A motor system identification method based on CMAWL-PINN

Abstract

The application relates to the technical field of numerical control, and particularly relates to a motor system identification method based on CMAWL-PINN. The permanent magnet synchronous motor system has the characteristics of strong nonlinearity, high data coupling, time lag, dynamicity and the like, so that it is difficult to realize accurate identification effect. In view of the above problems, the CMAWL-PINN system identification method is designed and invented. The method is composed of a double-layer LSTM network, a cascaded multi-attention mechanism, a physical information neural network (PINN) framework and an adaptive window adjustment mechanism based on residual statistics and evaluation indexes, and an Optuna super parameter optimization framework is used to find the optimal super parameter combination. Through experimental verification on a large amount of data sets and comparative experiments with traditional mainstream methods and current improved methods, the application method shows excellent performance and stronger robustness and adaptability under different working conditions.

Description

Technical Field

[0001] This invention belongs to the field of numerical control technology, specifically a motor system identification method based on CMAWL-PINN. Background Technology

[0002] As a core power source in modern industry, the performance and reliability of permanent magnet synchronous motors directly affect the operating efficiency and safety of the entire system. From the perspective of the equipment's operation, it is a typical nonlinear system, exhibiting nonlinear and data-coupled characteristics. Furthermore, most industrial systems also contain uncertainties such as time delays, external disturbances, and noise. It is difficult to establish accurate mathematical models, making the identification and research of such systems quite challenging.

[0003] Currently, system identification methods are mainly divided into two categories: those based on physical models and those based on data. Representative methods based on physical models include the Hammerstein-Wiener model and the Lur'e model, which estimate the target distribution by establishing a mathematical model of the entire system. These methods have clear physical meaning and strong interpretability, but suffer from problems such as numerous simplification assumptions and difficulties in parameter calibration. Data-driven methods, such as regression analysis, support vector machines, and neural networks, predict future data by learning patterns of change in historical data. These methods are characterized by simple modeling and strong adaptability, but they often neglect the physical laws governing changes in the motor system, resulting in limited generalization ability, especially when facing new operating conditions.

[0004] With the development of deep learning technology, Physical Information Neural Networks (PINNs), as a novel method integrating physical knowledge and deep learning, have attracted widespread attention in scientific computing and engineering fields in recent years. PINNs achieve respect for physical laws and good data fitting by embedding physical laws as constraints into the loss function of the neural network. In fields such as fluid mechanics, heat conduction, and structural mechanics, PINNs have been successfully applied to the modeling and prediction of complex physical systems, demonstrating significant advantages in solving partial differential equations and parameter identification problems. Therefore, combining PINNs with temporal deep learning models and introducing attention mechanisms to enhance the model's ability to extract key information is an effective way to improve the identification performance of motor systems. Summary of the Invention

[0005] To address the aforementioned problems, this invention proposes a motor system identification method based on the Cascaded Multi-Attention Adaptive Window CMAWL-PINN (CMAWL-PINN) mechanism.

[0006] First, to improve the model's ability to identify key factors affecting changes in the motor system, this invention designs a cascaded multi-attention mechanism, comprising three cascaded attention layers: a temporal attention layer, a feature attention layer, and a physical constraint attention layer. By focusing on different aspects of the data through different types of attention mechanisms, and by fusing the outputs of these attention mechanisms, a more comprehensive and richer feature representation can be obtained.

[0007] The core idea of ​​cascaded fusion is to connect multiple attention mechanisms sequentially, with the output of one attention mechanism serving as the input to the next, forming a hierarchical structure for feature extraction. This approach allows attention mechanisms at different layers to share some parameters, reducing the total number of parameters and memory consumption. Furthermore, through progressive feature refinement, it can capture more complex patterns.

[0008] Secondly, to ensure that the recognition results conform to the physical laws of the system, two approaches are taken. First, by cascading a physical constraint attention layer within a multi-attention model, physical laws are implicitly learned, achieving the integration of physical knowledge at the feature level. Furthermore, the feature representation is adjusted based on the physical laws between feature dimensions. Second, a Physical Information Neural Network (PINN) framework is introduced to integrate physical constraints into the deep learning model. During training, physical laws are explicitly enforced, and through physical constraint loss combined with adaptive weights, the recognition results are ensured to conform to physical laws, achieving the integration of physical knowledge at the output level. The combination of these two approaches improves the model's interpretability and generalization ability.

