Weak magnetic motor optimization method and system based on deep learning and storage medium

Abstract

The application relates to the technical field of weak magnetic motor optimization, and discloses a weak magnetic motor optimization method and system based on deep learning and a storage medium. The method comprises the following steps: collecting motor operation data under different rotating speeds, constructing a deep neural network model after denoising and normalization processing, intelligently adjusting direct-axis and quadrature-axis current reference values, dynamically calculating control parameters in combination with operation state data, and forming a closed-loop optimization control system. The application realizes high-performance and high-efficiency operation of the motor in a wide rotating speed range, can dynamically adjust control parameters in real time, and effectively balances multiple performance indexes such as system stability, response speed, energy efficiency and temperature rise.

Description

Technical Field

[0001] This application relates to the field of field weakening motor optimization technology, and in particular to a field weakening motor optimization method, system and storage medium based on deep learning. Background Technology

[0002] In modern industrial production, permanent magnet motors are widely used in various precision drive systems due to their high efficiency, compactness, and reliability, including in aerospace, new energy vehicles, machinery manufacturing, and industrial automation. Traditional permanent magnet motor control technology mainly relies on the classic PI (proportional-integral) control method, achieving field weakening control by adjusting stator current and rotor flux linkage to meet performance requirements under different speed and load conditions. Engineers have long been committed to improving the performance and energy efficiency of motors over a wide speed range.

[0003] However, traditional field weakening control methods have many inherent limitations. First, fixed-parameter PI controllers cannot effectively adapt to the complex dynamic characteristics of motors under different operating conditions, severely restricting control performance. Second, the lack of in-depth understanding and real-time modeling capabilities of the complex internal physical processes of the motor makes it difficult to precisely adjust the control strategy. Furthermore, existing methods struggle to simultaneously address multiple key performance indicators such as system stability, dynamic response, energy efficiency, and temperature rise, often requiring trade-offs and compromises among different metrics. Summary of the Invention

[0004] This application provides a deep learning-based optimization method, system, and storage medium for field weakening motors, which enables high-performance and high-efficiency operation of the motor over a wide speed range and allows for real-time dynamic adjustment of control parameters, effectively balancing multiple performance indicators such as system stability, response speed, energy efficiency, and temperature rise.

[0005] Firstly, this application provides a deep learning-based optimization method for a field-weakening motor. The method includes: collecting stator current, rotor position angle, voltage, and temperature data of a permanent magnet motor at different speeds to obtain raw field-weakening operation data; performing denoising and maximum / minimum value normalization on the raw field-weakening operation data to obtain standardized field-weakening control feature data; inputting the standardized field-weakening control feature data into a five-layer hidden-layer deep neural network for training to obtain a field-weakening control performance prediction model; calculating parameter adjustments for the direct-axis current reference value and quadrature-axis current reference value based on the output of the field-weakening control performance prediction model to obtain basic field-weakening control parameters; using the basic field-weakening control parameters combined with motor operating state data to calculate the real-time adjustment of the controller's proportional gain and integral gain to obtain dynamic control parameters; applying the dynamic control parameters to the motor's field-weakening control system and iteratively optimizing the field-weakening control performance prediction model through feedback data to form a closed-loop optimization control method for the field-weakening motor.

[0006] Secondly, this application provides a deep learning-based field-weakening motor optimization system, the deep learning-based field-weakening motor optimization system comprising:

[0007] The data acquisition module is used to collect stator current, rotor position angle, voltage and temperature data of permanent magnet motor at different speeds to obtain raw data of field weakening operation.

[0008] The processing module is used to denoise and normalize the maximum and minimum values ​​of the raw data of the field weakening operation to obtain standardized field weakening control feature data.

[0009] The training module is used to input the standardized field weakening control feature data into a five-layer hidden deep neural network for training, so as to obtain a field weakening control performance prediction model.

[0010] The adjustment module is used to perform parameter adjustment calculations on the direct-axis current reference value and the quadrature-axis current reference value based on the output results of the field weakening control performance prediction model, so as to obtain the basic parameters of field weakening control.

[0011] The calculation module is used to calculate the real-time adjustment of the controller's proportional gain and integral gain by combining the field weakening control basic parameters with the motor operating status data, thereby obtaining dynamic control parameters.

[0012] An optimization module is used to apply the dynamic control parameters to the motor field weakening control system and to iteratively optimize the field weakening control performance prediction model through feedback data, thereby forming a closed-loop optimization control method for the field weakening motor.

[0013] Thirdly, a computer device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the computer device to execute the aforementioned deep learning-based magnetic field weakening motor optimization method.

[0014] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the aforementioned deep learning-based method for optimizing a weak magnetic motor.

[0015] The technical solution provided in this application employs a high-frequency (10kHz) data acquisition strategy to comprehensively capture the multi-parameter features of motor operation, providing a rich and accurate data foundation for the deep learning model. The five-layer hidden layer structure of the deep neural network, through a progressively decreasing number of neurons, achieves multi-scale feature extraction of motor operation data, effectively capturing the complex nonlinear dynamic characteristics of the motor. Artificial intelligence algorithms play a crucial role in the solution, particularly the introduction of deep learning and reinforcement learning models, which breaks through the static parameter limitations of traditional control methods, constructing an intelligent control system capable of autonomous learning and dynamic optimization. The multi-objective optimization algorithm, through precisely designed weight coefficients (0.3 for stability margin and efficiency, and 0.2 for response time and temperature rise), achieves a multi-dimensional balance of system performance. The introduction of the Covariance Matrix Adaptive Evolutionary Strategy (CMA-ES) enables the parameter optimization process to intelligently search in a complex high-dimensional parameter space, effectively avoiding getting trapped in local optima. The design of the deep reinforcement learning model further elevates the adaptive adjustment of the controller to a new level, constructing an intelligent system capable of real-time learning and adjustment of the control strategy by defining a finely defined state space and action space. The attention mechanism, combined with a state prediction model based on a Long Short-Term Memory (LSTM) network, further enhances the system's predictive control capabilities, enabling early identification of potential unstable operating conditions. This significantly improves motor performance, achieving high-performance and high-efficiency operation over a wide speed range. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of one embodiment of the deep learning-based field weakening motor optimization method in this application.

[0018] Figure 2 This is a schematic diagram of one embodiment of the field weakening motor optimization system based on deep learning in this application.

[0019] Figure 3 This is a schematic block diagram of the structure of the computer device in an embodiment of the present invention. Detailed Implementation

[0020] This application provides a method, system, and storage medium for optimizing a field-weakening motor based on deep learning. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the deep learning-based field-weakening motor optimization method in this application includes:

[0022] Step S101: Collect stator current, rotor position angle, voltage and temperature data of permanent magnet motor at different speeds to obtain raw data of field weakening operation;

[0023] Step S102: Denoise and normalize the maximum and minimum values ​​of the raw data of the field weakening operation to obtain standardized field weakening control characteristic data.

[0024] Step S103: Input the standardized field weakening control feature data into a five-layer hidden deep neural network for training to obtain a field weakening control performance prediction model.

[0025] Step S104: Based on the output results of the field weakening control performance prediction model, perform parameter adjustment calculations on the direct-axis current reference value and the quadrature-axis current reference value to obtain the basic parameters of field weakening control.

[0026] Step S105: Using the basic parameters of field weakening control combined with the motor operating status data, calculate the real-time adjustment of the controller's proportional gain and integral gain to obtain the dynamic control parameters.

[0027] Step S106: Apply dynamic control parameters to the field weakening control system of the motor, and cyclically optimize the field weakening control performance prediction model through feedback data to form a closed-loop optimization control method for the field weakening motor.

[0028] It is understood that the executing entity of this application can be a deep learning-based field-weakening motor optimization system, or it can be a terminal or a server; the specific implementation is not limited here. This application's embodiments use a server as an example for illustration.

[0029] Specifically, test points were set within a range of 0.2 to 3 times the rated speed of the motor to collect parameters such as stator current, rotor position angle, voltage, and temperature. These data were categorized into three regions: constant torque, field weakening, and constant power, with specific indicators added to each. For the field weakening region data, a field weakening depth indicator was added, reflecting the degree to which the motor entered field weakening control; for the constant power region data, a power fluctuation indicator was added, characterizing the motor's power stability in the high-speed region. Simultaneously, by calculating the stable operating boundaries of each working region, the critical points and unstable operating regions of the field weakening region were determined, thereby filtering out data that met the stable operating conditions. The collected raw field weakening operation data needed to be denoised and normalized to obtain standardized field weakening control characteristic data. During denoising, outlier detection was first performed, using a three-standard-deviation principle to identify abnormal data points, which were then replaced with the average of adjacent normal data points. Then, a mean-value filtering algorithm was applied to preliminarily clean the data, selecting a seven-point window length to filter each signal channel separately, effectively removing sudden interference. Next, wavelet transform decomposition technology was used, selecting five-level de Beciétie wavelets for multi-scale analysis to extract signal features and reconstruct them, resulting in finely denoised data. Normalization involved calculating the maximum and minimum values ​​of each parameter, establishing a normalization transformation parameter table, and adjusting the numerical range to the interval between 0 and 1, which improves the training effect of the neural network.

