheating risks in the engine.

[0076] Subsequently, the predicted temperature data is compared with a preset temperature threshold. In this application, the temperature threshold can be set according to the engine's design limits, the cooling system's load-bearing capacity, and vehicle operation safety requirements, such as the maximum allowable temperature of the engine coolant and the safe operating temperature of the cylinder block material. In some embodiments, the temperature threshold can be set as a multi-level threshold, including a warning threshold and a danger threshold, to distinguish different degrees of temperature anomalies.

[0077] If the predicted temperature exceeds a preset temperature threshold, a temperature warning message is generated and displayed via the in-vehicle display interface, audible and visual alarm devices, or vehicle-to-everything (V2X) remote communication module, so that the driver or remote maintenance system can take timely measures. In some embodiments, when the predicted temperature exceeds a danger threshold, corresponding active safety strategies can also be triggered, such as reducing engine power output, activating forced cooling mode, or restricting vehicle operating status, thereby preventing the temperature from rising further.

[0078] If the predicted temperature does not exceed the preset temperature threshold, the engine temperature is determined to be within the normal range, and no warning is triggered; the vehicle continues to operate under the current conditions.

[0079] In this application, the training process of the temperature prediction model specifically includes: First, historical temperature data is sequentially input into each hidden layer of the initial model. The encoder of each hidden layer performs dimensionality reduction on the input features. Then, the dimensionality reduction features are increased through the corresponding decoder and the dimensionality increase layer of the graph convolutional network to obtain the reconstructed feature vectors corresponding to each hidden layer. Next, the comparison error between the reconstructed feature vector of each hidden layer and the input initial feature vector is calculated, and this is used as an indicator to measure the modeling ability of different hidden layers. Finally, based on the comparison error value of each hidden layer, the optimal number of hidden layers is dynamically determined to obtain the target number of hidden layers. The determined target number of hidden layers is then applied to the initial model to obtain the final temperature prediction model. The engine temperature prediction method proposed in this application can dynamically determine the optimal number of hidden layers by adaptively controlling the number of hidden layers based on the comparison error during model training, thereby effectively reducing model complexity while avoiding overfitting and underfitting. Based on the same inventive concept, a second aspect of the embodiments of this application provides an engine temperature prediction system, such as... Figure 5 As shown, the system includes: Input module 201 is used to input the engine's current temperature data into a pre-trained temperature prediction model. The temperature prediction model includes multiple preset target hidden layers, and each target hidden layer includes a set of encoders and decoders. The encoding / decoding and reconstruction processing module 202 is used to perform encoding / decoding and reconstruction processing on the temperature data at the current moment through the encoders and decoders of the multiple target hidden layers, so as to obtain the predicted temperature data of the engine at the next moment. The temperature prediction model includes: The encoding, decoding, and reconstruction processing submodule is used to sequentially input historical temperature data into each hidden layer of the initial model, perform dimensionality reduction processing through the encoder of the hidden layer, and perform dimensionality increase processing through the decoder of the hidden layer and the dimensionality increase layer of the graph convolutional network to obtain the reconstruction feature vector corresponding to each hidden layer. The first calculation submodule is used to calculate the comparison error between the reconstructed feature vector and the initial feature vector of each hidden layer, and obtain the comparison error of each hidden layer. The first determining submodule is used to determine the target number of hidden layers in the initial model based on the contrast error corresponding to each hidden layer, and to use the initial model with the determined target number of hidden layers as the temperature prediction model.

[0080] Optionally, the encoding / decoding reconstruction processing submodule includes: The first dimensionality reduction subunit is used to input the feature vector output by the encoder in the (N-1)th round into the encoder in the Nth round for dimensionality reduction processing, where N is an integer greater than 1; The first dimension-upgrading subunit is used to input the dimension-reduced feature vector of the encoder in the Nth round into the decoder in the Nth round to obtain the dimension-upgrading feature vector of the decoder in the Nth round. The second dimension-upgrading subunit is used to input the dimension-upgrading feature vector of the decoder in the Nth round into the dimension-upgrading layer of the graph convolutional network for dimension-upgrading processing to obtain the reconstructed feature vector in the Nth round. The dimension-upgrading layer of the graph convolutional network performs dimension-upgrading processing according to the number of rounds corresponding to the number of layers in the hidden layer where the decoder is located. The first determining subunit is used to determine the target number of hidden layers in the initial model based on the contrast error corresponding to each hidden layer, including: The second determining subunit is used to obtain the comparison error of the Nth round based on the reconstructed feature vector of the Nth round and the initial feature vector; The third determining subunit is used to determine the number of target hidden layers of the initial model based on the comparison error of the Nth round and the comparison error of the (N-1)th round.

