Model training method and device, temperature prediction method and device and computer storage medium

Abstract

The invention provides a model training method, a temperature prediction method, a temperature prediction device and a computer storage medium. The model training method comprises the following steps: acquiring a training feature matrix of rack training data of the permanent magnet synchronous motor; obtaining a feature vector of the training feature matrix and a reconstruction feature matrix; obtaining a temperature prediction value of the bench training data based on the feature vector; determining reconstruction loss based on the training feature matrix and the reconstruction feature matrix; determining a prediction loss based on a temperature prediction value and a temperature true value of the bench training data; and training a temperature prediction model by using the reconstruction loss and the prediction loss. According to the model training method, the temperature prediction model is trained through the deep neural network technology, real-time motor rotor temperature prediction in an actual application scene is achieved, and therefore the real-time performance and accuracy of prediction are improved.

Description

Technical Field

[0001] This application relates to the technical field of motor temperature detection, and particularly to a model training method, a temperature prediction method, a temperature prediction device, and a computer storage medium. Background Art

[0002] The rotor temperature prediction based on a theoretical model usually relies on magnetic flux calculation or parameter estimation, which makes it vulnerable to harsh working conditions or noise interference, resulting in a large deviation in the prediction result. In addition, this method often relies on expert knowledge or subjective experience, and the prediction process is complex and the generalization ability is weak. Summary of the Invention

[0003] To solve the above technical problems, this application proposes a model training method, a temperature prediction method, a temperature prediction device, and a computer storage medium.

[0004] To solve the above technical problems, this application proposes a model training method, and the model training method includes:

[0005] Obtain the training feature matrix of the bench test training data of the permanent magnet synchronous motor;

[0006] Obtain the eigenvector of the training feature matrix and the reconstructed feature matrix;

[0007] Obtain the temperature prediction value of the bench test training data based on the eigenvector;

[0008] Determine the reconstruction loss based on the training feature matrix and the reconstructed feature matrix;

[0009] Determine the prediction loss based on the temperature prediction value and the true temperature value of the bench test training data;

[0010] Train the temperature prediction model by using the reconstruction loss and the prediction loss.

[0011] Among them, the obtaining the training feature matrix of the bench test training data of the permanent magnet synchronous motor includes:

[0012] Extract the original features of the bench test training data of the permanent magnet synchronous motor;

[0013] Calculate the derived features based on the original features;

[0014] Fuse the original features and the derived features into a feature matrix.

[0015] Among them, before fusing the original features and the derived features into a feature matrix, the model training method further includes:

[0016] Obtain the feature mean and feature standard deviation of the original features and the derived features;

[0017] Calculate the standardized features of the original features and / or the derived features by using the feature mean value and the feature standard deviation.

[0018] Among them, the reconstruction losses of the training feature matrix and the reconstruction feature matrix include cosine similarity loss and Euclidean distance loss.

[0019] Among them, the training of the temperature prediction model by using the reconstruction loss and the prediction loss includes:

[0020] Obtain the weight decay term of the temperature prediction model;

[0021] Adjust the learning rate based on the first moment and the second moment of the gradient of the model parameters of the temperature prediction model;

[0022] Based on the weight decay term and the learning rate, iteratively update the model parameters of the temperature prediction model through the reconstruction loss and the prediction loss.

[0023] Among them, after training the temperature prediction model by using the reconstruction loss and the prediction loss, the model training method further includes:

[0024] Extract the bench feature vector of the bench data and the real vehicle feature vector of the real vehicle data by using the temperature prediction model;

[0025] Obtain the domain adaptation loss based on the bench feature vector and the real vehicle feature vector;

[0026] Train the temperature prediction model by using the domain adaptation loss.

[0027] Among them, the model training method further includes:

[0028] Obtain the temperature prediction value of the bench data based on the bench feature vector;

[0029] Determine the bench prediction loss based on the temperature prediction value and the true temperature value of the bench data;

[0030] The training of the temperature prediction model by using the domain adaptation loss includes:

[0031] Train the temperature prediction model by using the domain adaptation loss and the bench prediction loss.

[0032] To solve the above technical problems, the present application also proposes a temperature prediction method, and the temperature prediction method includes:

[0033] Obtain the operation data of the permanent magnet synchronous motor in the vehicle;

[0034] Input the operation data into a pre-trained temperature prediction model to obtain the rotor temperature of the permanent magnet synchronous motor;

[0035] Wherein, the temperature prediction model is obtained through the above-mentioned model training method.