[0030] Standardized field weakening control feature data were input into a five-layer hidden-layer deep neural network for training to construct a field weakening control performance prediction model. In this network structure, the number of neurons in the five hidden layers are 256, 192, 128, 96, and 64, forming a funnel-shaped structure that facilitates feature extraction and dimensionality reduction. A funnel-shaped decreasing activation function with a parameter of 0.01 was configured in the hidden layers to avoid the gradient vanishing problem; a linear activation function was set in the output layer to adapt to the continuous numerical characteristics of the performance indicators. During training, a weighted mean square error loss function was used, assigning weights of 0.3, 0.2, 0.3, and 0.2 to the four indicators of system stability margin, dynamic response time, efficiency, and temperature rise, respectively. An adaptive moment estimation algorithm with a learning rate of 0.001 was selected as the optimizer, with a batch size of 64 and a training cycle of 300. Based on the output results of the field weakening control performance prediction model, parameter adjustment calculations were performed on the direct-axis current reference values ​​and the quadrature-axis current reference values ​​to obtain the basic parameters of field weakening control. This process first inputs the motor's current speed, load torque, and voltage constraints into the prediction model to obtain predicted values ​​for four performance indicators. Then, a multi-objective optimization function is set, with maximizing stability margin, minimizing response time, maximizing efficiency, and minimizing temperature rise as optimization objectives, transforming it into a weighted single-objective function. An improved covariance matrix adaptive evolutionary strategy algorithm is used for parameter optimization, setting the population size to 50, the initial step size to 0.3, and the number of parents to 25, iterating 500 times to obtain the Pareto optimal solution set. The optimal configuration of the direct-axis current reference value and the quadrature-axis current reference value is extracted from this set to establish a current reference value calculation model.

[0031] By combining fundamental field weakening control parameters with motor operating state data, the real-time adjustments of the controller's proportional gain and integral gain are calculated to obtain dynamic control parameters. This step first defines a state space containing seven key variables and an action space containing four variables, constructing a state-action mapping relationship and a reward function. Based on this, a deep Q-network with a dual-network structure is constructed, with both networks employing four fully connected neural networks. An update mechanism is set up: the evaluation network updates once per step, and the target network updates once every 100 steps. An exploration strategy is designed with an initial exploration rate of 1.0, decaying to a minimum of 0.01 per step, a decay rate of 0.995, and a discount factor of 0.95. After training for 500,000 steps in a motor simulation environment, a Q-value network capable of real-time evaluation of the current motor state and outputting the optimal control action is obtained.

[0032] Dynamic control parameters are applied to the field weakening control system of a motor. The performance prediction model of the field weakening control is cyclically optimized using feedback data, forming a closed-loop optimization control method for the field weakening motor. Specifically, the dynamic control parameters and the basic field weakening control parameters are weighted and averaged to obtain comprehensive control parameters, which are then configured into the control system. The system's operating status is monitored in real time at a sampling frequency of 10kHz, and the error between the actual performance index value and the predicted value is calculated. A data buffer pool stores the latest 1000 sets of data, which are merged with the original training data for incremental training of the model. The learning rate is 0.0005, and the iteration period is 50. Taking a certain type of permanent magnet motor as an example, at 1.5 times the rated speed of 2000rpm, the collected direct-axis current is 10A, the quadrature-axis current is 15A, the direct-axis voltage is 120V, and the quadrature-axis voltage is 200V. After these data are filtered through a seven-point window, noise is effectively suppressed; and through maximum-minimum normalization, the values ​​are mapped to the range of 0 to 1. The processed data is input into a trained five-layer deep neural network, which predicts a system stability margin of 0.25, a dynamic response time of 0.08s, an efficiency of 93%, and a temperature rise of 40℃. Based on these predictions, the direct-axis current reference value is corrected to -12A using a covariance matrix adaptive evolution strategy algorithm, while the quadrature-axis current reference value is maintained at 15A. The deep Q-network further calculates a proportional gain adjustment of +0.5 and an integral gain adjustment of -0.2. After these parameters are applied to the control system, through continuous collection of feedback data and incremental model training, the closed-loop optimization system continuously adapts to changes in the motor's operating state, maintaining optimal control performance.

[0033] In this embodiment, a high-frequency (10kHz) data acquisition strategy is employed to comprehensively capture the multi-parameter features of motor operation, providing a rich and accurate data foundation for the deep learning model. The five-layer hidden layer structure of the deep neural network, through a progressively decreasing number of neurons, enables multi-scale feature extraction of the motor operation data, effectively capturing the complex nonlinear dynamic characteristics of the motor. Artificial intelligence algorithms play a crucial role in the solution, particularly the introduction of deep learning and reinforcement learning models, which overcomes the static parameter limitations of traditional control methods, constructing an intelligent control system capable of autonomous learning and dynamic optimization.

[0034] The multi-objective optimization algorithm achieves a multi-dimensional balance of system performance through precisely designed weight coefficients (0.3 each for stability margin and efficiency, and 0.2 each for response time and temperature rise). The introduction of the Covariance Matrix Adaptive Evolutionary Strategy (CMA-ES) enables the parameter optimization process to intelligently search within a complex high-dimensional parameter space, effectively avoiding getting trapped in local optima. The design of the deep reinforcement learning model elevates the controller's adaptive adjustment to a new level, constructing an intelligent system capable of real-time learning and adjustment of the control strategy by defining a finely defined state and action space. The attention mechanism combined with a Long Short-Term Memory (LSTM) network state prediction model further enhances the system's predictive control capability, enabling early identification of potential unstable operating conditions. This significantly improves the motor's performance indicators, achieving high-performance and high-efficiency operation over a wide speed range.

[0035] In one specific embodiment, the process of performing step S101 may specifically include the following steps:

[0036] Multiple test points were set up within a range of 0.2 to 3 times the rated speed of the motor. Data including stator current, rotor position angle, motor speed, stator voltage, field weakening current component, torque current component, direct axis current, quadrature axis current, power factor and efficiency index were collected at each test point to obtain raw data for multiple operating conditions.

[0037] The raw data under multiple operating conditions are classified according to the motor's operating region, into three categories: constant torque region, field weakening region, and constant power region, resulting in regional datasets.

[0038] The original characteristics of the data in the constant torque region are preserved, the field weakening depth index is added to the data in the field weakening region, and the power fluctuation index is added to the data in the constant power region to obtain an enhanced feature dataset.

[0039] Based on the enhanced feature dataset, the stable working boundary of each working region is calculated, the critical point and unstable operating region of the weak magnetic region are determined, and the operating boundary data of the weak magnetic region is obtained.

[0040] The raw collected data was filtered using the boundary data of the weak magnetic field region to remove abnormal data points that did not meet the stable operating conditions, thus obtaining preliminary purified data.

[0041] A sliding window mean filtering algorithm was applied to the preliminary purified data, with a window length of 5 sampling points, to eliminate high-frequency noise interference and obtain denoised data.

[0042] The denoised data is normalized based on the maximum and minimum values ​​of each parameter. The maximum and minimum value normalization transformation is performed on each parameter to obtain normalized data.

[0043] The normalized data was randomly split in an 8:2 ratio, with 80% of the data used as training data and 20% used as validation data, forming the original data for weak magnetic field operation.

[0044] Specifically, multiple test points are set up within a range of 0.2 to 3 times the rated speed of the permanent magnet motor. This setup ensures that the data covers various operating states of the motor from low to high speeds. For each test point, the acquisition system records basic electrical parameters such as stator current, rotor position angle, motor speed, and stator voltage, as well as field weakening current and torque current components. The field weakening current component refers to the current used to weaken the magnetic field, typically corresponding to the negative value of the direct-axis current; the torque current component is the current that generates electromagnetic torque, typically corresponding to the quadrature-axis current. The direct-axis and quadrature-axis currents are current components expressed in a rotating coordinate system, converted from three-phase current through coordinate transformation. In addition, power factor and efficiency indicators are also recorded simultaneously, forming comprehensive raw data for multiple operating conditions.

