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

Abstract

The invention 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 steps that motor operation data at different rotating speeds are collected and subjected to denoising and normalization processing, then a deep neural network model is constructed, intelligent adjustment of direct-axis and quadrature-axis current reference values is achieved, control parameters are dynamically calculated in combination with operation state data, and a closed-loop optimization control system is formed. According to the invention, high-performance and high-efficiency operation of the motor in a wide rotating speed range is realized, control parameters can be dynamically adjusted in real time, and multiple performance indexes such as system stability, response speed, energy efficiency and temperature rise are effectively balanced.

Description

Technical Field

[0001] The present application relates to the technical field of flux-weakening motor optimization, and in particular to a flux-weakening motor optimization method, system, and storage medium based on deep learning. Background Art

[0002] In modern industrial production, permanent magnet motors (PMMs) 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 PMM control technology relies primarily on the classic PI (proportional-integral) control method, which achieves field weakening by adjusting the stator current and rotor flux to meet performance requirements under varying speed and load conditions. Engineers have long been committed to improving the performance and energy efficiency of motors across a wide speed range.

[0003] However, traditional flux-weakening control methods have many inherent limitations. First, the fixed-parameter PI controller cannot effectively adapt to the complex dynamic characteristics of the motor under different operating conditions, which severely restricts control performance. Second, the lack of in-depth understanding of the complex physical processes within the motor and the ability to model them in real time make 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 between different indicators. Summary of the Invention

[0004] This application provides a deep learning-based weak-magnetic motor optimization method, system and storage medium for achieving high-performance and high-efficiency operation of the motor within a wide speed range, and can dynamically adjust control parameters in real time to effectively balance multiple performance indicators such as system stability, response speed, energy efficiency and temperature rise.

[0005] In the first aspect, the present application provides a method for optimizing a weak magnetic motor based on deep learning, and the method for optimizing a weak magnetic motor based on deep learning includes: collecting the stator current, rotor position angle, voltage and temperature data of the permanent magnet motor at different speeds to obtain the original data of weak magnetic operation; denoising and normalizing the original data of weak magnetic operation to obtain standardized weak magnetic control feature data; inputting the standardized weak magnetic control feature data into a five-layer hidden layer deep neural network for training to obtain a weak magnetic control performance prediction model; according to the output result of the weak magnetic control performance prediction model, performing parameter adjustment calculation on the direct-axis current reference value and the quadrature-axis current reference value to obtain basic parameters of weak magnetic control; using the basic parameters of weak magnetic control in combination with the motor operating status data to calculate the real-time adjustment amount of the controller proportional gain and integral gain to obtain dynamic control parameters; applying the dynamic control parameters to the motor weak magnetic control system, and optimizing the weak magnetic control performance prediction model through feedback data loop to form a closed-loop optimization control method for a weak magnetic motor.

[0006] In a second aspect, the present application provides a deep learning-based flux weakening motor optimization system, the deep learning-based flux weakening motor optimization system comprising:

[0007] The acquisition module is used to collect the stator current, rotor position angle, voltage and temperature data of the permanent magnet motor at different speeds to obtain the original data of the weak magnetic operation;

[0008] A processing module is used to perform denoising and maximum and minimum value normalization on the raw data of the weak magnetic operation to obtain standardized weak magnetic control characteristic data;

[0009] A training module, configured to input the standardized magnetic weakening control characteristic data into a five-hidden-layer deep neural network for training to obtain a magnetic weakening control performance prediction model;

[0010] an adjustment module, configured to perform parameter adjustment calculation on a direct-axis current reference value and a quadrature-axis current reference value according to an output result of the flux-weakening control performance prediction model, so as to obtain basic flux-weakening control parameters;

[0011] A calculation module is used to calculate the real-time adjustment amount of the controller proportional gain and integral gain by using the basic parameters of the weak magnetic control in combination with the motor operating state data to obtain dynamic control parameters;

[0012] The optimization module is used to apply the dynamic control parameters to the motor weakening control system and optimize the weakening control performance prediction model through feedback data loop to form a weakening motor closed-loop optimization control method.

[0013] In a third aspect, a computer device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned deep learning-based weak magnetic motor optimization method.

[0014] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, enables the computer to execute the above-mentioned deep learning-based weak magnetic motor optimization method.