[0081] Optionally, the encoding / decoding reconstruction processing submodule further includes: The first calculation subunit is used to calculate the similarity between two adjacent target hidden layers based on each of the comparison errors, and obtain the similarity matrix of the temperature prediction model; The second calculation subunit is used to perform a difference operation between the similarity matrix and the preset identity matrix to obtain a comparison error fusion matrix; The feature fusion subunit is used to take the contrast error fusion matrix as an adjacency matrix and input it into the graph convolutional network along with the reconstructed feature vectors corresponding to each of the target hidden layers to generate a fused feature vector. The parameter optimization subunit is used to optimize the parameters of the temperature prediction model based on the fused feature vector.

[0082] Optionally, the step of determining the number of target hidden layers of the initial model based on the comparison error in the Nth round and the comparison error in the (N-1)th round, the third determining subunit includes: The fourth determining subunit is used to determine the number of rounds in the Nth round as the number of layers of the target hidden layer when the comparison error in the Nth round is greater than the comparison error in the (N-1)th round. The encoding / decoding reconstruction subunit is used to input the feature vector after dimensionality reduction of the encoder in the Nth round into the N+1th hidden layer for encoding / decoding reconstruction processing when the comparison error in the Nth round is less than or equal to the comparison error in the (N-1)th round.

[0083] Optionally, the input feature vector is reduced in dimensionality by an encoder, and the encoding / decoding reconstruction processing submodule includes: The mapping subunit is used to map the input feature vector to a low-dimensional space through the encoder, and to perform initial dimensionality reduction processing on the feature vector in the low-dimensional space to obtain a low-dimensional space feature vector. The random sampling subunit is used to randomly sample the feature vectors in the low-dimensional space to obtain the feature vectors after dimensionality reduction by the encoder.

[0084] Optionally, the feature vector after dimensionality reduction by the encoder is subjected to dimensionality increase processing to obtain a reconstructed feature vector. The encoding / decoding reconstruction processing submodule includes: The initial dimensionality-up subunit is used to perform initial dimensionality-up processing on the feature vector after dimensionality reduction by the encoder through the decoder to obtain the initial reconstructed feature vector. The feature structure of the initial reconstructed feature vector is the same as the feature structure of the feature vector after dimensionality reduction by the encoder. The multi-round dimensionality-upgrading subunit is used to perform dimensionality-upgrading on the initial reconstructed feature vector through the dimensionality-upgrading layer of the graph convolutional network for multiple rounds based on the local topological structure information of the initial reconstructed feature vector, so as to obtain the reconstructed feature vector. The number of rounds of dimensionality-upgrading through the dimensionality-upgrading layer is the same as the number of layers of the hidden layer where the decoder is located, and the feature structure of the reconstructed feature vector is the same as the feature structure of the initial feature vector.

[0085] Optionally, the encoding / decoding reconstruction processing module 202, which performs encoding / decoding reconstruction processing on the temperature data at the current moment through the encoders and decoders of the multiple target hidden layers to obtain the predicted temperature data of the engine at the next moment, includes: The feature extraction submodule is used to extract features from the temperature data at the current moment to obtain the initial feature vector of the temperature data at the current moment. The second determining submodule is used to sequentially input the initial feature vector of the temperature data at the current moment into the encoder of each target hidden layer for dimensionality reduction processing, and input the feature vector after dimensionality reduction processing by each encoder into its corresponding decoder to obtain multiple feature vectors after dimensionality increase processing by each decoder. The third determining submodule is used to input the multiple feature vectors after being upscaled by each of the decoders into the upscaled layer of the temperature prediction model, and to upscale the multiple feature vectors after being upscaled by each of the decoders according to the layer value of their respective target hidden layers in the upscaled layer, so as to obtain the reconstructed feature vectors corresponding to each target hidden layer. The feature fusion submodule is used to perform feature fusion on the reconstructed feature vector based on the contrast error fusion matrix determined during the model training phase, so as to obtain a fused feature vector. The fourth determination submodule is used to determine the predicted temperature data of the engine at the next moment based on the fused feature vector.