[0045] The collected raw data under various operating conditions needs to be categorized and organized according to the motor's operating region. The operating regions of permanent magnet motors are mainly divided into three categories: constant torque region, field weakening region, and constant power region. The constant torque region refers to the area where the motor operates below its rated speed, at which point it can output a constant maximum torque. The field weakening region refers to the area where, after the motor speed exceeds the rated speed, field weakening control is needed to reduce the magnetic field of the permanent magnet to maintain voltage balance. The constant power region is the area where the output power of the motor tends to be constant during high-speed operation. By dividing the data using speed thresholds and voltage utilization parameters, the raw data is classified into these three regions, forming regional datasets.

[0046] For data from different operating regions, targeted feature enhancement processing is required. For data in the constant torque region, since the control strategy is relatively simple, its original characteristics can be maintained. For data in the field weakening region, a field weakening depth index is added, which is obtained by calculating the ratio of the direct-axis current to the rated current, characterizing the degree of field weakening in the motor. For data in the constant power region, a power fluctuation index is added, which is obtained by calculating the ratio of the standard deviation of the power to the average value, characterizing the stability of the motor power. This feature enhancement makes the data better reflect the control characteristics of each region, forming an enhanced feature dataset.

[0047] Based on the enhanced feature dataset, it is necessary to calculate the stable operating boundaries of each working region. The calculation of the stable operating boundaries mainly relies on the motor's voltage and torque equations. By solving these equations, the current range within which the motor can operate stably at different speeds is determined. Particularly for the field weakening region, it is necessary to determine the critical point for field weakening control, i.e., the speed at which field weakening control should be implemented; simultaneously, it is also necessary to identify unstable operating regions, such as the conditions prone to loss of synchronization in deep field weakening regions. These calculation results form the operating boundary data for the field weakening region.

[0048] The calculated weak magnetic field region operating boundary data is used to filter and purify the original collected data. This primarily involves removing data points falling within unstable operating regions and outliers that significantly deviate from normal operating boundaries. This filtering process can be achieved by setting thresholds, such as marking data points exceeding a certain percentage of stable operating boundaries as outliers and removing them from the dataset. This process makes the data more reliable, forming preliminary purified data.

[0049] After initial data purification, a sliding window mean filtering algorithm is applied for further denoising. Sliding window mean filtering is a commonly used digital signal processing method. Its principle is to take the average of the current point and a total of five points (the length of the window before and after it) as the filtered result. Setting the window length to 5 sampling points means that each processed data point is calculated as the average of the current point and two points before and after it (a total of 5 points). This method is simple and effective, smoothing out random fluctuations in the data, eliminating high-frequency noise interference, and obtaining smoother denoised data.

[0050] Max-min normalization is a standard preprocessing step in deep learning, involving normalizing denoised data. For each parameter, its maximum and minimum values ​​are found across the entire dataset, and then the original values ​​are linearly mapped to the [0,1] interval. This normalization helps eliminate the influence of different parameter units, making neural network training more stable and efficient. The normalized data obtained after processing has all feature values ​​constrained to the same range. The normalized data is then randomly split in an 8:2 ratio, with 80% allocated as training data for training the deep neural network and 20% as validation data for validating model performance and tuning hyperparameters. This split is a common practice in machine learning, balancing the sufficiency of the training set with the representativeness of the validation set.

[0051] In one specific embodiment, the process of performing step S102 may specifically include the following steps:

[0052] Outlier detection was performed on the raw data of weak magnetic field operation. Data points that exceeded the normal range were identified using the three-standard-deviation principle, and outlier-marked data were obtained.

[0053] The raw data of weak magnetic field operation is filtered based on the outlier marking data. Data points marked as outliers are replaced with the average value of adjacent normal data points to obtain preliminary cleaned data.

[0054] The preliminary cleaned data was processed using a median filtering algorithm. A seven-point window was selected to filter each signal channel separately to obtain noise-suppressed data.

[0055] Wavelet transform decomposition was applied to the noise-suppressed data, and five-level de Becié wavelets were selected for multi-scale analysis to extract signal features and reconstruct them, resulting in finely denoised data.

[0056] The maximum and minimum values ​​of each parameter are calculated based on the finely denoised data, and a normalized transformation parameter table is established to obtain the normalized baseline data.

[0057] By using normalized baseline data, a maximum-minimum normalization transformation is performed on each parameter in the finely denoised data to adjust the numerical range to the interval between zero and one, thus obtaining normalized feature data.

[0058] The normalized feature data is balanced by performing data balancing processing, and the data in the weak magnetic field area is sampled in a key manner to ensure that the data proportion of each working area is balanced, thus obtaining balanced feature data.

[0059] The equilibrium characteristic data is divided into multiple data segments according to the time series. Each segment contains state parameters at twenty consecutive time points, forming standardized weak magnetic control characteristic data.

[0060] Specifically, outlier detection is first performed on the raw data of field weakening operation using the three-standard-deviation principle, a commonly used statistical method for anomaly detection. For each parameter's data sequence, its mean and standard deviation are calculated first, and data points deviating from the mean by more than three standard deviations are marked as outliers. The three-standard-deviation principle is based on the normal distribution theory. In a normal distribution, the probability of data falling within the range of plus or minus three standard deviations of the mean is approximately 99.7%, therefore, data points exceeding this range are likely outliers. Using this method, outliers in the stator current, rotor position angle, voltage, and temperature data collected from the motor at various speeds are marked, forming outlier-marked data. Based on the outlier-marked data, the raw field weakening operation data is filtered. For data points marked as outliers, they are not simply deleted, but replaced with the average of adjacent normal data points. This processing method considers the continuous characteristics of time series data; by taking the average of several normal points before and after the outlier as a replacement value, it can maintain the smoothness of the data without destroying the overall distribution characteristics of the data. The replaced dataset is called the initial cleaned data. This dataset has had obvious outliers removed, but may still contain a lot of random noise.

[0061] Further noise suppression processing is needed for the initially cleaned data, and the median filtering algorithm is used here. Median filtering is a non-linear filtering method. For each data point, the median value of all points within its window, sorted by size, is taken as the filtering result. Choosing a seven-point window length means that for each point in the time series, the median value of that point and the three points before and after it (a total of seven points) is used to replace the original value. Median filtering is particularly suitable for removing impulse noise and spike interference from motor operating data, while also preserving the edge characteristics of the data well, unlike mean filtering which can cause blurring of data edges. Median filtering is applied to each signal channel (such as stator current, rotor position angle, etc.) to obtain data with good noise suppression performance.

[0062] Further denoising is performed on the noise-suppressed data using wavelet transform decomposition. Wavelet transform is a time-frequency analysis method that provides the local time-domain characteristics of a signal at different frequencies. A five-level de Beciétie wavelet (usually denoted as the DB wavelet) is selected for multi-scale analysis. The de Beciétie wavelet is a wavelet basis function with good orthogonality and tight support, making it particularly suitable for analyzing the non-stationary characteristics of motor operating data. Five-level decomposition means decomposing the signal into wavelet coefficients at five different frequency scales, including one approximation coefficient and five detail coefficients. The wavelet transform decomposition can be expressed as:

[0063]

[0064] Where S(t) represents the original signal, A5(t) represents the fifth-level approximation coefficient, and D... j (t) represents the detail coefficients of the j-th layer. During denoising, thresholding is applied to the detail coefficients of each layer to retain key feature information and suppress noise influence. The signal is then reconstructed to obtain finely denoised data. Based on the finely denoised data, the maximum and minimum values ​​of each parameter are calculated, and a normalization transformation parameter table is established. The normalization transformation parameter table includes the name of each parameter, its original maximum and minimum values, and the corresponding normalization coefficient. These parameters will be used for the normalization processing of all subsequent data to ensure data processing consistency. The normalized baseline data is the basis for subsequent normalization operations and an important reference for data preprocessing during model deployment.

[0065] Using normalized baseline data, a minimum-maximum normalization transformation is performed on each parameter in the finely denoised data. Minimum-maximum normalization is a linear transformation that maps the original data linearly to the [0,1] interval. Through this transformation, parameters of different dimensions and orders of magnitude are unified to the same range, which is beneficial for the training process of the neural network. The normalized data is called normalized feature data.

[0066] Data balancing of normalized feature data primarily addresses the imbalance in the number of samples across different operating regions within the dataset. In actual motor operation, data from the field weakening region is often less than that from the constant torque region, leading to insufficient prediction accuracy for the trained model in the field weakening region. By focusing on sampling data from the field weakening region and employing oversampling techniques such as SMOTE (Synthetic Minority Over-sampling Technique) to generate new sample points for the field weakening region, the data proportion across all operating regions is balanced. This data balancing process results in balanced feature data. The balanced feature data is then divided into multiple data segments according to the time series, with each segment containing state parameters at twenty consecutive time points. This processing method considers the temporal relationships in motor control, enabling the model to learn the dynamic characteristics of parameter changes over time. Each data segment includes both the current state and the evolution of previous states, aiding in the model's accurate prediction of the motor's dynamic response. The data obtained in this way is called standardized field weakening control feature data, which serves as input for subsequent deep learning model training.