[0015] In the technical solution provided by this application, a high-frequency (10kHz) data acquisition strategy is adopted to fully 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 realizes multi-scale feature extraction of motor operation data by decreasing the number of neurons layer by layer, effectively capturing the complex nonlinear dynamic characteristics of the motor. Artificial intelligence algorithms play a key role in the solution, especially the introduction of deep learning and reinforcement learning models, which breaks through the static parameter limitations of traditional control methods and builds an intelligent control system capable of autonomous learning and dynamic optimization. The multi-objective optimization algorithm achieves a multi-dimensional balance of system performance through precisely designed weight coefficients (stability margin and efficiency are 0.3 each, response time and temperature rise are 0.2 each). The introduction of the covariance matrix adaptive evolutionary strategy (CMA-ES) enables the parameter optimization process to perform intelligent search in a complex high-dimensional parameter space, effectively avoiding falling into a local optimal solution. The design of the deep reinforcement learning model has elevated the controller adaptive adjustment to a whole new level. By defining a fine state space and action space, an intelligent system capable of real-time learning and adjustment of control strategies is constructed. The attention mechanism, combined with the state prediction model of the 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 the motor's performance indicators, enabling high-performance and high-efficiency operation across a wide speed range. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 Schematic diagram of an embodiment of a method for optimizing a weak magnetic motor based on deep learning in an embodiment of the present application;

[0018] Figure 2 Schematic diagram of an embodiment of a flux-weakening motor optimization system based on deep learning in an embodiment of the present application;

[0019] Figure 3 It is a schematic block diagram of the structure of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The embodiments of the present application provide a method, system and storage medium for optimizing a weak magnetic motor based on deep learning. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or that are inherent to these processes, methods, products or devices.

[0021] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiments of the present application, an embodiment of a method for optimizing a weak magnetic motor based on deep learning includes:

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

[0023] Step S102: De-noising and maximum and minimum value normalization processing are performed on the original magnetic weakening operation data to obtain standardized magnetic weakening control characteristic data;

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

[0025] Step S104: performing parameter adjustment calculation on the direct-axis current reference value and the quadrature-axis current reference value according to the output result of the field-weakening control performance prediction model to obtain basic field-weakening control parameters;

[0026] Step S105: Calculate the real-time adjustment amount of the controller proportional gain and integral gain by using the basic parameters of the weak magnetic control and the motor operating state data to obtain the dynamic control parameters;

[0027] Step S106 : applying the dynamic control parameters to the motor flux weakening control system, and optimizing the flux weakening control performance prediction model through feedback data loop to form a flux weakening motor closed-loop optimization control method.

[0028] It is understandable that the execution subject of this application can be a weak magnetic motor optimization system based on deep learning, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0029] Specifically, test points were set within a range of 0.2 to 3 times the motor's rated speed to collect parameters such as stator current, rotor position angle, voltage, and temperature. This data was categorized into three zones: constant torque, field weakening, and constant power, with specific indicators added to each. For the field weakening zone data, a field weakening depth indicator was added to reflect the extent to which the motor entered field weakening control. For the constant power zone data, a power fluctuation indicator was added to characterize the motor's power stability in the high-speed range. Simultaneously, by calculating the stable operating boundaries of each operating zone, the critical points and unstable operating regions of the field weakening zone were determined, thereby filtering out data that met stable operating conditions. The collected raw field weakening data was denoised and normalized to obtain standardized field weakening control characteristic data. During the denoising process, the data was first detected for outliers. Anomalous data points were identified using the triple standard deviation principle and replaced with the average of the adjacent normal data points. The median filter algorithm was then applied to the preliminarily cleaned data, with a seven-point window length selected to filter each signal channel separately to effectively remove sudden interference. Then, using wavelet transform decomposition technology, a five-layer Debesie wavelet was selected for multi-scale analysis, extracting signal features and reconstructing them to obtain refined denoised data. Normalization processing involves calculating the maximum and minimum values of each parameter, establishing a normalized conversion parameter table, and adjusting the value range to the range of 0 to 1, which helps improve the training effect of the neural network.