[0009] Furthermore, in order to adapt to the dynamic characteristics of the system under different operating conditions of the motor, an adaptive window adjustment mechanism based on residual statistics and evaluation indicators is proposed. By analyzing and identifying the statistical characteristics of the residuals and the trend of model performance changes, the size of the time window is dynamically adjusted to improve the model's ability to capture features at different time scales.

[0010] The core idea is to determine whether the current window size is appropriate based on the relationship between the residual mean and residual variance. The residual mean reflects the systematic bias of the model, while the residual variance reflects the random error of the model. When the residual mean is greater than a threshold multiple of the residual variance, it indicates that systematic bias is dominant, and the window size needs to be increased to capture longer-term dependencies. When the residual mean is less than a threshold multiple of the residual variance, it indicates that random error is dominant, and the window size needs to be decreased to improve sensitivity to short-term changes.

[0011] Finally, in order to balance data fitting and physical consistency, an adaptive physical constraint weighting strategy is designed to dynamically adjust the influence of physical constraints based on identification error and performance, thereby balancing data fitting and physical consistency.

[0012] To improve model performance and reduce manual hyperparameter tuning, this method employs the Optuna hyperparameter optimization framework to automatically optimize key hyperparameters. Specifically, the Tree-structured Parzen Estimator (TPE) algorithm and early stopping strategy from Optuna are selected, along with its adaptive sampling strategy, to quickly approach and find the optimal hyperparameter combination under limited computational resources.

[0013] This invention adopts the following technical solution: a motor system identification method based on CMAWL-PINN, comprising the following steps:

[0014] Historical motor state data is acquired and preprocessed to obtain a motor state sequence;

[0015] Construct the CMAWL-PINN model and train it based on the motor state sequence;

[0016] Based on the motor state sequence, the motor temperature of the motor system is identified using the trained CMAWL-PINN model.

[0017] The construction of the CMAWL-PINN model includes:

[0018] LSTM is used to extract features from the input motor state sequence to obtain the input for the next attention mechanism.

[0019] A cascaded multi-attention mechanism is used to obtain feature representations based on the output of the LSTM, thereby improving the model's ability to identify key factors that have the greatest impact on motor state changes.

[0020] The entire connected hierarchy is used to represent the predicted motor temperature based on characteristics.

[0021] The physical information neural network is used to construct current constraints and heat conduction constraints based on the current, voltage, and ambient temperature in the input motor state sequence, and obtain the expression of the total physical constraints.

[0022] Adaptive physical constraint weights are used to construct the total loss function based on the expression of the total physical constraints, which is then used for model training.

[0023] An adaptive window adjustment mechanism is used to adaptively adjust the input window of historical motor state data based on the decision coefficients trained on the model.

[0024] The motor status includes motor temperature, current, voltage, speed, torque, coolant temperature, and ambient temperature. The motor temperature includes stator winding temperature, stator tooth temperature, stator yoke temperature, and permanent magnet surface temperature.

[0025] The cascaded multi-attention mechanism includes the following steps:

[0026] The temporal attention layer is used to identify the feature dimensions that have the greatest impact on different prediction targets. The process is as follows:

[0027]

[0028] Among them, g i This represents the attention score at the i-th time step. This is the output of the LSTM at time step i, where n is the vector dimension in the motor state sequence. and It is a learnable parameter, α i It is the attention weight at time step i. It is the weighted context vector;

[0029] The feature attention layer is used to identify the most important feature dimensions for different prediction targets, and the process is as follows:

[0030] β=σ(W b c)c′=β⊙c

[0031] in, It is the output of the time attention layer. These are learnable parameters, and σ is the sigmoid activation function. These are the attention weights along the feature dimension. It is the weighted eigenvector;

[0032] The physically constrained attention layer receives the output of the feature attention layer and further adjusts the feature representation according to the physical laws between feature dimensions. The process is as follows:

[0033] γ=tanh(W p c′)c″=c′⊙(1+γ)

[0034] in, It is the output of the feature attention layer. These are learnable parameters, and tanh is the hyperbolic tangent activation function. It is the attention weight of physical constraints. This is the final feature representation, where tanh represents the activation function and ⊙ represents the Hadamard product.