[0036] To solve the above technical problems, the present application also proposes a temperature prediction device, which includes a memory and a processor coupled to the memory; wherein, the memory is used to store program data, and the processor is used to execute the program data to implement the model training method and / or the temperature prediction method as described above.

[0037] To solve the above technical problems, the present application also proposes a computer storage medium, which is used to store program data. When the program data is executed by a computer, it is used to implement the model training method and / or the temperature prediction method as described above.

[0038] Compared with the prior art, the beneficial effects of the present application are as follows: The temperature prediction device obtains the training feature matrix of the bench test data of the permanent magnet synchronous motor; obtains the eigenvector and the reconstructed feature matrix of the training feature matrix; obtains the temperature prediction value of the bench test data based on the eigenvector; determines the reconstruction loss based on the training feature matrix and the reconstructed feature matrix; determines the prediction loss based on the temperature prediction value and the true temperature value of the bench test data; and trains the temperature prediction model using the reconstruction loss and the prediction loss. Through the above model training method, the temperature prediction model is trained by deep neural network technology to realize real-time prediction of the motor rotor temperature in the actual application scenario, thereby improving the real-time performance and accuracy of the prediction. Description of the Drawings

[0039] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0040] Wherein:

[0041] Figure 1 is a schematic flowchart of an embodiment of the model training method provided by the present application;

[0042] Figure 2 is a schematic flowchart of the pre-training of the temperature prediction model provided by the present application;

[0043] Figure 3 is a schematic diagram of the DAE-LSTM training structure provided by the present application;

[0044] Figure 4 It is a schematic flowchart of another embodiment of the model training method provided by this application;

[0045] Figure 5 It is a schematic diagram of the domain adaptation training process provided by this application;

[0046] Figure 6 It is a domain adaptation flowchart of the temperature prediction model provided by this application;

[0047] Figure 7 It is a schematic flowchart of an embodiment of the temperature prediction method provided by this application;

[0048] Figure 8 It is a schematic structural diagram of an embodiment of the temperature prediction device provided by this application;

[0049] Figure 9 It is a schematic structural diagram of an embodiment of the computer storage medium provided by this application. Specific Embodiments

[0050] Next, the technical solutions in the embodiments of this application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of this application.

[0051] The terms "first", "second", "third", "fourth", etc. (if any) in the specification, claims and above-mentioned drawings of this application are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of this application described here, for example, can be implemented in an order different from those illustrated or described here. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0052] Permanent magnet synchronous motors are widely used in fields such as electric vehicles and industrial robots due to their high power density, high efficiency, wide speed regulation range and high reliability. A permanent magnet synchronous motor mainly consists of a stator, a rotor, and an end cover, etc. Among them, the stator is similar to a common induction motor and uses a laminated design to reduce iron loss, while the rotor can adopt a solid or laminated structure. The winding design of the motor is diverse, including concentrated full pitch, distributed short pitch or unconventional forms, etc.

[0053] Although permanent magnet synchronous motors have many advantages, their performance depends to a large extent on the control and management of their internal temperature. Accurately predicting the temperature of permanent magnet synchronous motors is crucial for ensuring their safe and efficient operation, extending their service life, and improving the performance of drive systems. However, the direct measurement of rotor temperature remains challenging. This is mainly because the rotor is located inside the motor, making it inconvenient to install traditional temperature sensors, resulting in difficulty in directly measuring the rotor temperature. In addition, calculation methods relying on fixed-value assumptions cannot meet the requirements of high-precision control. Therefore, it is particularly important to design an efficient rotor temperature prediction method. Real-time prediction of rotor temperature has become an effective means to ensure the reliability of the motor, which helps to improve the stability of the motor output.

[0054] The rotor temperature directly affects the efficiency and life of the motor. Excessive temperature may cause the aging, performance degradation, and even damage of insulating materials. Therefore, real-time monitoring and prediction of rotor temperature are crucial. Accurate rotor temperature prediction can help design better cooling strategies and control algorithms, thereby improving the performance and reliability of the motor. Through preventive maintenance and real-time monitoring, it can reduce faults caused by overheating and extend the service life of the motor.