[0030] Standardized field-weakening control feature data was fed into a five-hidden-layer deep neural network for training to construct a field-weakening control performance prediction model. In this network, the five hidden layers have 256, 192, 128, 96, and 64 neurons, respectively, 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 for the hidden layers to prevent the vanishing gradient problem. A linear activation function was used in the output layer to accommodate the continuous numerical nature of the performance metrics. During training, a weighted mean square error (WMSE) loss function was used, assigning weights of 0.3, 0.2, 0.3, and 0.2 to the system stability margin, dynamic response time, efficiency, and temperature rise, respectively. The optimizer used an adaptive moment estimation algorithm with a learning rate of 0.001, a batch size of 64, and a training cycle of 300. Based on the output of the field-weakening control performance prediction model, the direct-axis and quadrature-axis current reference values were adjusted and calculated to obtain the basic field-weakening control parameters. 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. A multi-objective optimization function is then established, with maximizing stability margin, minimizing response time, maximizing efficiency, and minimizing temperature rise as optimization objectives. This is then converted into a weighted single-objective function. Parameter optimization is performed using an improved covariance matrix adaptive evolutionary strategy algorithm, with a population size of 50, an initial step size of 0.3, and a number of parents of 25. Optimization is performed 500 times through iterations to obtain a Pareto-optimal solution set. The optimal configuration of the direct-axis and quadrature-axis current reference values is extracted from this, and a current reference value calculation model is established.

[0031] By combining the basic parameters of weak magnetic control with the motor operating status data, the real-time adjustment of the controller's proportional gain and integral gain is calculated to obtain the dynamic control parameters. This step first defines the state space containing 7 key variables and the action space containing 4 variables, and constructs the state-action mapping relationship and reward function. Based on this, a deep Q network with a dual network structure is constructed, and both networks use a 4-layer fully connected neural network. An update mechanism is set up, the evaluation network is updated once per step, and the target network is updated every 100 steps. An exploration strategy is designed with an initial exploration rate of 1.0, which decays to a minimum value of 0.01 according to the number of steps, 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 is obtained that can evaluate the current motor state in real time and output the optimal control action.

[0032] Dynamic control parameters are applied to the motor's field-weakening control system. The field-weakening control performance prediction model is optimized through feedback data loops, resulting in a closed-loop optimization control method for field-weakening motors. Specifically, a weighted average fusion of the dynamic control parameters and the basic field-weakening control parameters is performed to obtain comprehensive control parameters, which are then configured in the control system. The system's operating status is monitored in real time with a sampling frequency of 10kHz. The error between the actual performance indicators and the predicted values is calculated. A data buffer is set up to store the latest 1000 sets of data. This data is then merged with the existing training data for incremental model training with a learning rate of 0.0005 and an iteration cycle of 50. For example, at a speed 1.5 times the rated speed of 2000 rpm, the collected data are a direct-axis current of 10A, a quadrature-axis current of 15A, a direct-axis voltage of 120V, and a quadrature-axis voltage of 200V. These data are filtered through a seven-point window median filter to effectively suppress noise. Maximum and minimum value normalization is then used to map the values to a range between 0 and 1. The processed data was fed into a trained five-layer deep neural network, which predicted a system stability margin of 0.25, a dynamic response time of 0.08s, an efficiency of 93%, and a temperature rise of 40°C. Based on these predictions, the covariance matrix adaptive evolutionary strategy algorithm calculated that the direct-axis current reference value was corrected to -12A, while the quadrature-axis current reference value remained at 15A. The deep Q network further calculated a proportional gain adjustment of +0.5 and an integral gain adjustment of -0.2. After these parameters were applied to the control system, through the continuous collection of feedback data and incremental model training, the closed-loop optimization system continuously adapted to changes in the motor's operating state, maintaining optimal control results.

[0033] In the embodiment of the present application, a high-frequency (10kHz) data acquisition strategy is adopted to comprehensively capture the multi-parameter characteristics of the 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 realizes the multi-scale feature extraction of the motor operation data by decreasing the number of neurons layer by layer, and effectively captures the complex nonlinear dynamic characteristics of the motor. Artificial intelligence algorithms play a key role in the solution, especially the introduction of deep learning and reinforcement learning models, which breaks through the static parameter limitations of traditional control methods and builds 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 for stability margin and efficiency, 0.2 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 in a complex high-dimensional parameter space, effectively avoiding falling into local optimal solutions. The design of the deep reinforcement learning model has taken controller adaptive adjustment to a whole new level. By defining a detailed state space and action space, an intelligent system capable of learning and adjusting control strategies in real time has been constructed. The attention mechanism combined with the state prediction model of the long short-term memory network (LSTM) further enhances the system's predictive control capabilities and can identify potential unstable working conditions in advance. The performance indicators of the motor have been significantly improved, and high-performance and high-efficiency operation of the motor over a wide speed range has been achieved.