[0067] For example, during the optimization of field weakening control for a certain type of permanent magnet synchronous motor, raw data containing 10 parameters, including stator current, rotor position angle, direct-axis current, and quadrature-axis current, were collected from the test platform. Anomaly detection of the direct-axis current data was performed using the three-standard-deviation principle. The calculated mean of the direct-axis current data was -32A, and the standard deviation was 4.5A. Therefore, data points exceeding the range [-45.5A, -18.5A] were marked as anomalies. For a data point marked as anomaly (-55A), the average of the five normal points before and after it (-34A) was used as the replacement. Median filtering with a window length of 7 was applied to the replaced data, transforming the [...,-33A,-32A,-34A,-52A,-36A,-35A,-34A,...] data into [...,-33A,-32A,-34A,-36A,-35A,-34A,...] data, effectively removing sudden interference. The data was then subjected to a five-level de Beciétius wavelet decomposition to separate the main features and noise components, resulting in a smoother signal after reconstruction. After maximum-minimum normalization, the range of the direct-axis current was mapped from [-60A, 0A] to the [0, 1] interval. The processed data was then organized into data segments containing 20 consecutive time points according to the time series, forming standardized magnetic weakening control feature data for deep learning model training.

[0068] In one specific embodiment, the process of executing step S103 may specifically include the following steps:

[0069] A deep neural network structure consisting of an input layer, five hidden layers, and an output layer is constructed. The number of neurons in the input layer is the same as the feature dimension. The number of neurons in the five hidden layers are 256, 192, 128, 96, and 64, respectively. The number of neurons in the output layer is 4, thus obtaining the basic neural network architecture.

[0070] A funnel-shaped structure was set for the hidden layers of the neural network infrastructure, and a batch normalization layer and a dropout layer were added after each layer. The dropout rate was set to 20%, resulting in an anti-overfitting network structure.

[0071] Funnel-shaped decreasing activation functions are configured for the hidden layer neurons in the anti-overfitting network structure. A funnel-shaped decreasing activation function with a parameter of 0.01 is used to obtain nonlinear mapping units.

[0072] Based on the feature distribution of standardized field weakening control feature data, a linear activation function is set for the output layer, and a weighted mean square error is defined as the loss function to obtain the training objective function;

[0073] Weights are assigned to each performance metric in the training objective function: system stability margin is weighted at 0.3, dynamic response time at 0.2, efficiency at 0.3, and temperature rise at 0.2, resulting in a weighted training objective.

[0074] Based on the weighted training objective, an adaptive moment estimation algorithm with a learning rate of 0.001 was selected as the optimizer, and the batch size was set to 64 to obtain the network training strategy.

[0075] The neural network was trained using a network training strategy, with a training cycle of 300. An early stopping strategy was adopted, and training was stopped when the validation set loss did not improve for 15 consecutive cycles, thus obtaining the initial training model.

[0076] The initial trained model is tested for performance, the prediction error on the test set is calculated, and the network structure is fine-tuned through Bayesian hyperparameter optimization to form a prediction model for weak magnetic control performance.

[0077] Specifically, a deep neural network structure is constructed, comprising an input layer, five hidden layers, and an output layer. The number of neurons in the input layer directly corresponds to the dimensionality of the standardized field weakening control feature data. For field weakening control of a permanent magnet motor, the input features include multiple parameters such as stator current, rotor position angle, motor speed, stator voltage, field weakening current component, and torque current component. Typically, the feature dimension after data preprocessing is between 30 and 50. The five hidden layers use a gradually decreasing number of neurons: 256, 192, 128, 96, and 64, forming a structure that gradually narrows from input to output. The output layer is designed with four neurons, corresponding to four key performance indicators: system stability margin, dynamic response time, efficiency, and temperature rise. This hierarchical and rationally structured network architecture forms the basis for subsequent optimizations.

[0078] A funnel-shaped structure was used for the hidden layers of the neural network architecture. This structure refers to a network design where the number of neurons gradually decreases from the input layer to the output layer. The advantage of the funnel-shaped structure is that the earlier layers can extract rich features from the data, while the later layers focus on filtering and fusing these features, achieving effective dimensionality reduction and abstraction. Batch normalization and dropout layers were added after each hidden layer. Batch normalization standardizes each batch of data, keeping the data distribution within the sensitive region of the activation function and accelerating the network training process. Dropout layers prevent the network from overfitting to the training data by randomly shutting down a certain percentage of neurons. A dropout rate of 20% means that 20% of the neurons are randomly shut down during training, while all neurons are used during testing. This significantly improves the model's generalization ability and avoids over-reliance on the training data.

[0079] A funnel-shaped decreasing activation function was configured for the hidden layer neurons in the anti-overfitting network structure. This funnel-shaped decreasing activation function refers to the Leaky ReLU (Leaked Corrected Linear Unit) function, where the leakage parameter is set to 0.01. The mathematical expression of the Leaky ReLU function is f(x) = max(0.01x, x). Compared with the standard ReLU function, Leaky ReLU still has a small slope when the input is negative, thus preventing the gradient from completely vanishing during backpropagation, which helps solve the gradient vanishing problem in deep network training. The parameter of 0.01 means that the slope of the function is 0.01 on the negative half-axis. This is an empirical setting that effectively prevents neurons from "dying" (i.e., some neurons are never activated) while maintaining the non-linearity of the activation function.

[0080] Based on the characteristic distribution of standardized magnetic weakening control feature data, a linear activation function was set for the output layer instead of the commonly used nonlinear activation function. The linear activation function directly outputs a weighted sum of neurons, suitable for predicting continuous numerical outputs, such as system stability margin and dynamic response time. Simultaneously, a weighted mean squared error (MSE) was defined as the loss function. MSE is an improved version of the standard mean squared error (SMSE), considering the importance weights of different output metrics. While the standard mean squared error calculates the sum of squares of the differences between predicted and true values, the weighted mean squared error assigns different weight coefficients to the errors of different metrics, thus focusing more on the prediction accuracy of important metrics during training.

[0081] Different weights were assigned to the four performance metrics in the training objective function: system stability margin (0.3), dynamic response time (0.2), efficiency (0.3), and temperature rise (0.2). These weights reflect the relative importance of different performance metrics in field weakening control. System stability margin and efficiency were given higher weights (0.3), indicating that these two metrics have a greater impact on the quality of field weakening control; while dynamic response time and temperature rise are relatively less important, with a weight of 0.2. Through this differentiated weighting, the network training process focuses more on improving the prediction accuracy of key metrics.

[0082] Based on the weighted training objective, the Adaptive Moments Estimation (Adam) algorithm with a learning rate of 0.001 was selected as the optimizer. The Adam algorithm is an optimization algorithm that combines the advantages of momentum and RMSProp, adaptively adjusting the learning rate of each parameter and exhibiting good adaptability to gradients at different scales. Setting the learning rate to 0.001 is a conservative choice, ensuring stable training and a moderate convergence speed. A batch size of 64 was set, meaning that 64 samples are used for each gradient update, achieving a good balance between training efficiency and memory usage.

[0083] The neural network was trained using the aforementioned training strategy, with 300 training epochs and an early stopping strategy. Early stopping is a common technique to prevent overfitting; it monitors performance metrics on the validation set and terminates training early when performance no longer improves. Specifically, training stops when the validation set loss fails to improve for 15 consecutive epochs. This ensures sufficient model training while avoiding overtraining that could lead to a decline in generalization ability. The model obtained after this initial training is called the initial training model.

[0084] The initial trained model is tested for performance, and the prediction error on the test set is calculated to evaluate the model's generalization ability. Then, the network structure is fine-tuned using Bayesian hyperparameter optimization. Bayesian optimization is a black-box optimization method based on a probabilistic model, capable of efficiently searching for optimal configurations in a complex parameter space. Hyperparameter optimization targets include the number of network layers, the number of neurons in each layer, the dropout rate, and the learning rate. By continuously trying different combinations of hyperparameters, the configuration that performs best on the validation set is found. Through this series of construction, training, and optimization processes, a deep neural network model capable of accurately predicting the performance of magnetic weakening control is finally formed.