[0086] Based on the same inventive concept, a third aspect of the embodiments of this application provides a method as follows: Figure 6 The illustrated electronic device 100 includes a processor 120, a memory 110, and a program or instructions stored in the memory 110 and executable on the processor 120, wherein the program or instructions, when executed by the processor 120, implement the steps of the engine temperature prediction method as described in the first aspect of this application.

[0087] Based on the same inventive concept, in a fourth aspect of this application, a readable storage medium is provided, on which a program or instructions are stored, which, when executed by a processor, implement the steps of the engine temperature prediction method as described in the first aspect of this application.

[0088] Each embodiment in this specification focuses on the differences from other embodiments. For the same or similar parts between the embodiments, please refer to each other.

[0089] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, embodiments of this application can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of this application can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0090] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0091] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0092] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0093] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0094] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0095] The above provides a detailed description of the engine temperature prediction method, system, electronic device, and storage medium. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting engine temperature, characterized in that, The method includes: The engine's current temperature data is input into a pre-trained temperature prediction model, which includes multiple preset target hidden layers, each of which includes a set of encoders and decoders. The temperature data at the current moment is reconstructed by encoding and decoding each of the encoders and decoders of the multiple target hidden layers to obtain the predicted temperature data of the engine at the next moment. The training process of the temperature prediction model includes: Historical temperature data is sequentially input into each hidden layer of the initial model. The encoder of the hidden layer performs dimensionality reduction, and the decoder of the hidden layer and the dimensionality-upgrading layer of the graph convolutional network perform dimensionality-upgrading to obtain the reconstructed feature vector corresponding to each hidden layer. Calculate the comparison error between the reconstructed feature vector and the initial feature vector for each hidden layer to obtain the comparison error for each hidden layer. Based on the contrast error corresponding to each hidden layer, the target number of hidden layers in the initial model is determined, and the initial model with the determined target number of hidden layers is used as the temperature prediction model.

2. The engine temperature prediction method according to claim 1, characterized in that, The encoding, decoding, and reconstruction process of the temperature prediction model in any round includes: The feature vector output by the encoder in the (N-1)th round is input into the encoder in the Nth round for dimensionality reduction, where N is an integer greater than 1; The feature vector after dimensionality reduction of the encoder in the Nth round is input into the decoder in the Nth round to obtain the feature vector after dimensionality increase of the decoder in the Nth round. The feature vector after dimensionality increase of the decoder in the Nth round is input into the dimensionality increase layer of the graph convolutional network for dimensionality increase processing to obtain the reconstructed feature vector in the Nth round. The dimensionality increase layer of the graph convolutional network performs dimensionality increase processing according to the number of rounds corresponding to the number of layers in the hidden layer where the decoder is located. The step of determining the target number of hidden layers in the initial model based on the contrast error corresponding to each hidden layer includes: Based on the reconstructed feature vector of the Nth round and the initial feature vector, the comparison error of the Nth round is obtained; The number of target hidden layers in the initial model is determined based on the comparison error in the Nth round and the comparison error in the (N-1)th round.

3. The engine temperature prediction method according to claim 1, characterized in that, After determining the temperature prediction model, the training steps of the temperature prediction model further include: Based on each of the aforementioned comparison errors, the similarity between each pair of adjacent target hidden layers is calculated to obtain the similarity matrix of the temperature prediction model; The similarity matrix is ​​compared with the preset identity matrix to obtain the comparison error fusion matrix. The contrast error fusion matrix is ​​used as an adjacency matrix and input into the graph convolutional network along with the reconstructed feature vectors corresponding to each of the target hidden layers to perform feature fusion and generate a fused feature vector. Based on the fused feature vector, the parameters of the temperature prediction model are optimized.