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

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

[0037] The original data of multiple working conditions are classified according to the motor working area, and divided into three types of data: constant torque area, weak magnetic area and constant power area, to obtain regional data sets;

[0038] The original features of the data in the constant torque area are maintained, the weak magnetic depth index is added to the data in the weak magnetic area, and the power fluctuation index is added to the data in the constant power area to obtain an enhanced feature data set;

[0039] Calculate the stable working boundary of each working area based on the enhanced feature data set, determine the critical point and unstable operating area of the weakening magnetic area, and obtain the operating boundary data of the weakening magnetic area;

[0040] The original collected data is screened using the operating boundary data of the weak magnetic area, and abnormal data points that do not meet the stable operating conditions are eliminated to obtain preliminary purified data;

[0041] Apply the sliding window mean filter algorithm to the preliminary cleaned data, set the window length to 5 sampling points, eliminate high-frequency noise interference, and obtain denoised data;

[0042] The denoised data is normalized according to the maximum and minimum values of each parameter, and the maximum and minimum normalization conversion is performed on each parameter to obtain normalized data;

[0043] The normalized data are randomly divided in a ratio of 8:2, with 80% of the data divided into training data and 20% of the data divided into verification data to form the original data of weak magnetic operation.

[0044] Specifically, multiple test points are set in the range of 0.2 to 3 times the rated speed of the permanent magnet motor. This setting ensures that the data covers various working conditions of the motor from low speed to high speed. For each test point, the acquisition system will record basic electrical parameters such as stator current, rotor position angle, motor speed, stator voltage, as well as weak magnetic current components and torque current components. The weak magnetic current component refers to the current used to weaken the magnetic field, which usually corresponds to the negative value of the direct-axis current; the torque current component is the current that generates electromagnetic torque, which usually corresponds to the quadrature-axis current. The direct-axis current and quadrature-axis current are current components expressed in a rotating coordinate system, which are converted from the three-phase current through coordinate transformation. In addition, the power factor and efficiency indicators are also recorded at the same time to form comprehensive multi-working condition raw data.

[0045] The collected raw data from multiple operating conditions needs to be categorized and organized according to the motor's operating region. The operating regions of permanent magnet motors are primarily divided into three categories: constant torque, field weakening, and constant power. The constant torque region refers to the area below the rated speed where the motor can output a constant maximum torque. The field weakening region occurs when the motor speed exceeds the rated speed and field weakening control is used to weaken the permanent magnet's magnetic field to maintain voltage balance. The constant power region is where the motor's output power remains constant at high speeds. Data is divided into these three regions using speed thresholds and voltage utilization parameters, forming regional datasets.

[0046] Targeted feature enhancement is required for data from different operating areas. For data in the constant torque zone, the original features can be retained due to the relatively simple control strategy. For data in the field weakening zone, a field weakening depth indicator is added. This indicator, calculated by calculating the ratio of the direct-axis current to the rated current, characterizes the degree of field weakening in the motor. For data in the constant power zone, a power fluctuation indicator is added. This indicator, calculated by calculating the ratio of the standard deviation to the mean power, characterizes the stability of the motor power. This feature enhancement allows the data to better reflect the control characteristics of each zone, forming an enhanced feature dataset.

[0047] Based on the enhanced feature dataset, the stable operating boundaries of each operating region need to be calculated. This calculation is primarily based on the motor's voltage and torque equations. Solving these equations determines the current range within which the motor can operate stably at different speeds. Specifically for the field weakening region, the critical point for field weakening control—the speed at which field weakening control begins—needs to be determined. It is also necessary to identify unstable operating regions, such as those prone to step loss in deep field weakening. These calculation results form the field weakening region operating boundary data.