[0035] The physical constraints of the physical information neural network framework include current constraints and heat conduction constraints:

[0036] The current constraints are as follows:

[0037]

[0038] Among them, i d and iq These are the normalized d-axis and q-axis currents, respectively, and N is the batch size.

[0039] The thermal conduction constraints are as follows:

[0040]

[0041] Among them, T i It is the predicted motor temperature T output by the fully connected layer at the current moment. i-1 It is the predicted temperature at the previous moment, Δt is the time step, and k is the thermal conductivity coefficient;

[0042] The expression for the total physical constraints is:

[0043] L physics =λ1L current +λ2L heat

[0044] λ1 and λ2 are weighting coefficients used to balance the effects of different physical constraints.

[0045] The adaptive physical constraint weighting strategy is as follows:

[0046] The expression for the total loss function during model training is:

[0047] L total =L data +λ adaptive ·L physics

[0048] Among them, L data It is the data loss, i.e., the mean square error, L physics It is the physical constraint loss, λ adaptive It is an adaptive weight, and its calculation formula is:

[0049]

[0050] Where λ0 is the initial weights, epoch is the number of training iterations of the model, and R... 2 It is the determination coefficient for the current batch, and α and β are adjustment parameters.

[0051] The adaptive window adjustment mechanism includes the following steps:

[0052] 1) Calculate the direction of change:

[0053]

[0054] Among them, e i It is the mean residual of the current batch. It represents the residual variance, and θ is a parameter; a direction value of 1 indicates that the sequence size increases, and a direction value of -1 indicates that the sequence size decreases.

[0055] 2) Based on the coefficient of determination R in the most recent rounds of training 2 The change in the value of the model is compared with a preset threshold to determine whether the model performance has stagnated, as follows:

[0056]

[0057] Where, ΔR 2 This indicates the change in the coefficient of determination. The coefficient of determination for the currently trained model. The coefficient of determination of the model from the previous training round;

[0058]

[0059] Where is_stagnation represents a boolean value, True indicates that the model performance has stagnated, and False indicates that the model performance has not stagnated and that the training is normal.

[0060] 3) When model performance stagnates for multiple consecutive rounds, a window adjustment strategy is used to calculate the window change Δw:

[0061]

[0062] Among them, w t It is the current window size;

[0063] When normal model training is detected, the window size is adjusted based on continuous changes in residual statistics:

[0064]

[0065] Where γ is the adjustment step size parameter;

[0066] 4) Introduce the minimum adjustment mechanism as follows:

[0067]

[0068] Where, Δw min The threshold is epoch, which represents the number of training iterations of the model; epoch / 8 = 0 indicates that the epoch is a multiple of 8.

[0069] Ultimately, the window size update rule is as follows:

[0070]

[0071] Among them, w t+1 w represents the window size at time t+1.min and w max These are the lower and upper limits of the window size, respectively.

[0072] A motor system identification device based on CMAWL-PINN, comprising:

[0073] The data acquisition unit is used to acquire historical motor state data and perform preprocessing to obtain a motor state sequence;

[0074] The model training unit is used to build the CMAWL-PINN model and train the model based on the motor state sequence.

[0075] The state identification unit is used to identify the motor temperature of the motor system based on the motor state sequence using a trained CMAWL-PINN model.

[0076] The CMAWL-PINN model includes:

[0077] The LSTM layer is used to extract features from the input motor state sequence, which then becomes the input for the attention mechanism in the next step.

[0078] A cascaded multi-attention mechanism module is used to obtain feature representations based on the output of LSTM, thereby improving the model's ability to identify key factors in motor state changes.

[0079] The entire connected hierarchy is used to represent the predicted motor temperature based on characteristics.

[0080] The physical information neural network is used to construct current constraints and heat conduction constraints based on the current, voltage, and ambient temperature in the input motor state sequence, and obtain the expression of the total physical constraints.