[0085] Taking the field weakening control optimization of a certain type of permanent magnet synchronous motor as an example, the constructed deep neural network input layer contains 42 neurons, corresponding to the preprocessed feature dimensions. The features include sequential data at 20 time points, such as motor speed, stator current, direct-axis current, and quadrature-axis current. During network training, the training data is first input into the network in batches of 64 samples. Forward propagation calculates the predicted values, and then the weighted mean square error loss is calculated. For example, for a certain sample, the network predicts a system stability margin of 0.75 (true value 0.82), a dynamic response time of 0.15 seconds (true value 0.12 seconds), an efficiency of 92% (true value 95%), and a temperature rise of 38℃ (true value 35℃). The weighted mean square error is calculated based on the set weights. During backpropagation, the Adam optimizer is used to update the network parameters with a learning rate of 0.001. The validation set loss is monitored during training. It is found that after the 178th cycle, the validation set loss has not decreased for 15 consecutive cycles, triggering an early stop mechanism to stop training. Finally, through Bayesian optimization, it was found that adjusting the number of neurons in the fourth hidden layer from 96 to 112 and the dropout rate from 20% to 15% improved the model's performance on the test set. This configuration was ultimately determined as the final structure of the weak magnetic control performance prediction model.

[0086] In one specific embodiment, the process of executing step S104 may specifically include the following steps:

[0087] Input the current motor speed, load torque and voltage limit conditions into the field weakening control performance prediction model to obtain the predicted values ​​of four performance indicators: system stability margin, dynamic response time, efficiency and temperature rise, and obtain the operating condition performance characteristic data.

[0088] A multi-objective optimization function is set for the operating condition performance characteristic data, with the optimization objectives being to maximize stability margin, minimize response time, maximize efficiency, and minimize temperature rise, resulting in the objective function expression.

[0089] The objective function expression is converted into a weighted single objective function, and the weight coefficients for the four indicators are set to 0.3, 0.2, 0.3 and 0.2 respectively, to obtain the comprehensive optimization objective.

[0090] An improved covariance matrix adaptive evolution strategy algorithm is constructed based on the comprehensive optimization objective. The population size is set to 50, the initial step size is 0.3, and the number of parents is 25, resulting in a parameter optimization algorithm.

[0091] The direct-axis current reference value and the quadrature-axis current reference value were iteratively optimized using a parameter optimization algorithm. The number of iterations was set to 500 to obtain the Pareto optimal solution set.

[0092] The Pareto optimal solution set is analyzed to extract the optimal configuration of the direct-axis current reference value and the quadrature-axis current reference value, thus obtaining the current reference value configuration scheme.

[0093] Based on the current reference value configuration scheme, establish the direct-axis current reference value correction formula and the quadrature-axis current reference value correction formula, which are respectively expressed as functions of direct-axis current and rotational speed and functions of quadrature-axis current and torque demand, to obtain the current reference value calculation model;

[0094] The current reference value calculation model was verified and tested. The current reference value was calculated under different operating conditions, and the boundary was corrected based on the verification results to form the basic parameters for field weakening control.

[0095] Specifically, using key motor operating parameters as input sources, a sophisticated performance mapping model is constructed through a deep neural network. Input parameters include the motor's current speed, load torque, and voltage limits. These parameters undergo complex nonlinear feature extraction and mapping to transform them into predicted values ​​for four core performance indicators. The first stage of data processing involves the feature extraction mechanism of the deep neural network. The network employs a five-layer hidden structure, each layer equipped with a funnel-shaped decreasing activation function, to achieve progressive abstraction and transformation of the input data. Specifically, the input layer receives the original parameters, which are then processed hierarchically by 256, 192, 128, 96, and 64 neurons. Each layer enhances the model's generalization ability through batch normalization and dropout strategies. This process is essentially a complex data dimensionality reduction and feature reconstruction technique capable of capturing subtle nonlinear relationships between motor operating parameters. The multi-objective optimization objective is constructed as a comprehensive balance across four dimensions: system stability margin, dynamic response time, efficiency, and temperature rise. Through a weighted single-objective function transformation, stability margin and efficiency each account for 0.3 weights, while response time and temperature rise each account for 0.2 weights. This weighting design not only reflects the differentiated importance of each performance index in engineering practice, but also provides a precise objective function for subsequent optimization.

[0096] The Covariance Matrix Adaptive Evolutionary Strategy (CMA-ES) is the core of the optimization algorithm, and its computational process is extremely complex. The algorithm starts with an initial population of 50, using an initial step size of 0.3 and 25 parent selections to perform intelligent search in a high-dimensional parameter space. The specific computational process involves dynamic adjustment of the covariance matrix, normal distribution random sampling, and adaptive evolution. Each iteration adjusts the search strategy based on the performance evaluation results of the current population, gradually approaching the optimal solution. The entire process involves 500 iterations, ultimately generating a Pareto optimal solution set. Analyzing the Pareto optimal solution set first involves multi-dimensional analysis to extract the optimal configuration of the direct-axis and quadrature-axis current reference values. Then, two core functions are constructed: one describes the mapping relationship between the direct-axis current reference value and the rotational speed, and the other characterizes the functional dependency between the quadrature-axis current reference value and the torque demand. The construction of these two functions requires not only complex mathematical modeling but also full consideration of the motor's physical characteristics and operating boundaries. The current reference values ​​are repeatedly calculated under various operating conditions, and the parameter boundaries are continuously corrected by comparing actual operating data with predicted results.

[0097] In one specific embodiment, the process of executing step S105 may specifically include the following steps:

[0098] Define the state space, including seven key variables: motor speed, direct-axis current, quadrature-axis current, direct-axis voltage, quadrature-axis voltage, direct-axis voltage margin, and system stability margin index, and obtain the state observation vector;

[0099] Define the action space, which includes four variables: the proportional gain adjustment of the field weakening controller, the integral gain adjustment, the direct-axis current reference value correction coefficient, and the quadrature-axis current reference value correction coefficient, to obtain the control action vector;

[0100] By combining the state observation vector and the basic parameters of the weak magnetic control, a state-action mapping relationship is constructed, and a reward function is designed to obtain a reinforcement learning environment model.

[0101] A deep Q-network with a dual network structure was constructed based on a reinforcement learning environment model. Both networks use a 4-layer fully connected neural network with 128, 96 and 64 hidden layer neurons, respectively, to obtain the Q-value estimation network.

[0102] An update mechanism is set up for the Q-value estimation network. The evaluation network is updated once every step, and the target network is updated once every 100 steps. The capacity of the experience replay buffer is set to 10,000, and the network training rules are obtained.

[0103] An exploration strategy was designed based on the network training rules. The initial exploration rate was set to 1.0, and it was decayed to a minimum of 0.01 with each step. The decay rate was 0.995, and the discount factor was set to 0.95, resulting in an exploration-utilization balance strategy.

[0104] The Q-value estimation network and the exploration-balanced strategy were combined and trained for 500,000 steps in a motor simulation environment to obtain the trained Q-value network.

[0105] The trained Q-value network is used to evaluate the current motor state in real time, output the optimal control action, calculate the adjustment of the proportional gain and integral gain of the controller, and form dynamic control parameters.

[0106] Specifically, the state space is defined, containing seven key variables: motor speed representing the current operating speed of the motor; direct-axis current, the current component along the magnetic field direction in the vector control of the permanent magnet synchronous motor, used for field weakening control; quadrature-axis current, the current component perpendicular to the magnetic field direction, mainly controlling the motor torque; direct-axis voltage and quadrature-axis voltage representing the two orthogonal components of the stator voltage in the rotating coordinate system; direct-axis voltage margin, the margin between the direct-axis voltage and the maximum allowable voltage of the system, reflecting the system's tolerance to voltage fluctuations; and system stability margin, which quantifies the degree to which the system is close to an unstable state under the current operating conditions. These seven variables form the state observation vector, comprehensively describing the operating state of the motor field weakening control system. In defining the action space, four core control variables are selected: the proportional gain adjustment of the field weakening controller, representing the real-time adjustment magnitude of the proportional controller gain; the integral gain adjustment, representing the real-time adjustment magnitude of the integral controller gain; the direct-axis current reference value correction coefficient, used to correct the basic direct-axis current reference value for more precise field weakening control; and the quadrature-axis current reference value correction coefficient, used to correct the basic quadrature-axis current reference value for optimized torque output. These four variables together constitute the control action vector, which directly affects the adjustment of key parameters of the field weakening control system.

[0107] By combining the state observation vector and the fundamental parameters of the field weakening control, a state-action mapping relationship is constructed. This mapping relationship represents the system's response after taking a certain control action in a specific state. A reward function is also designed, which comprehensively considers four aspects: system stability, dynamic response speed, efficiency, and temperature rise. Positive rewards are given for control actions that improve stability, shorten response time, improve efficiency, and reduce temperature rise, while negative rewards are given for actions that do not. Through the definition of the state space, action space, and reward function, a complete reinforcement learning environment model is formed, providing a training scenario for deep reinforcement learning algorithms.