[0048] The calculated magnetic field weakening boundary data is used to filter and purify the raw data. This primarily removes data points that fall within the unstable operating region and those that significantly deviate from the normal operating boundary. This filtering process can be accomplished by setting thresholds, such as marking data points that exceed a certain percentage of the stable operating boundary as abnormal and removing them from the dataset. This results in more reliable data and forms the basis for preliminary cleansing.

[0049] The data after preliminary cleansing is further denoised using a sliding window mean filter. Sliding window mean filtering is a commonly used digital signal processing method that takes the average of the current point and the number of points before and after it (a total of window length) 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 the two points before and after it (a total of five points). This simple and effective method can smooth random fluctuations in the data, eliminate high-frequency noise interference, and produce smoother denoised data.

[0050] The denoised data is normalized to its maximum and minimum values, a standard preprocessing step for deep learning. For each parameter, its maximum and minimum values across the entire dataset are found, and the original values are then linearly mapped to the interval [0, 1]. This normalization helps eliminate the effects of varying parameter scales, making neural network training more stable and efficient. All eigenvalues of the resulting normalized data are constrained to fall within the same range. The normalized data is randomly split in an 8:2 ratio, with 80% of the data used as training data for training the deep neural network and 20% as validation data for verifying model performance and adjusting hyperparameters. This split is a common practice in machine learning and ensures both a sufficient training set and a representative validation set.

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

[0052] Perform outlier detection on the raw data of weak magnetic operation, use the principle of three times standard deviation to identify data points that exceed the normal range, and obtain outlier marked data;

[0053] The raw data of the weak magnetic operation are screened according to the outlier mark data, and the data points marked as abnormal are replaced with the average value of the adjacent normal data points to obtain the preliminary cleaned data;

[0054] The preliminary cleaned data were processed using the median filter algorithm, and a seven-point window length was selected to filter each signal channel separately to obtain noise-suppressed data;

[0055] Wavelet transform is applied to the noise suppression data for decomposition, and five-layer Debesy wavelet is selected for multi-scale analysis to extract signal features and reconstruct them to obtain fine denoised data;

[0056] Calculate the maximum and minimum values of each parameter based on the refined denoising data, establish a normalized conversion parameter table, and obtain normalized benchmark data;

[0057] Using the normalized benchmark data, a maximum and minimum normalization transformation is performed on each parameter in the refined denoised data, and the value range is adjusted to the interval of zero to one to obtain normalized feature data;

[0058] Perform data balancing on the normalized feature data and focus on sampling the weak magnetic area data to ensure that the data proportion of each working area is balanced and obtain balanced feature data;

[0059] The balance characteristic data are divided into multiple data segments according to the time series, each segment contains the state parameters of twenty consecutive time points, forming standardized magnetic weakening control characteristic data.

[0060] Specifically, the raw field-weakening data is first detected for outliers using the triple standard deviation principle, a common statistical anomaly detection method. For each parameter's data series, the mean and standard deviation are calculated. Data points that deviate from the mean by more than three standard deviations are then labeled as outliers. The triple standard deviation principle is based on 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 outside this range are likely outliers. This method identifies outliers in the stator current, rotor position angle, voltage, and temperature data collected at various motor speeds, generating outlier marker data. Based on this outlier marker data, the raw field-weakening data is filtered and processed. Data points marked as outliers are not simply deleted but replaced with the average of the adjacent normal data points. This processing method takes into account the continuity of time series data. By taking the average of several normal points before and after the outlier as the replacement value, it maintains data smoothness while preserving the overall distribution of the data. The replaced dataset is called preliminary cleaned data. The dataset has removed obvious outliers but may still contain more random noise.

[0061] The preliminary cleaned data needs to be further processed for noise suppression, and the median filter algorithm is used here. Median filtering is a nonlinear filtering method. For each data point, the median value of all points in the window in which it is located, sorted by size, is taken as the filtering result. Selecting a seven-point window length means that for each point in the time series, the median value of the point and the three points before and after it, a total of seven points, is taken to replace the original value. Median filtering is particularly suitable for removing pulse noise and spike interference in motor operation data. At the same time, it can well maintain the edge characteristics of the data and will not cause blurring of the data edge like mean filtering. Median filtering is performed on each signal channel (such as stator current, rotor position angle, etc.) to obtain data with better noise suppression effect.