[0055] Applying deep learning to motor temperature prediction not only helps to promote intelligent manufacturing but also significantly improves production efficiency and product quality. The estimation accuracy of the rotor temperature model is of great significance in fault diagnosis and maintenance during operation. By accurately predicting the rotor temperature, the cloud service system can monitor the working state of the rotor in real time, detect potential faults in a timely manner, and take corresponding repair measures. If the prediction accuracy of the rotor temperature model is insufficient, it may lead to misjudgment of the rotor operating state, thus delaying fault diagnosis and repair. In addition, in the field of electric vehicles, optimizing motor temperature management is a key factor in improving the driving range and safety. Therefore, this research has important industrial application value.

[0056] In response to this, a rotor temperature prediction method for permanent magnet synchronous motors based on domain adaptation is proposed. This method includes model pre-training on bench data, feature alignment, and model migration on real vehicle data.

[0057] The following first introduces the part of model pre-training on bench data. For details, please refer to Figure 1 and Figure 2 , Figure 1 is a schematic flowchart of an embodiment of the model training method provided by this application, Figure 2 is a schematic flowchart of the pre-training of the temperature prediction model provided by this application.

[0058] The model training method of the present application is applied to a temperature prediction device. Among them, the temperature prediction device of the present application can be a server, a terminal device, or a system in which the server and the terminal device cooperate with each other. Correspondingly, each part included in the temperature prediction device, such as each unit, subunit, module, and submodule, can be all set in the server, all set in the terminal device, or respectively set in the server and the terminal device.

[0059] Furthermore, the above-mentioned server can be hardware or software. When the server is hardware, it can be implemented as a distributed server cluster composed of multiple servers or as a single server. When the server is software, it can be implemented as multiple software or software modules, such as software or software modules used to provide a distributed server, or as a single software or software module, which is not specifically limited here.

[0060] As Figure 1 shown, the specific steps are as follows:

[0061] Step S11: Obtain the training feature matrix of the bench test data of the permanent magnet synchronous motor.

[0062] In the embodiment of the present application, Figure 2 the source domain data preprocessing is the acquisition and data preprocessing of the bench test data by the temperature prediction device.

[0063] Specifically, the temperature prediction device collects the inlet water temperature, controller temperature, oil pump temperature, oil pump speed, oil pump current, oil pump bus voltage, d-axis current I d , q-axis current I q , d-axis voltage U d , q-axis voltage U q , speed w, torque Tor, rotor temperature, etc. of the permanent magnet synchronous motor.

[0064] Among them, the rotor temperature is only collected on the bench by installing a slip ring device, and the rotor temperature does not need to be collected on the actual vehicle. The bench test data collection includes a large amount of motor data under different working conditions, including motor data at different speeds, different torques, different inlet water temperatures, and different inlet water flows.

[0065] Therefore, the training input data of the temperature prediction model of the present application includes 12 original features such as inlet water temperature, controller temperature, oil pump temperature, oil pump speed, oil pump current, oil pump bus voltage, d-axis current, q-axis current, d-axis voltage, q-axis voltage, speed, and torque, and 3 derived features such as total current I s , total voltage U s , power loss P loss , with a total of 15 input features.

[0066] Among them, the calculation formula for the above-derived features is as follows:

[0067]

[0068]

[0069] Furthermore, in order to achieve feature extraction between different types of data, the temperature prediction device also needs to perform data preprocessing on the above features.

[0070] Specifically, the temperature prediction device converts the collected data into a distribution with a mean of 0 and a variance of 1 through data standardization. The standardized data has the same scale, which can effectively reduce the influence of inconsistent dimensions between features. The process of data standardization is as follows:

[0071]

[0072] Among them, α is the mean of the feature, and β is the standard deviation of the feature.

[0073] Finally, the temperature prediction device converts all the features after data preprocessing into a feature matrix, that is, it divides the time series of all features with a unit time window to form a feature matrix with a size of (number of windows, window length, 15) for input to the neural network.

[0074] This application combines multiple sensor input sources (such as inlet temperature, controller temperature, d-axis current, q-axis current, d-axis voltage, q-axis voltage, speed, and torque) for prediction, thereby comprehensively considering the influence of different factors on the rotor temperature. This multi-variable fusion learning method implemented using a deep neural network can improve the stability and accuracy of prediction. When the motor is in a stationary state or running at a low speed, relatively accurate rotor temperature prediction can still be achieved.

[0075] Step S12: Obtain the feature vectors of the training feature matrix and the reconstructed feature matrix.