4. The engine temperature prediction method according to claim 2, characterized in that, The step of determining the number of target hidden layers in the initial model based on the comparison error in the Nth round and the comparison error in the (N-1)th round includes: If the comparison error in the Nth round is greater than the comparison error in the (N-1)th round, the number of rounds in the Nth round is determined as the number of the target hidden layer. If the comparison error in the Nth round is less than or equal to the comparison error in the (N-1)th round, the feature vector of the encoder after dimensionality reduction in the Nth round is input into the (N+1)th hidden layer for encoding and decoding reconstruction processing in the (N+1)th round.

5. The engine temperature prediction method according to claim 1, characterized in that, The encoder performs dimensionality reduction on the input feature vector, including: The encoder maps the input feature vector to a low-dimensional space, and performs initial dimensionality reduction on the feature vector in the low-dimensional space to obtain a low-dimensional space feature vector. The feature vectors in the low-dimensional space are randomly sampled to obtain the feature vectors after dimensionality reduction by the encoder.

6. The engine temperature prediction method according to claim 1, characterized in that, The feature vectors after dimensionality reduction by the encoder are then subjected to dimensionality upscaling to obtain reconstructed feature vectors, including: The decoder performs initial up-dimensionality processing on the feature vector after dimensionality reduction by the encoder to obtain an initial reconstructed feature vector. The feature structure of the initial reconstructed feature vector is the same as the feature structure of the feature vector after dimensionality reduction by the encoder. Based on the local topological structure information of the initial reconstructed feature vector, the initial reconstructed feature vector is increased in dimensionality through multiple rounds of the dimensionality-increasing layer of the graph convolutional network to obtain the reconstructed feature vector. The number of rounds of dimensionality increase through the dimensionality-increasing layer is the same as the number of layers in the hidden layer where the decoder is located, and the feature structure of the reconstructed feature vector is the same as the feature structure of the initial feature vector.

7. The engine temperature prediction method according to claim 1, characterized in that, The process of encoding, decoding, and reconstructing the temperature data at the current moment using the encoders and decoders of the multiple target hidden layers to obtain the predicted temperature data of the engine at the next moment includes: The temperature data at the current moment is subjected to feature extraction to obtain the initial feature vector of the temperature data at the current moment; The initial feature vector of the temperature data at the current moment is sequentially input into the encoder of each target hidden layer for dimensionality reduction, and the feature vector after dimensionality reduction by each encoder is input into its corresponding decoder to obtain multiple feature vectors after dimensionality increase by each decoder. The multiple feature vectors after being upscaled by each of the decoders are respectively input into the upscaled layer of the temperature prediction model. In the upscaled layer, the multiple feature vectors after being upscaled by each of the decoders are upscaled according to the layer value of their respective target hidden layers to obtain the reconstructed feature vectors corresponding to each target hidden layer. Based on the contrast error fusion matrix determined during the model training phase, feature fusion is performed on the reconstructed feature vector to obtain the fused feature vector; Based on the fused feature vector, the predicted temperature data of the engine at the next moment is determined.

8. An engine temperature prediction system, characterized in that, The system includes: The input module is used to input the engine's current temperature data into a pre-trained temperature prediction model. The temperature prediction model includes multiple preset target hidden layers, and each target hidden layer includes a set of encoders and decoders. The encoding / decoding and reconstruction processing module is used to encode and decode the temperature data at the current moment through the encoders and decoders of the multiple target hidden layers to obtain the predicted temperature data of the engine at the next moment. The temperature prediction model includes: The encoding, decoding, and reconstruction processing submodule is used to sequentially input historical temperature data into each hidden layer of the initial model, perform dimensionality reduction processing through the encoder of the hidden layer, and perform dimensionality increase processing through the decoder of the hidden layer and the dimensionality increase layer of the graph convolutional network to obtain the reconstruction feature vector corresponding to each hidden layer. The first calculation submodule is used to calculate the comparison error between the reconstructed feature vector and the initial feature vector of each hidden layer, and obtain the comparison error of each hidden layer. The first determining submodule is used to determine the target number of hidden layers in the initial model based on the contrast error corresponding to each hidden layer, and to use the initial model with the determined target number of hidden layers as the temperature prediction model.

9. An electronic device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the engine temperature prediction method as described in any one of claims 1-7.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the engine temperature prediction method as described in any one of claims 1-7.

Images