[0062] The noise-suppressed data is subjected to a more refined denoising process using wavelet transform decomposition technology. Wavelet transform is a time-frequency analysis method that can provide the local characteristics of the signal in the time domain at different frequencies. A five-layer Debetsy wavelet (usually referred to as DB wavelet) is selected for multi-scale analysis. The Debetsy wavelet is a wavelet basis function with good orthogonality and tight support, and is particularly suitable for analyzing the non-stationary characteristics of motor operation data. Five-layer decomposition means decomposing the signal into wavelet coefficients at five different frequency scales, including one approximate coefficient and five detail coefficients. Wavelet transform decomposition can be expressed as:

[0063]

[0064] Among them, S(t) represents the original signal, A5(t) represents the fifth layer approximation coefficient, and D j (t) represents the detail coefficient of the jth layer. During the denoising process, thresholding is applied to the detail coefficients of each layer to retain the key feature information and suppress the influence of noise. The signal is then reconstructed to obtain refined denoised data. Based on the refined denoised data, the maximum and minimum values of each parameter are calculated to create a normalization conversion parameter table. The normalization conversion parameter table contains the name, original maximum and minimum values, and corresponding normalization coefficients of each parameter. These parameters will be used for the normalization of all subsequent data to ensure consistency in data processing. The normalized baseline data serves as the basis for subsequent normalization operations and is an important reference for data preprocessing during model deployment.

[0065] Using normalized benchmark data, we perform a maximum-minimum normalization transformation on each parameter in the refined denoised data. Maximum-minimum normalization is a linear transformation that linearly maps the original data to the interval [0, 1]. This transformation brings parameters of different dimensions and orders of magnitude into the same range, facilitating neural network training. The normalized data is called normalized feature data.

[0066] Data balancing of the 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 data from the constant-torque region, which can lead to insufficient prediction accuracy for the trained model in this region. By focusing on the field-weakening region data and employing oversampling techniques such as the Synthetic Minority Over-sampling Technique (SMOTE) algorithm to generate new field-weakening region samples, the data distribution across all operating regions is balanced. This data balancing process generates balanced feature data. The balanced feature data is then divided into multiple data segments based on the time series, with each segment containing state parameters at twenty consecutive time points. This processing approach accounts for the temporal relationships in motor control and enables the model to learn the dynamic characteristics of parameters over time. Each data segment contains both the current state and the evolution of previous states, helping the model accurately predict the motor's dynamic response. The resulting data is called normalized field-weakening control feature data and serves as input for subsequent deep learning model training.

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

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

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

[0070] 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% to obtain an anti-overfitting network structure.

[0071] A funnel-type decreasing activation function is configured for the hidden layer neurons in the anti-overfitting network structure, and a funnel-type decreasing activation function with a parameter of 0.01 is used to obtain a nonlinear mapping unit;

[0072] According to the characteristic distribution of the standardized weak magnetic control characteristic data, a linear activation function is set for the output layer, and the weighted mean square error is defined as the loss function to obtain the training objective function;

[0073] Set weights for each performance indicator in the training objective function. The system stability margin weight is 0.3, the dynamic response time weight is 0.2, the efficiency weight is 0.3, and the temperature rise weight is 0.2. The weighted training objective is obtained.

[0074] Based on the weighted training objective, the 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;

[0075] The neural network is trained using the network training strategy. The training cycle is set to 300, and the early stopping strategy is adopted. When the validation set loss does not improve for 15 consecutive cycles, the training is stopped to obtain the initial training model.

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

[0077] Specifically, 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 directly corresponds to the dimension of the standardized weak magnetic control feature data. For the weak magnetic control of permanent magnet motors, the input features include multiple parameters such as stator current, rotor position angle, motor speed, stator voltage, weak magnetic current component, torque current component, etc. The feature dimension after data preprocessing is usually between 30 and 50. The five hidden layers are set with a gradually decreasing number of neurons, namely 256, 192, 128, 96, and 64, forming a structure that gradually narrows from input to output. The output layer is designed to have 4 neurons, corresponding to the four key performance indicators of system stability margin, dynamic response time, efficiency, and temperature rise. This hierarchical and well-structured network architecture becomes the basis for subsequent optimization.