[0076] In the embodiment of this application, the temperature prediction device performs model training on the bench test training data, and inputs the feature matrix obtained by preprocessing the 12 original features (such as inlet temperature, controller temperature, oil pump temperature, oil pump speed, oil pump current, oil pump bus voltage, d-axis current, q-axis current, d-axis voltage, q-axis voltage, speed, and torque) and 3 derived features into the deep neural network to predict the rotor temperature.

[0077] Step S13: Obtain the temperature prediction value of the bench test training data based on the feature vectors.

[0078] In the embodiment of the present application, taking the deep neural network of Deep Auto-Encoder Long Short-Term Memory (DAE-LSTM) as the model basis of the temperature prediction model as an example, the training process is continued to be introduced:

[0079] For details, please refer to Figure 3 , Figure 3 which is the schematic diagram of the DAE-LSTM training structure provided by the present application.

[0080] As Figure 3 shown, the temperature prediction device inputs the training feature matrix X that has undergone data preprocessing into the encoder to extract the feature vectors of the training feature matrix. On the one hand, the feature vectors are input into the fully connected layer to obtain the temperature prediction value of the bench training data; on the other hand, the feature vectors are input into the decoder to obtain the reconstructed feature matrix Xr of the training feature matrix X.

[0081] Step S14: Determine the reconstruction loss based on the training feature matrix and the reconstructed feature matrix.

[0082] In the embodiment of the present application, the temperature prediction device determines the reconstruction loss of the bench training data according to the training feature matrix and the reconstructed feature matrix. As Figure 3 shown, the reconstruction loss of DAE-LSTM consists of two parts: cosine similarity loss and Euclidean distance loss.

[0083] Among them, the reconstruction loss L rec is calculated as follows:

[0084]

[0085] where m is the number of samples of the bench training data, d is the number of features of each sample; L s is the cosine similarity loss, and L d is the Euclidean distance loss.

[0086] Step S15: Determine the prediction loss based on the temperature prediction value and the true temperature value of the bench training data.

[0087] In the embodiment of the present application, the prediction loss L MSE is calculated as follows:

[0088]

[0089] where y i is the true temperature of the rotor of the motor, is the predicted value of the rotor temperature output by the model.

[0090] Step S16: Train the temperature prediction model using the reconstruction loss and the prediction loss.

[0091] In the embodiment of the present application, the temperature prediction device trains the temperature prediction model using the reconstruction loss and the prediction loss calculated in the above steps. The iterative process of the network model parameters of the temperature prediction model is as follows:

[0092] The temperature prediction device mainly uses AdamW (Adam Weight Decay, an Adam optimizer with weight decay) for network model iteration. By improving the way of weight decay (L2 regularization), the optimization effect is enhanced. AdamW directly applies weight decay (L2 regularization) to parameter updates instead of adding a regularization term to the loss function. The weight decay is independent of the gradient calculation, making the weight updates more stable. This can avoid introducing biases in parameter updates and improve the generalization ability of the model.

[0093] In addition, AdamW also uses an adaptive learning rate mechanism to dynamically adjust the learning rate according to the first moment (average value) and the second moment (variance) of the gradient of each parameter. This feature enables the optimizer to better handle sparse gradients and parameters of different scales. Through more effective weight decay and dynamic adjustment of the learning rate, AdamW can help the model better prevent overfitting, especially when the training data is less.

[0094] The following is the process of AdamW using the loss gradient backpropagation to iteratively adjust the network parameters. The update process is as follows:

[0095] Bias-corrected momentum estimation:

[0096]

[0097] Bias correction:

[0098]

[0099] Network parameter update:

[0100]

[0101] Among them, the learning rate is denoted as η, and ò is a very small constant to prevent division by zero. The first moment estimate and the second moment estimate at the current time step t are respectively represented by ; λ is the weight decay coefficient.

[0102] In a specific embodiment, the temperature prediction device sets the target number of training epochs to 100, and the initial learning rate is 1e -3, the learning rate dynamic adjustment strategy is CosineAnnealingLR (cosine annealing to adjust the learning rate). When starting to train the model, after each epoch ends, calculate the loss of the current model on the validation set. When the epoch reaches the target number of training epochs or the loss on the validation set has not decreased for 10 epochs, stop training and save the current network parameters.