[0081] The adaptive physics constraint weight module is used to construct the total loss function based on the expression of the total physics constraints, which is used for model training.

[0082] The adaptive window adjustment mechanism module is used to adaptively adjust the input window of historical motor state data based on the determination coefficients of model training.

[0083] The present invention has the following advantages:

[0084] 1. Multi-dimensional feature extraction capability: Through a three-level structure of time attention, feature attention and physical constraint attention, the key information in time series data is extracted layer by layer, which effectively solves the characteristics of high coupling and strong nonlinearity of permanent magnet synchronous motor data and improves the model's ability to capture key information.

[0085] 2. Higher interpretability and physical meaning: By incorporating the physical constraints of the motor as a regularization term into the loss function, the identification results are ensured to conform to physical laws. Even under conditions of sparse data or high noise, the identification accuracy can still be maintained, while improving the interpretability and physical meaning of the model.

[0086] 3. Better adaptability and robustness: The time window size is dynamically adjusted based on residual statistics and evaluation indicators. It can reduce the window size to capture short-term changes under rapidly changing operating conditions and increase the window size to capture long-term dependencies under stable operating conditions, which significantly improves the model's adaptability and robustness to different operating conditions. Attached Figure Description

[0087] Figure 1 This is a diagram illustrating the overall architecture of the model method of this invention.

[0088] Figure 2 This is a flowchart of the method of the present invention;

[0089] Figure 3 This is a schematic diagram illustrating the changes in the attention layer weights in the model method;

[0090] Figure 4 This is a schematic diagram of the adaptive window adjustment mechanism in the model method. Detailed Implementation

[0091] The method of the present invention will be further described in detail below with reference to the accompanying drawings.

[0092] This invention relates to the field of CNC technology, specifically an adaptive window CMAWL-PINN motor parameter identification method based on a cascaded multi-attention mechanism. Permanent magnet synchronous motor systems (PMSMs) exhibit strong nonlinearity, high data coupling, time delay, and dynamic characteristics, making accurate identification difficult. To address these issues, a CMAWL-PINN system identification method was designed and invented. This method comprises a two-layer LSTM network, a cascaded multi-attention mechanism, a Physical Information Neural Network (PINN) framework, and an adaptive window adjustment mechanism based on residual statistics and evaluation indicators. It further utilizes the Optuna hyperparameter optimization framework to find the optimal hyperparameter combination. Experimental verification on a large dataset, as well as comparative experiments with traditional mainstream methods and current improved methods, demonstrates the invention's excellent performance and stronger robustness and adaptability under different operating conditions.

[0093] like Figure 1As shown, this invention acquires historical motor state data and preprocesses it to obtain a motor state sequence; it constructs a CMAWL-PINN model and trains the model based on the motor state sequence; based on the motor state sequence, it identifies the motor temperature of the motor system using the trained CMAWL-PINN model. The motor state includes motor temperature, current, voltage, speed, torque, coolant temperature, and ambient temperature. The predicted target is the motor temperature, including stator winding temperature, stator tooth temperature, stator yoke temperature, and permanent magnet surface temperature.

[0094] like Figure 2 As shown, the overall architecture of the method of the present invention mainly includes the following steps:

[0095] 1. Cascaded multi-attention mechanism

[0096] The model method employs a cascaded multi-attention mechanism, consisting of three cascaded attention layers: a temporal attention layer, a feature attention layer, and a physical constraint attention layer. For example... Figure 3 Taking one round of training as an example, the weight changes of the three attention layers in the cascaded multi-attention mechanism are shown.

[0097] The temporal attention layer aims to identify the most important time steps for target prediction within a time window, and its calculation process is as follows:

[0098]

[0099] Among them, g i This represents the attention score at the i-th time step. This is the output of the LSTM at time step i, where n is the vector dimension in the motor state sequence. and It is a learnable parameter, α i It is the attention weight at time step i. It is the weighted context vector;

[0100] The time attention layer can automatically identify the time steps that have the greatest impact on changes in the target, such as load mutation points and velocity change points, thereby improving the model's sensitivity to critical moments.