[0108] Based on the constructed reinforcement learning environment model, a double-network structure Deep Q-Network (DQN) is used for parameter optimization. Deep Q-Network is a type of reinforcement learning algorithm that approximates the Q-value function through a neural network to learn the optimal policy. The double-network structure consists of two parts: an evaluation network and a target network. The two networks have identical structures but different parameter update frequencies. Both networks use a 4-layer fully connected neural network, with the input layer corresponding to the state space dimension and the output layer corresponding to the action space dimension. Three hidden layers are set in between, with 128, 96, and 64 neurons respectively, forming a progressively shrinking structure. The evaluation network is responsible for estimating the Q-value of the current state-action pair, while the target network is used to calculate the target Q-value, stabilizing the training process.

[0109] An update mechanism is implemented for the Q-value estimation network. The evaluation network updates immediately after each interaction with the environment, while the target network parameters are updated every 100 steps, copied from the evaluation network's parameters. This asynchronous update mechanism effectively prevents oscillations and divergences during training. The experience replay buffer is a crucial component of deep reinforcement learning, used to store historical data of the agent's interactions with the environment, including a quadruple of state, action, reward, and next state. With a capacity of 10,000, the experience replay buffer stores the interaction data from the most recent 10,000 steps. A batch of data is randomly sampled from this buffer for training each time the network is updated, breaking the temporal correlation between data and improving training stability.

[0110] An exploration strategy was designed based on the network training rules, employing an ε-greedy strategy to balance exploration and exploitation. The initial exploration rate was set to 1.0, indicating completely random action selection in the early stages to fully explore the environment. As the number of training steps increased, the exploration rate decreased by a decay rate of 0.995, eventually reaching a minimum of 0.01. At this minimum, the agent selects the currently estimated optimal action with a 99% probability and explores randomly with a 1% probability. A discount factor of 0.95 was set, representing the discount rate for future rewards, controlling the algorithm's trade-off between short-term and long-term gains. This exploration-exploitation balancing strategy ensures that the algorithm can both fully explore the action space and gradually converge to the optimal strategy.

[0111] A Q-value estimation network and an exploration-balanced strategy were combined and trained iteratively in a motor simulation environment for a total of 500,000 interaction steps. In each interaction, the agent observes the current state, selects an action according to the exploration strategy, and the environment executes the action, transitioning to a new state and providing a reward. The agent stores this experience in a replay buffer. Then, a batch of experience data is randomly sampled from the buffer, the target Q-value is calculated, and the evaluation network parameters are updated. In this way, the Q-network gradually learns the value of various control actions in different states, ultimately forming a policy network capable of selecting the optimal control action based on the motor state.

[0112] After training, the resulting Q-value network is applied to the actual control process. In each control cycle, the current motor state data, including parameters such as speed, current, and voltage, is first acquired to form a state observation vector. This vector is then input into the trained Q-value network, which outputs value estimates for various possible actions. The action with the highest value is selected as the control decision, resulting in adjustments to the controller's proportional and integral gains, as well as correction coefficients for the current reference value. Finally, the controller parameters and current reference value are modified based on these adjustments to achieve dynamic optimization of the controller parameters.

[0113] Taking a certain type of permanent magnet synchronous motor as an example, during field weakening control, when the detected motor speed is 4500 rpm (1.5 times the rated speed), the direct-axis current is -40A, the quadrature-axis current is 60A, the direct-axis voltage is 150V, the quadrature-axis voltage is 200V, the direct-axis voltage margin is 50V, and the system stability margin is 0.25, this state observation vector is input into the trained Q-value network. After forward calculation, the network outputs value estimates for four control adjustments and selects the action with the highest value: a proportional gain adjustment of +0.8, an integral gain adjustment of -0.5, a direct-axis current reference value correction coefficient of 1.05, and a quadrature-axis current reference value correction coefficient of 0.98. This means that under the current state, the proportional gain should be appropriately increased, the integral gain decreased, the negative value of the direct-axis current slightly increased to enhance the field weakening effect, and the quadrature-axis current slightly decreased to optimize the torque output. After these adjustments were applied to the controller, the motor's operating state became more stable, the direct-axis voltage margin increased to 65V, the system stability margin improved to 0.35, and the dynamic response became faster, achieving the effect of intelligent adaptive control.

[0114] In one specific embodiment, the process of executing step S106 may specifically include the following steps:

[0115] The dynamic control parameters and the basic parameters of the field weakening control are integrated, and the final control parameters are calculated using a weighted average method to obtain the comprehensive control parameters.

[0116] The integrated control parameters are configured into the motor field weakening control system, and the proportional gain and integral gain of the controller are updated to obtain the parameter update control system.

[0117] The operating status of the parameter update control system is monitored in real time. The sampling frequency is set to 10kHz, and the stator current, voltage, speed and temperature are collected to obtain operating feedback data.

[0118] The system performance indicators, including stability margin, dynamic response time, efficiency, and temperature rise, are calculated using operational feedback data to obtain actual performance indicator values.

[0119] The actual performance index values ​​are compared with the predicted values ​​of the field weakening control performance prediction model, the prediction error is calculated, and the model accuracy evaluation data is obtained.

[0120] Set up a data buffer pool based on the model accuracy evaluation data, and store the latest 1000 sets of data according to the first-in-first-out principle to obtain the incremental training dataset;

[0121] The incremental training dataset is merged with the original training data, and representative samples are selected and retained to obtain the updated training dataset.

[0122] The field weakening control performance prediction model was incrementally trained using the updated training dataset, with a learning rate of 0.0005 and an iteration period of 50, thus forming a closed-loop optimization control method for the field weakening motor.

[0123] Specifically, the dynamic control parameters and the fundamental parameters of the field weakening control are not simply linearly superimposed, but deeply coupled through a complex weight allocation algorithm based on information entropy. Each parameter is treated as a multidimensional feature vector, and its weight is determined by the instantaneous state of the motor operation, historical performance data, and the confidence level of the prediction model. A sampling frequency of 10kHz means that 10,000 data points are captured per second. The four parameters of stator current, voltage, speed, and temperature are considered as orthogonal fundamental features of the motor's operating state. Each sampling point is not merely a numerical value, but a multidimensional information carrier containing rich dynamic characteristics of the motor's operation.

[0124] The calculation of performance metrics is an extremely complex signal processing and feature extraction process. The calculation of the four metrics—stability margin, dynamic response time, efficiency, and temperature rise—is not a simple numerical operation, but rather obtained through a series of highly nonlinear signal transformations and feature mapping algorithms. The algorithm needs to simultaneously consider instantaneous sampled data, historical operating data, and the physical characteristic model of the motor, constructing a multi-dimensional performance evaluation hyperspace. Model accuracy evaluation is a complex comparative analysis process based on information theory. Error analysis between actual measured metrics and model predicted metrics is not limited to simple numerical comparisons, but involves a comprehensive evaluation using advanced statistical indicators such as information divergence and entropy change in the multi-dimensional error space. Each error is not merely a numerical value, but a quantitative indicator of the model's cognitive ability. The design of the data buffer pool embodies a dynamic learning strategy based on information value. The first-in, first-out (FIFO) mechanism is not just simple data storage, but a complex data value assessment and selection process. Through multi-dimensional indicators such as information entropy, data sparsity, and representativeness, the most cognitively valuable samples are extracted from 1000 sets of data. This selection mechanism ensures that the model can continuously learn from the most critical and information-dense operating data. Incremental training of the model is a highly complex adaptive learning process. The tiny learning rate of 0.0005 means that the model adjusts its parameters at an extremely cautious pace. The 50 training iterations are not just simple numerical iterations, but a dynamic and self-organizing learning process in which the model continuously adjusts its cognitive boundaries through tiny parameter perturbations.

[0125] For example, under a specific motor operating condition, the algorithm captures a set of multi-dimensional operating data. This data not only includes instantaneous values ​​of stator current, voltage, speed, and temperature, but more importantly, the complex correlations between these data. Through multi-dimensional nonlinear mapping, the algorithm extracts a set of implicit feature vectors. These feature vectors undergo complex information processing and are ultimately transformed into predicted values ​​of performance indicators. When there is a slight deviation between the predicted and actual measured values, this deviation itself becomes a key source of information for further model learning and optimization.

[0126] The above describes the deep learning-based field-weakening motor optimization method in the embodiments of this application. The following describes the deep learning-based field-weakening motor optimization system in the embodiments of this application. Please refer to [link / reference]. Figure 2 One embodiment of the deep learning-based field-weakening motor optimization system in this application includes:

[0127] The data acquisition module is used to collect stator current, rotor position angle, voltage and temperature data of permanent magnet motor at different speeds to obtain raw data of field weakening operation.