[0103] In this application, the temperature prediction device obtains the training feature matrix of the bench test data of the permanent magnet synchronous motor; obtains the eigenvector and the reconstructed feature matrix of the training feature matrix; obtains the temperature prediction value of the bench test data based on the eigenvector; determines the reconstruction loss based on the training feature matrix and the reconstructed feature matrix; determines the prediction loss based on the temperature prediction value and the true temperature value of the bench test data; and trains the temperature prediction model using the reconstruction loss and the prediction loss. Through the above model training method, the temperature prediction model is trained by deep neural network technology to realize real-time motor rotor temperature prediction in actual application scenarios, thereby improving the real-time performance and accuracy of the prediction.

[0104] Furthermore, existing rotor temperature prediction methods based on machine learning usually train with data collected under specific working conditions. However, due to the large difference between experimental conditions and actual working conditions, the performance of the model trained with data obtained from bench tests is often poor on real data.

[0105] Traditional rotor temperature prediction methods mostly design models for specific types of motors, resulting in the lack of generality of these models. When it is necessary to predict for other types of motors, it is often necessary to redesign the model, thus reducing the development efficiency.

[0106] Therefore, based on the pre-training of the temperature prediction model in the above embodiments, this application uses domain adaptation technology to transfer the pre-trained model to real vehicle data. Among them, the basic idea of domain adaptation is to establish a shared model between the source domain and the target domain, and realize the transfer from the source domain to the target domain through feature transformation and model fine-tuning. Here, the source domain refers to bench data, and the target domain refers to real vehicle data.

[0107] By introducing domain adaptation technology, on the one hand, it can improve the performance of the model trained with data obtained from bench tests on real data, and on the other hand, it can adaptively adjust the model for different types of motor models at any time without redesigning the model, greatly improving the development efficiency.

[0108] For details, please refer to Figure 4 、 Figure 5 and Figure 6 , Figure 4 is the schematic flow chart of another embodiment of the model training method provided by this application,Figure 5 It is a schematic diagram of the domain adaptation training process provided by this application. Figure 6 It is the domain adaptation flowchart of the temperature prediction model provided by this application.

[0109] As Figure 4 shown, the specific steps are as follows:

[0110] Step S21: Use the temperature prediction model to extract the bench feature vector of the bench data and the real vehicle feature vector of the real vehicle data.

[0111] In the embodiment of this application, the temperature prediction device loads the pre-trained network parameters, and then uses the pre-trained model on the bench data, that is, the training result of the above embodiment, to extract the bench feature vector F X of the bench data and the real vehicle feature vector F Y of the real vehicle data in parallel.

[0112] Step S22: Based on the bench feature vector and the real vehicle feature vector, obtain the domain adaptation loss.

[0113] In the embodiment of this application, the temperature prediction device uses the Gaussian kernel function to map the features of the source domain to the feature space of the target domain, and uses MMSD (Multimodal Sentiment Analysis Loss) to measure the maximum difference between the bench data and the real vehicle data feature vectors. In further training, minimize the maximum difference between the bench data and the real vehicle data feature vectors, so as to achieve domain adaptation. The domain-adapted model can accurately predict the rotor temperature of the bench data and the rotor temperature of the measured data at the same time.

[0114] Among them, the representation of the MMSD loss function is as follows:

[0115]

[0116] Among them,

[0117]

[0118] Among them, XX T , XY T , YY T is the inner product matrix of X and Y, X T is the transpose of the X matrix, the diag(·) function is used to extract the diagonal elements of the given matrix, w k is the hyperparameter of the Gaussian kernel function, and σ 2 is the variance of the input feature vector. MMSD not only focuses on the mean difference, but also considers the dispersion of the samples, and strengthens the penalty for large differences by calculating the squared kernel function.

[0119] Based on this, the bench feature vector F X and the real vehicle feature vector F Y extracted in parallel by the pre-trained model of this application are used to calculate the MMSD as follows:

[0120]

[0121]

[0122] where F X is the feature vector of the bench data extracted by the pre-trained model, and F Y is the feature vector of the real vehicle data extracted by the pre-trained model. μ X , μ Y represent the means of and respectively. ‖·‖ F represents the calculation of the Frobenius norm. m is the number of samples of the bench data, and n is the number of samples of the real vehicle data.