[0101] Among them, h i The output of the received LSTM layer; the LSTM layer extracts features from the input motor state sequence, handles long-term dependencies in long sequence data, and obtains the output sequence H of the hidden states. w is the current time window size, and 64 is the hidden layer dimension of the LSTM.

[0102] The feature attention layer receives the output of the temporal attention layer and further identifies the most important feature dimensions for target prediction. Its calculation process is as follows:

[0103] β=σ(W b c)c′=β⊙c

[0104] in, It is the output of the time attention layer. These are learnable parameters, and σ is the sigmoid activation function. These are the attention weights along the feature dimension. This is the weighted feature vector. The feature attention layer can automatically identify the feature dimensions that have the greatest impact on different targets, thereby improving the model's sensitivity to key features.

[0105] The physical constraint attention layer receives the output of the feature attention layer and further adjusts the feature representation according to the physical laws between feature dimensions. Its calculation process is as follows:

[0106] γ=tanh(W p c′)c″=c′⊙(1+γ)

[0107] in, It is the output of the feature attention layer. These are learnable parameters, and tanh is the hyperbolic tangent activation function. It is the attention weight of physical constraints. This is the final feature representation. The physically constrained attention layer uses the tanh activation function, which allows for positive and negative adjustment. This can enhance or suppress specific features, ensuring that the feature representation conforms to physical laws, thereby improving the physical consistency and generalization ability of the model.

[0108] 2. Physical Information Neural Network (PINN) Framework

[0109] The physical constraint design of the Physical Information Neural Network (PINN) framework mainly integrates physical knowledge at the output level. Using thermodynamic principles as an example, this module's function in a motor system is illustrated, primarily comprising current constraints and thermal conduction constraints.

[0110] On the one hand, the temperature rise of the motor is partly caused by the current, and large fluctuations or abnormalities in the current can significantly affect the change in motor temperature. Therefore, the current constraint is designed as follows:

[0111]

[0112] Among them, i d and i q These are the normalized d-axis and q-axis currents, respectively, and N is the batch size. On the other hand, according to basic thermodynamic principles, the rate of temperature change of the motor is proportional to the temperature difference. Therefore, the heat conduction constraints are designed as follows:

[0113]

[0114] Among them, T i It is the predicted temperature at the current moment, T i-1 Here, Δt is the predicted temperature from the previous moment, Δt is the time step, and k is the thermal conductivity coefficient. The expression for the total physical constraints is:

[0115] L physics =λ1L current +λ2L heat

[0116] λ1 and λ2 are weighting coefficients used to balance the effects of different physical constraints.

[0117] 3. Adaptive window adjustment mechanism

[0118] The adaptive window adjustment mechanism in the model method design falls into two categories: adjustment strategies when performance stagnates and adjustment strategies during normal training. Both strategies share the commonality of calculating the direction of change.

[0119]

[0120] The criterion for judging performance stagnation is the coefficient of determination R in the most recent training rounds. 2 The change is less than a preset threshold:

[0121]

[0122] When the model performance stagnates for multiple consecutive rounds, the system adopts a more aggressive window adjustment strategy. By adjusting the window size, the performance bottleneck can be broken and local optima can be avoided. The formula for calculating the window size change Δw is as follows:

[0123]

[0124] Among them, e i It is the mean residual of the current batch. It is the residual variance, θ is the threshold parameter, and w t This is the current window size.

[0125] During normal training, the window size is adjusted based on the continuous change in residual statistics:

[0126]

[0127] Where γ is the adjustment step size parameter. To avoid frequent small changes in the window size, a minimum adjustment mechanism is introduced, as shown in the following formula, where Δw min The general setting is 1:

[0128]

[0129] Ultimately, the window size update rule is as follows:

[0130]

[0131] Among them, w min and w max These are the lower and upper limits of the window size, respectively.

[0132] This adaptive window adjustment mechanism, combining residual statistics and performance monitoring, increases the window size when the residual mean is greater than a threshold multiple of the residual variance to capture longer-term dependencies; conversely, it decreases the window size when the residual mean is less than a threshold multiple of the residual variance to improve sensitivity to short-term changes. Throughout the training process, the window size is dynamically adjusted to adapt to the changing system characteristics of the motor under different operating conditions. Figure 4 This is a graph showing the window size, residual statistics, and changes over the first 50 training epochs. As the number of training epochs increases, the mean and variance of the residuals continuously decrease, while the window size changes continuously according to the prescribed adaptive adjustment mechanism to adapt to the changes in training. At the same time, the evaluation metrics maintain dynamic growth.