[0128] The processing module is used to denoise and normalize the maximum and minimum values ​​of the raw data of the field weakening operation to obtain standardized field weakening control feature data.

[0129] The training module is used to input the standardized field weakening control feature data into a five-layer hidden deep neural network for training, so as to obtain a field weakening control performance prediction model.

[0130] The adjustment module is used to perform parameter adjustment calculations on the direct-axis current reference value and the quadrature-axis current reference value based on the output results of the field weakening control performance prediction model, so as to obtain the basic parameters of field weakening control.

[0131] The calculation module is used to calculate the real-time adjustment of the controller's proportional gain and integral gain by combining the field weakening control basic parameters with the motor operating status data, thereby obtaining dynamic control parameters.

[0132] An optimization module is used to apply the dynamic control parameters to the motor field weakening control system and to iteratively optimize the field weakening control performance prediction model through feedback data, thereby forming a closed-loop optimization control method for the field weakening motor.

[0133] Through the collaborative efforts of the aforementioned components, a high-frequency (10kHz) data acquisition strategy was employed to comprehensively capture the multi-parameter characteristics of motor operation, providing a rich and accurate data foundation for the deep learning model. The five-layer hidden layer structure of the deep neural network, by progressively decreasing the number of neurons in each layer, enabled multi-scale feature extraction of the motor's operating data, effectively capturing the complex nonlinear dynamic characteristics of the motor. Artificial intelligence algorithms played a crucial role in the solution, particularly the introduction of deep learning and reinforcement learning models, which broke through the static parameter limitations of traditional control methods, constructing an intelligent control system capable of autonomous learning and dynamic optimization.

[0134] The multi-objective optimization algorithm achieves a multi-dimensional balance of system performance through precisely designed weight coefficients (0.3 each for stability margin and efficiency, and 0.2 each for response time and temperature rise). The introduction of the Covariance Matrix Adaptive Evolutionary Strategy (CMA-ES) enables the parameter optimization process to intelligently search within a complex high-dimensional parameter space, effectively avoiding getting trapped in local optima. The design of the deep reinforcement learning model elevates the controller's adaptive adjustment to a new level, constructing an intelligent system capable of real-time learning and adjustment of the control strategy by defining a finely defined state and action space. The attention mechanism combined with a Long Short-Term Memory (LSTM) network state prediction model further enhances the system's predictive control capability, enabling early identification of potential unstable operating conditions. This significantly improves the motor's performance indicators, achieving high-performance and high-efficiency operation over a wide speed range.

[0135] Reference Figure 3 This invention also provides a computer device, which can be a server, and its internal structure can be as follows: Figure 3As shown, the computer device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores the data corresponding to this embodiment. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0136] Those skilled in the art will understand that Figure 3 The structures shown are merely block diagrams of some structures related to the present invention and do not constitute a limitation on the computer devices on which the present invention is applied.

[0137] An embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. It is understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0138] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A deep learning-based optimization method for field-weakening motors, characterized in that, The deep learning-based optimization method for field-weakening motors includes: A high-frequency data acquisition strategy was adopted to collect stator current, rotor position angle, voltage and temperature data of permanent magnet motor at different speeds to obtain raw data of field weakening operation; The raw data of the field weakening operation is denoised and normalized for maximum and minimum values ​​to obtain standardized field weakening control feature data. The standardized field weakening control feature data is input into a five-layer hidden deep neural network for training to obtain a field weakening control performance prediction model. Based on the output of the field weakening control performance prediction model, parameter adjustment calculations are performed on the direct-axis current reference value and the quadrature-axis current reference value to obtain the basic parameters of field weakening control. These include: inputting the current motor speed, load torque, and voltage limits into the field weakening control performance prediction model to obtain predicted values ​​for four performance indicators: system stability margin, dynamic response time, efficiency, and temperature rise, thus obtaining operating condition performance characteristic data; setting a multi-objective optimization function for the operating condition performance characteristic data, taking maximizing stability margin, minimizing response time, maximizing efficiency, and minimizing temperature rise as optimization objectives, thus obtaining the objective function expression; converting the objective function expression into a weighted single-objective function, setting weight coefficients for each of the four indicators, thus obtaining the comprehensive optimization objective; and constructing an improved covariance matrix based on the comprehensive optimization objective. An adaptive evolutionary strategy algorithm is used to set the population size, initial step size, and number of parents to obtain a parameter optimization algorithm. This algorithm is then used to iteratively optimize the direct-axis current reference value and the quadrature-axis current reference value to obtain a Pareto optimal solution set. The Pareto optimal solution set is analyzed to extract the optimal configuration of the direct-axis current reference value and the quadrature-axis current reference value, resulting in a current reference value configuration scheme. Based on this configuration scheme, correction formulas for the direct-axis current reference value and the quadrature-axis current reference value are established, representing the direct-axis current as a function of rotational speed and the quadrature-axis current as a function of torque demand, respectively, thus obtaining a current reference value calculation model. This model is then validated and tested under different operating conditions, and boundary adjustments are made based on the validation results to form the fundamental parameters for field weakening control. By combining the aforementioned field weakening control basic parameters with motor operating status data, the real-time adjustment amounts of the controller's proportional gain and integral gain are calculated to obtain dynamic control parameters. The dynamic control parameters are applied to the field weakening control system of the motor, and the field weakening control performance prediction model is cyclically optimized through feedback data to form a closed-loop optimization control method for the field weakening motor.

2. The deep learning-based field-weakening motor optimization method according to claim 1, characterized in that, The process involves collecting stator current, rotor position angle, voltage, and temperature data of the permanent magnet motor at different speeds to obtain raw data for field weakening operation, including: Multiple test points were set up within a range of 0.2 to 3 times the rated speed of the motor. Data including stator current, rotor position angle, motor speed, stator voltage, field weakening current component, torque current component, direct axis current, quadrature axis current, power factor and efficiency index were collected at each test point to obtain raw data for multiple operating conditions. The original data under multiple operating conditions is classified according to the motor's operating region, into three categories: constant torque region, field weakening region, and constant power region, to obtain regional datasets. The data in the constant torque region retains its original characteristics, the data in the field weakening region is supplemented with a field weakening depth index, and the data in the constant power region is supplemented with a power fluctuation index to obtain an enhanced feature dataset. Based on the enhanced feature dataset, the stable working boundary of each working region is calculated, the critical point and unstable operating region of the weak magnetic region are determined, and the operating boundary data of the weak magnetic region is obtained. The raw collected data is filtered using the boundary data of the weak magnetic field region to remove abnormal data points that do not meet the stable operating conditions, thus obtaining preliminary purified data. The sliding window mean filtering algorithm is applied to the preliminary purified data, with a window length of 5 sampling points, to eliminate high-frequency noise interference and obtain denoised data; The denoised data is normalized according to the maximum and minimum values ​​of each parameter, and the maximum and minimum value normalization transformation is performed on each parameter to obtain normalized data. The normalized data is randomly divided in an 8:2 ratio, with 80% of the data assigned as training data and 20% as validation data, forming the original data for weak magnetic field operation.

3. The deep learning-based field-weakening motor optimization method according to claim 1, characterized in that, The process of denoising and normalizing the maximum and minimum values ​​of the raw data for field weakening operation yields standardized field weakening control characteristic data, including: The raw data of the weak magnetic field operation is subjected to outlier detection. Data points that exceed the normal range are identified using the three-standard-deviation principle, and outlier-marked data are obtained. Based on the outlier marking data, the raw data of the weak magnetic field operation is filtered, and the data points marked as outliers are replaced with the average value of adjacent normal data points to obtain preliminary cleaned data; The preliminary cleaning data is processed using a median filtering algorithm. A seven-point window is selected to filter each signal channel separately to obtain noise-suppressed data. Wavelet transform decomposition was applied to the noise-suppressed data, and five-level de Becié wavelets were selected for multi-scale analysis to extract signal features and reconstruct them, resulting in finely denoised data. Based on the finely denoised data, the maximum and minimum values ​​of each parameter are calculated, a normalized transformation parameter table is established, and normalized baseline data is obtained. The normalized baseline data is used to perform a maximum-minimum normalization transformation on each parameter in the finely denoised data, adjusting the numerical range to the interval between zero and one, to obtain normalized feature data; The normalized feature data is subjected to data balancing processing, and the data in the weak magnetic region is sampled in a key manner to ensure that the data proportion of each working region is balanced, thus obtaining balanced feature data. The equilibrium characteristic data is divided into multiple data segments according to the time series, and each segment contains state parameters at twenty consecutive time points to form standardized weak magnetic control characteristic data.