[0123] The domain adaptation loss of this application consists of two parts: MMSD and CORAL. Among them, MMSD measures the difference between distributions by maximizing the mean difference and dispersion degree, while CORAL (orrelation Regularization Loss, correlation alignment loss) aligns the covariance matrices of the source domain and the target domain, making the high-order statistics (especially the covariance) of the source domain and the target domain consistent. In this way, CORAL can perform a more fine-grained alignment between the source domain and the target domain. Especially for those cases with similar means but different covariance structures, CORAL can effectively improve the alignment effect.

[0124] Step S23: Train the temperature prediction model using the domain adaptation loss.

[0125] In the embodiment of this application, as Figure 5 shown, during the domain adaptation training process, the temperature prediction device can also train the temperature prediction model by combining the prediction loss of the bench data with the domain adaptation loss determined in step S22. Therefore, the training objective during the domain adaptation process is:

[0126]

[0127] Furthermore, as Figure 6 shown, the temperature prediction device can also perform gradient backpropagation and update the model parameters using AdamW and CosineAnnealingLR, and the process is not elaborated here.

[0128] Please continue to refer to Figure 7 , Figure 7It is a schematic flowchart of an embodiment of the temperature prediction method provided by this application.

[0129] As Figure 7 shown, the specific steps are as follows:

[0130] Step S31: Obtain the operation data of the permanent magnet synchronous motor in the vehicle.

[0131] Step S32: Input the operation data into a pre-trained temperature prediction model to obtain the rotor temperature of the permanent magnet synchronous motor.

[0132] Among them, the pre-trained temperature prediction model can be Figure 1 or Figure 4 the temperature prediction model obtained by training in the shown embodiment.

[0133] The temperature prediction method of this application combines deep neural network and domain adaptation technology, and then realizes real-time electronic rotor temperature prediction in actual application scenarios, thereby improving the real-time performance and accuracy of prediction. This can not only improve the precision of motor control, but also extend the service life of the motor and ensure its reliability under high-load conditions.

[0134] This application provides an innovative design for the data acquisition scheme under complex working conditions: a multi-dimensional data acquisition scheme suitable for real scenarios is proposed, and the data acquisition scheme under complex working conditions and different heat dissipation conditions in real scenarios is used to enrich the training data set as much as possible, enriching the diversity and authenticity of the training data, and providing strong data support for the subsequent process of domain adaptation.

[0135] This application provides a multi-feature input model and optimization strategy: the combination of sensor data such as inlet temperature, controller temperature, d-axis current, q-axis current, d-axis voltage, q-axis voltage, speed, torque, and derived features is used, and 15 features are selected as inputs, as well as the dynamic adjustment strategy of the AdamW optimizer and CosineAnnealingLR (cosine annealing to adjust the learning rate), ensuring the robustness and generalization ability of the model under complex conditions. This design can effectively enhance the prediction accuracy of the model in a multi-domain data environment.

[0136] This application uses the DAE-LSTM structure for feature dimension reduction and time series prediction to improve the rotor temperature prediction ability, and uses the combined measurement of cosine similarity and Euclidean distance to reconstruct the loss.

[0137] This application uses the joint optimization of the MMSD, CORAL, and MSE (Mean Squared Error Loss) loss functions to solve the feature alignment problem and enhance the domain adaptation ability. It ensures the alignment of the features of bench data and real vehicle data, and realizes the efficient migration and feature alignment between the source domain and the target domain through this method.

[0138] This application provides a method and process for feature alignment in cross-domain migration: It protects the methods and steps for implementing model migration between different data domains, especially the technical details of minimizing the differences between bench data (source domain or pre-training scenario) and (target domain or target migration scenario) through feature transformation during the domain adaptation process, so as to ensure the robustness and prediction stability of the model in cross-domain scenarios.

[0139] Those skilled in the art can understand that in the above method of the specific implementation manner, the writing order of each step does not mean a strict execution order that constitutes any limitation to the implementation process, and the specific execution order of each step should be determined by its function and possible internal logic.

[0140] To implement the above model training method, and / or temperature prediction method, this application also proposes a temperature prediction device. For details, please refer to Figure 8 , Figure 8 which is a schematic structural diagram of an embodiment of the temperature prediction device provided by this application.

[0141] The temperature prediction device 400 of this embodiment includes a processor 41, a memory 42, an input / output device 43, and a bus 44.

[0142] The processor 41, the memory 42, and the input / output device 43 are respectively connected to the bus 44. The memory 42 stores program data, and the processor 41 is used to execute the program data to implement the model training method and / or temperature prediction method described in the above embodiment.