[0133] 4. Adaptive Physical Constraint Weighting Strategy

[0134] The total loss function expression for the model method is:

[0135] L total =L data +λ adaptive ·L physics

[0136] Among them, L data It is data loss (mean squared error), L physics It is the physical constraint loss, λ adaptive It is an adaptive weight, and its calculation formula is:

[0137]

[0138] Where λ0 is the initial weight, typically set to 1.0, epoch is the number of training iterations, and R0 is the weight R0. 2 It is the determination coefficient for the current batch, and α and β are adjustment parameters.

[0139] This adaptive weight strategy has the following characteristics: In the early stage of training, the identification accuracy is low and the physical constraint weights are set too large. At this time, the physical characteristics are fitted first, so that the model can quickly converge to the direction that conforms to the physical constraints and improve the model performance rapidly. When the identification accuracy is high, the physical constraint weights are gradually reduced. At this time, the physical consistency is satisfied, and the focus is on fitting the data characteristics to gradually optimize the model performance.

[0140] 5. Optuna hyperparameter optimization

[0141] In this embodiment, hyperparameters such as the number of fully connected layers, the dropout rate of the Dropout layer, the initial window size, and the learning rate are optimized.

[0142] In hyperparameter optimization problems, TPE (Transformation of Hyperparameters) is a Bayesian optimization-based method. Its core idea is to guide the search by modeling the probabilistic relationship between hyperparameters and the objective function. The mathematical expression of the TPE algorithm is as follows:

[0143]

[0144] Where x is a hyperparameter, y is the objective function value, and y is the hyperparameter. * It is a threshold, usually set as a certain quantile (e.g., the 25th percentile) of the observed target function value, where l(x) is the value when the target function value is relatively good (y). <y * The hyperparameter distribution of ), g(x) is the distribution when the objective function value is poor (y≥y). * The hyperparameter distribution of ).

[0145] The TPE algorithm selects the next set of hyperparameters by maximizing the following expected improvement (EI):

[0146]

[0147] The advantage of this algorithm lies in its ability to effectively utilize historical evaluation results, gradually narrowing the search space and improving search efficiency. Furthermore, the TPE algorithm exhibits good adaptability to mixed-type (continuous, discrete, categorical) hyperparameter spaces, making it highly suitable for hyperparameter optimization of deep learning models. The objective function for hyperparameter optimization uses the coefficient of determination (R²) as its key parameter. 2 ) Optimization objective:

[0148]

[0149] Where θ is the set of hyperparameters, f θ It is the target model configured using hyperparameter θ.

[0150] 6. Evaluation Indicators

[0151] To comprehensively evaluate the performance of the model method, and in conjunction with the adaptive window adjustment mechanism and adaptive physical constraint weight strategy, this invention employs five evaluation metrics: the coefficient of determination (R²). 2The evaluation metrics include relative accuracy (ACC), mean absolute error (MAE), root mean square error (RMSE), and bias. The coefficient of determination (CDO) is the primary evaluation metric, measuring the fit between the model's predicted and actual values; a value closer to 1 indicates a better fit. Relative accuracy, calculated from the relative error, better reflects prediction accuracy across different target feature ranges. Mean absolute error measures the average absolute deviation between predicted and actual values, is insensitive to outliers, and stably reflects the overall model performance. RMSE is more sensitive to larger errors and reflects the model's performance in extreme cases. Bias is calculated as the average difference between predicted and actual values, reflecting the systematic bias in the model's predictions; ideally, it should be close to zero. The formulas for these metrics are as follows, where y... i For the true value, For predicted values, is the mean of the true values, and m is the sample size.