4. The deep learning-based field-weakening motor optimization method according to claim 1, characterized in that, The process of inputting the standardized field weakening control feature data into a five-layer hidden deep neural network for training to obtain a field weakening control performance prediction model includes: A deep neural network structure consisting of an input layer, five hidden layers, and an output layer is constructed. The number of neurons in the input layer is the same as the feature dimension. The number of neurons in the five hidden layers are 256, 192, 128, 96, and 64, respectively. The number of neurons in the output layer is 4, thus obtaining the basic neural network architecture. A funnel-shaped structure is set for the hidden layers of the neural network infrastructure, and a batch normalization layer and a dropout layer are added after each layer. The dropout rate is set to 20%, resulting in an anti-overfitting network structure. Funnel-shaped decreasing activation functions are configured for the hidden layer neurons in the anti-overfitting network structure. A funnel-shaped decreasing activation function with a parameter of 0.01 is used to obtain nonlinear mapping units. Based on the feature distribution of the standardized field weakening control feature data, a linear activation function is set for the output layer, and a weighted mean square error is defined as the loss function to obtain the training objective function; Weights are assigned to each performance index in the training objective function: system stability margin is weighted at 0.3, dynamic response time at 0.2, efficiency at 0.3, and temperature rise at 0.2, to obtain the weighted training objective. Based on the weighted training objective, an adaptive moment estimation algorithm with a learning rate of 0.001 is selected as the optimizer, and the batch size is set to 64 to obtain the network training strategy. The neural network was trained using the aforementioned network training strategy. The training cycle was set to 300, and an early stopping strategy was adopted. Training was stopped when the validation set loss did not improve for 15 consecutive cycles, thus obtaining the initial training model. The initial training model is subjected to performance testing, the prediction error on the test set is calculated, and the network structure is fine-tuned through Bayesian hyperparameter optimization to form a magnetic field weakening control performance prediction model.

5. The deep learning-based field-weakening motor optimization method according to claim 1, characterized in that, Based on the output results of the field weakening control performance prediction model, parameter adjustment calculations are performed on the direct-axis current reference value and the quadrature-axis current reference value to obtain the basic parameters of field weakening control, including: Input the current motor speed, load torque and voltage limit conditions into the field weakening control performance prediction model to obtain the predicted values ​​of four performance indicators: system stability margin, dynamic response time, efficiency and temperature rise, and obtain the operating condition performance characteristic data. A multi-objective optimization function is set for the operating condition performance characteristic data, with maximizing stability margin, minimizing response time, maximizing efficiency, and minimizing temperature rise as optimization objectives, resulting in the objective function expression; The objective function expression is converted into a weighted single objective function, and weight coefficients of 0.3, 0.2, 0.3 and 0.2 are set for the four indicators respectively to obtain the comprehensive optimization objective; An improved covariance matrix adaptive evolution strategy algorithm is constructed based on the comprehensive optimization objective. The population size is set to 50, the initial step size is 0.3, and the number of parents is 25, resulting in a parameter optimization algorithm. The parameter optimization algorithm was used to iteratively optimize the direct-axis current reference value and the quadrature-axis current reference value. The number of iterations was set to 500 to obtain the Pareto optimal solution set. The Pareto optimal solution set is analyzed to extract the optimal configuration of the direct-axis current reference value and the quadrature-axis current reference value, thus obtaining the current reference value configuration scheme. Based on the current reference value configuration scheme, establish the direct-axis current reference value correction formula and the quadrature-axis current reference value correction formula, which are respectively expressed as functions of direct-axis current and rotational speed and functions of quadrature-axis current and torque demand, to obtain the current reference value calculation model; The current reference value calculation model was verified and tested. The current reference value was calculated under different operating conditions, and boundary corrections were made based on the verification results to form the basic parameters for field weakening control.

6. The deep learning-based field-weakening motor optimization method according to claim 1, characterized in that, The process of using the field weakening control basic parameters combined with motor operating status data to calculate the real-time adjustment of the controller's proportional gain and integral gain to obtain dynamic control parameters includes: Define the state space, including seven key variables: motor speed, direct-axis current, quadrature-axis current, direct-axis voltage, quadrature-axis voltage, direct-axis voltage margin, and system stability margin index, and obtain the state observation vector; Define the action space, which includes four variables: the proportional gain adjustment of the field weakening controller, the integral gain adjustment, the direct-axis current reference value correction coefficient, and the quadrature-axis current reference value correction coefficient, to obtain the control action vector; By combining the state observation vector and the basic parameters of the weak magnetic field control, a state-action mapping relationship is constructed, and a reward function is designed to obtain a reinforcement learning environment model. Based on the reinforcement learning environment model, a deep Q-network with a dual network structure is constructed. Both networks use a 4-layer fully connected neural network with 128, 96 and 64 hidden layer neurons, respectively, to obtain the Q-value estimation network. An update mechanism is set up for the Q-value estimation network, the evaluation network is updated once per step, the target network is updated once every 100 steps, the experience replay buffer capacity is set to 10000, and the network training rules are obtained. Based on the network training rules, an exploration strategy was designed. The initial exploration rate was set to 1.0, and it was decayed to a minimum of 0.01 with each step. The decay rate was 0.995, and the discount factor was set to 0.95, resulting in an exploration-utilization balance strategy. The Q-value estimation network and the exploration-utilization balancing strategy are combined, and 500,000 training iterations are performed in a motor simulation environment to obtain the trained Q-value network. The trained Q-value network is used to evaluate the current motor state in real time, output the optimal control action, calculate the adjustment of the proportional gain and integral gain of the controller, and form dynamic control parameters.

7. The deep learning-based field-weakening motor optimization method according to claim 1, characterized in that, The method of applying the dynamic control parameters to the field weakening control system of the motor and cyclically optimizing the field weakening control performance prediction model through feedback data to form a closed-loop optimization control method for the field weakening motor includes: The dynamic control parameters are fused with the basic parameters of field weakening control, and the final control parameters are calculated using a weighted average method to obtain the comprehensive control parameters. The integrated control parameters are configured into the motor field weakening control system, and the proportional gain and integral gain of the controller are updated to obtain the parameter update control system. The operating status of the parameter update control system is monitored in real time. The sampling frequency is set to 10kHz, and stator current, voltage, speed and temperature are collected to obtain operating feedback data. The system performance indicators, including stability margin, dynamic response time, efficiency, and temperature rise, are calculated using the operational feedback data to obtain actual performance indicator values. The actual performance index values ​​are compared with the predicted values ​​of the field weakening control performance prediction model, the prediction error is calculated, and the model accuracy evaluation data is obtained. Based on the model accuracy evaluation data, a data buffer pool is set up, and the latest 1000 sets of data are stored according to the first-in-first-out principle to obtain the incremental training dataset; The incremental training dataset is merged with the original training data, and representative samples are selected and retained to obtain the updated training dataset. The field weakening control performance prediction model is incrementally trained using the updated training dataset, with a learning rate of 0.0005 and an iteration period of 50, thus forming a closed-loop optimization control method for the field weakening motor.

8. A deep learning-based field-weakening motor optimization system, used to implement the deep learning-based field-weakening motor optimization method as described in any one of claims 1-7, characterized in that, The deep learning-based field-weakening motor optimization system includes: The data acquisition module is used to collect stator current, rotor position angle, voltage and temperature data of permanent magnet motor at different speeds to obtain raw data of field weakening operation. The processing module is used to perform noise reduction and maximum / minimum value normalization processing on the raw data of the field weakening operation to obtain standardized field weakening control feature data. The training module is used to input the standardized field weakening control feature data into a five-layer hidden deep neural network for training, so as to obtain a field weakening control performance prediction model. The adjustment module is used to perform parameter adjustment calculations on the direct-axis current reference value and the quadrature-axis current reference value based on the output results of the field weakening control performance prediction model, so as to obtain the basic parameters of field weakening control. The calculation module is used to calculate the real-time adjustment of the controller's proportional gain and integral gain by combining the field weakening control basic parameters with the motor operating status data, thereby obtaining dynamic control parameters. An optimization module is used to apply the dynamic control parameters to the motor field weakening control system and to iteratively optimize the field weakening control performance prediction model through feedback data, thereby forming a closed-loop optimization control method for the field weakening motor.

9. A computer device, characterized in that, It includes a memory and a processor, the memory storing a computer program that can run on the processor, and the processor executing the computer program to implement the deep learning-based magnetic field weakening motor optimization method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when run by a processor, causes the processor to execute the deep learning-based magnetic field weakening motor optimization method as described in any one of claims 1 to 7.

Images