[0143] In the embodiment of the present application, the processor 41 can also be referred to as a CPU (Central Processing Unit). The processor 41 may be an integrated circuit chip with signal processing capabilities. The processor 41 can also be a general-purpose processor, a digital signal processor (DSP, Digital Signal Process), an application-specific integrated circuit (ASIC, Application Specific Integrated Circuit), a field-programmable gate array (FPGA, Field Programmable Gate Array), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The general-purpose processor can be a microprocessor or the processor 41 can also be any conventional processor, etc.

[0144] The present application also provides a computer storage medium. Please continue to refer to Figure 9 , Figure 9 FIG. is a schematic structural diagram of an embodiment of the computer storage medium provided by the present application. The computer storage medium 600 stores a computer program 61. When the computer program 61 is executed by a processor, it is used to implement the model training method and / or the temperature prediction method in the above embodiments.

[0145] When the embodiments of the present application are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the methods described in various embodiments of the present application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0146] The above are only the embodiments of the present application, and do not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present application, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present application.

Claims

1. A model training method, characterized in that, The model training method includes: Obtain the training feature matrix of the bench test data of the permanent magnet synchronous motor; Obtain the eigenvector of the training feature matrix and the reconstructed feature matrix; Obtain the temperature prediction value of the bench test data based on the eigenvector; Determine the reconstruction loss based on the training feature matrix and the reconstructed feature matrix; Determine the prediction loss based on the temperature prediction value and the true temperature value of the bench test data; Train the temperature prediction model using the reconstruction loss and the prediction loss.

2. The model training method according to claim 1, wherein: The obtaining the training feature matrix of the bench test data of the permanent magnet synchronous motor includes: Extract the original features of the bench test data of the permanent magnet synchronous motor; Calculate the derived features based on the original features; Fuse the original features and the derived features into a feature matrix.

3. The model training method according to claim 2, wherein: Before fusing the original features and the derived features into a feature matrix, the model training method further includes: Obtain the feature mean and feature standard deviation of the original features and the derived features; Calculate the standardized features of the original features and / or the derived features using the feature mean and the feature standard deviation.

4. The model training method according to claim 1, wherein: The reconstruction loss between the training feature matrix and the reconstructed feature matrix includes the cosine similarity loss and the Euclidean distance loss.

5. The model training method according to claim 1, wherein: The training the temperature prediction model using the reconstruction loss and the prediction loss includes: Obtain the weight decay term of the temperature prediction model; Adjust the learning rate based on the first moment and the second moment of the gradient of the model parameters of the temperature prediction model; Iteratively update the model parameters of the temperature prediction model through the reconstruction loss and the prediction loss based on the weight decay term and the learning rate.

6. The model training method according to claim 1, wherein: After training the temperature prediction model using the reconstruction loss and the prediction loss, the model training method further includes: Extract the bench feature vector of the bench data and the real vehicle feature vector of the real vehicle data using the temperature prediction model; Obtain the domain adaptation loss based on the bench feature vector and the real vehicle feature vector; Train the temperature prediction model using the domain adaptation loss.

7. The model training method according to claim 6, wherein: The model training method further includes: Obtain the temperature prediction value of the bench data based on the bench feature vector; Determine the bench prediction loss based on the temperature prediction value and the true temperature value of the bench data; The training the temperature prediction model using the domain adaptation loss includes: Train the temperature prediction model using the domain adaptation loss and the bench prediction loss.

8. A temperature prediction method, characterized in that, The temperature prediction method includes: Obtain the operation data of the permanent magnet synchronous motor in the vehicle; Input the operation data into the pre-trained temperature prediction model to obtain the rotor temperature of the permanent magnet synchronous motor. Among them, the temperature prediction model is obtained through the model training method according to any one of claims 1 to 7.

9. A temperature prediction device, characterized in that, The temperature prediction device includes a memory and a processor coupled to the memory; Among them, the memory is used to store program data, and the processor is used to execute the program data to implement the model training method according to any one of claims 1 to 7, and / or the temperature prediction method according to claim 8.

10. A computer storage medium, characterized in that, The computer storage medium is used to store program data, and when the program data is executed by a computer, it is used to implement the model training method according to any one of claims 1 to 7, and / or the temperature prediction method according to claim 8.

Images