[0152]

[0153] Finally, through experimental analysis and comparison, the proposed model method (CMAWL-PINN) demonstrates significant performance advantages over traditional machine learning models RFR, as well as mainstream neural network models RNN and GRU. CMAWL-PINN's bias is closer to 0; R... 2 The R² value was 0.9925, an average improvement of 7.23%; the ACC was 0.9758, an average improvement of 5.71%; the MAE was 1.2703, an average decrease of 61.46%; and the RMSE was 2.2071, an average decrease of 62.33%. Meanwhile, compared to some existing improved models, CMAWL-PINN's R² value was significantly higher than CRNN-DBO. 2 It's closer to 1 than ACC, R 2 It improved by 0.0367, ACC improved by 0.0182, and MAE and RMSE were smaller, with MAE decreasing by 1.2113 and RMSE decreasing by 1.8574; compared to E-ALIR, CMAWL-PINN's R... 2 It's also closer to 1 than ACC, R 2 The performance improved by 0.0048, ACC improved by 0.0094, MAE decreased by 0.5487, and RMSE decreased by 0.6582. Comprehensive analysis demonstrates that the proposed model method exhibits superior performance in motor system identification tasks.

[0154] In addition, this invention also conducted ablation experiments for comparison, with the comparison groups including: removing the adaptive window, removing the multi-attention mechanism, removing physical constraints, and the basic LSTM model. Compared to these ablation variants, CMAWL-PINN outperformed all evaluation metrics, effectively demonstrating the effectiveness and necessity of each component in CMAWL-PINN. The three key components—adaptive window, multi-attention mechanism, and physical constraints—work synergistically to improve the model's predictive performance.

[0155] Finally, this method has broad applicability. In other identification fields, only the physical constraint module needs to be modified accordingly. For example, for the identification problem of motion control, it can be replaced with kinematic physical constraints to complete the model adaptation.

Claims

1. A method for identifying motor systems based on CMAWL-PINN, characterized in that... This includes the following steps: Historical motor state data is acquired and preprocessed to obtain a motor state sequence; Construct the CMAWL-PINN model and train it based on the motor state sequence; Based on the motor state sequence, the motor temperature of the motor system is identified using the trained CMAWL-PINN model.

2. The motor system identification method based on CMAWL-PINN according to claim 1, characterized in that, The construction of the CMAWL-PINN model includes: LSTM is used to extract features from the input motor state sequence to obtain the input for the next attention mechanism. A cascaded multi-attention mechanism is used to obtain feature representations based on the output of the LSTM, thereby improving the model's ability to identify key factors in motor state changes. The entire connected hierarchy is used to represent the predicted motor temperature based on characteristics. The physical information neural network is used to construct current constraints and heat conduction constraints based on the current, voltage, and ambient temperature in the input motor state sequence, and obtain the expression of the total physical constraints. Adaptive physical constraint weights are used to construct the total loss function based on the expression of the total physical constraints, which is then used for model training. An adaptive window adjustment mechanism is used to adaptively adjust the input window of historical motor state data based on the decision coefficients trained on the model.

3. The motor system identification method based on CMAWL-PINN according to claim 1, characterized in that, The motor status includes motor temperature, current, voltage, speed, torque, coolant temperature, and ambient temperature.

4. The motor system identification method based on CMAWL-PINN according to claim 1, characterized in that, The motor temperature includes the stator winding temperature, stator tooth temperature, stator yoke temperature, and permanent magnet surface temperature.

5. The motor system identification method based on CMAWL-PINN according to claim 1, characterized in that, The cascaded multi-attention mechanism includes the following steps: The temporal attention layer is used to identify the feature dimensions that have the greatest impact on different prediction targets, and the process is as follows: Among them, g i This represents the attention score at the i-th time step. This is the output of the LSTM at time step i, where n is the vector dimension in the motor state sequence. and It is a learnable parameter, α i It is the attention weight at time step i. It is the weighted context vector; The feature attention layer is used to identify the most important feature dimensions for different prediction targets, and the process is as follows: β=σ(W b c) c′=β⊙c in, It is the output of the time attention layer. These are learnable parameters, and σ is the sigmoid activation function. These are the attention weights along the feature dimension. It is the weighted eigenvector; The physically constrained attention layer receives the output of the feature attention layer and further adjusts the feature representation according to the physical laws between feature dimensions. The process is as follows: γ=tanh(W p c′) c″=c′⊙(1+γ) in, It is the output of the feature attention layer. These are learnable parameters, and tanh is the hyperbolic tangent activation function. It is the attention weight of physical constraints. This is the final feature representation, where tanh represents the activation function and ⊙ represents the Hadamard product.

6. The motor system identification method based on CMAWL-PINN according to claim 1, characterized in that, The physical constraints of the physical information neural network framework include current constraints and heat conduction constraints: the current constraints are as follows: Among them, i d and i q These are the normalized d-axis and q-axis currents, respectively, and N is the batch size. The thermal conduction constraints are as follows: Among them, T i It is the predicted motor temperature T output by the fully connected layer at the current moment. i-1 It is the predicted temperature at the previous moment, Δt is the time step, and k is the thermal conductivity coefficient; The expression for the total physical constraints is: L physics =λ1L current +λ2L heat λ1 and λ2 are weighting coefficients used to balance the effects of different physical constraints.

7. The motor system identification method based on CMAWL-PINN according to claim 1, characterized in that, The adaptive physical constraint weighting strategy is as follows: The expression for the total loss function during model training is: L total =L data +λ adaptive ·L physics Among them, L data It is the data loss, i.e., the mean square error, L physics It is the physical constraint loss, λ adaptive It is an adaptive weight, and its calculation formula is: Where λ0 is the initial weights, epoch is the number of training iterations of the model, and R... 2 It is the determination coefficient for the current batch, and α and β are adjustment parameters.

8. The motor system identification method based on CMAWL-PINN according to claim 1, characterized in that, The adaptive window adjustment mechanism includes the following steps: 1) Calculate the direction of change: Among them, e i It is the mean residual of the current batch. It represents the residual variance, and θ is a parameter; a direction value of 1 indicates that the sequence size increases, and a direction value of -1 indicates that the sequence size decreases. 2) Based on the coefficient of determination R in the most recent rounds of training 2 The change in the value of the model is compared with a preset threshold to determine whether the model performance has stagnated, as follows: Where, ΔR 2 This indicates the change in the coefficient of determination. The coefficient of determination for the currently trained model. The coefficient of determination of the model from the previous training round; Where is_stagnation represents a boolean value, True indicates that the model performance has stagnated, and False indicates that the model performance has not stagnated and that the training is normal. 3) When model performance stagnates for multiple consecutive rounds, a window adjustment strategy is used to calculate the window change Δw: Among them, w t It is the current window size; When normal model training is detected, the window size is adjusted based on continuous changes in residual statistics: Where γ is the adjustment step size parameter; 4) Introduce the minimum adjustment mechanism as follows: Where, Δw min The threshold is epoch, which represents the number of training iterations of the model; epoch / 8 = 0 indicates that the epoch is a multiple of 8. Ultimately, the window size update rule is as follows: Among them, w t+1 w represents the window size at time t+1. min and w max These are the lower and upper limits of the window size, respectively.

9. A motor system identification device based on CMAWL-PINN, characterized in that... ,include: The data acquisition unit is used to acquire historical motor state data and perform preprocessing to obtain a motor state sequence; The model training unit is used to build the CMAWL-PINN model and train the model based on the motor state sequence. The state identification unit is used to identify the motor temperature of the motor system based on the motor state sequence using a trained CMAWL-PINN model.

10. A motor system identification device based on CMAWL-PINN according to claim 9, characterized in that, The CMAWL-PINN model includes: The LSTM layer is used to extract features from the input motor state sequence, which then becomes the input for the attention mechanism in the next step. A cascaded multi-attention mechanism module is used to obtain feature representations based on the output of LSTM, thereby improving the model's ability to identify key factors in motor state changes. The entire connected hierarchy is used to represent the predicted motor temperature based on characteristics. The physical information neural network is used to construct current constraints and heat conduction constraints based on the current, voltage, and ambient temperature in the input motor state sequence, and obtain the expression of the total physical constraints. The adaptive physics constraint weight module is used to construct the total loss function based on the expression of the total physics constraints, which is used for model training. The adaptive window adjustment mechanism module is used to adaptively adjust the input window of historical motor state data based on the determination coefficients of model training.

Images