[0078] The hidden layers of the neural network infrastructure adopt a funnel-shaped structure. This refers to a network design in which the number of neurons gradually decreases from the input layer to the output layer. The advantage of this funnel-shaped structure is that the front layers can extract rich features from the data, while the back layers focus on filtering and fusing these features, achieving effective dimensionality reduction and abstraction. Batch normalization and dropout layers are also added after each hidden layer. The batch normalization layer normalizes each batch of data, keeping the data distribution within the sensitive range of the activation function and accelerating network training. The dropout layer prevents the network from overfitting to the training data by randomly shutting down a certain percentage of neurons. Setting the dropout rate to 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 training data.

[0079] A funnel-type decreasing activation function is configured for the hidden layer neurons in the anti-overfitting network structure. The funnel-type decreasing activation function here refers to the Leaky ReLU (leaky rectified 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, the LeakyReLU still has a small slope when the input is negative. This prevents the gradient from completely disappearing during backpropagation, helping to solve the gradient vanishing problem in deep network training. A parameter of 0.01 means that on the negative half-axis, the slope of the function is 0.01. This is an empirical setting that can effectively prevent neuron "death" (i.e., some neurons are never activated) while maintaining the nonlinear characteristics of the activation function.

[0080] Based on the characteristic distribution of the standardized weak magnetic control feature data, a linear activation function is set for the output layer instead of the commonly used nonlinear activation function. The linear activation function directly outputs the weighted sum of neurons and is suitable for predicting continuous numerical outputs, such as performance indicators such as system stability margin and dynamic response time. At the same time, the weighted mean square error is defined as the loss function. The weighted mean square error is an improved version of the standard mean square error that takes into account the importance weights of different output indicators. The standard mean square error calculates the sum of the squares of the differences between the predicted value and the true value, while the weighted mean square error assigns different weight coefficients to the errors of different indicators, thereby paying more attention to the prediction accuracy of important indicators during the training process.

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

[0082] Based on the weighted training objective, the Adaptive Moment Estimation (Adam) algorithm with a learning rate of 0.001 was selected as the optimizer. The Adam algorithm combines the advantages of the momentum method and the RMSProp algorithm. It can adaptively adjust the learning rate of each parameter and has good adaptability to gradients of different scales. A learning rate of 0.001 is a conservative choice that ensures a stable training process and a moderate convergence rate. The batch size is set to 64, which means that each gradient update is calculated using 64 samples. This batch size strikes a good balance between training efficiency and memory usage.

[0083] The neural network was trained using the aforementioned network training strategy, with a training epoch of 300 and an early stopping strategy. Early stopping is a common technique for preventing overfitting. It monitors performance metrics on the validation set and terminates training early if performance stops improving. Specifically, training is stopped when the validation set loss has not improved for 15 consecutive epochs. This ensures that the model is fully trained while preventing a loss in generalization due to overtraining. The resulting model is called the initial training model.

[0084] The initial trained model is performance-tested, and the prediction error on the test set is calculated to assess the model's generalization ability. The network structure is then fine-tuned using Bayesian hyperparameter optimization. Bayesian optimization is a black-box optimization method based on probabilistic models that efficiently searches for the optimal configuration 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 repeatedly trying different hyperparameter combinations, the configuration with the best performance on the validation set is found. This series of construction, training, and optimization processes ultimately results in a deep neural network model capable of accurately predicting magnetic field weakening control performance.

[0085] Taking the optimization of flux-weakening control for 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, including 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. Predictions are calculated through forward propagation, and then the weighted mean squared error (MSE) loss is calculated. For example, for a given 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°C (true value 35°C). The weighted mean squared error (MSE) 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. During training, the validation set loss is monitored. After the 178th epoch, the validation set loss has not decreased for 15 consecutive epochs, triggering the early stopping mechanism to terminate training. Finally, through Bayesian optimization, it was found that after adjusting the number of neurons in the fourth hidden layer from 96 to 112 and the dropout rate from 20% to 15%, the performance of the model on the test set was improved. This configuration was finally determined as the final structure of the weak magnetic control performance prediction model.

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

[0087] The motor's current speed, load torque, and voltage limit conditions are input 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 performance characteristic data;

[0088] A multi-objective optimization function is set for the operating performance characteristic data, with maximizing the stability margin, minimizing the response time, maximizing the efficiency, and minimizing the temperature rise as the optimization objectives, and the objective function expression is obtained;

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

[0090] An improved covariance matrix adaptive evolutionary strategy algorithm was constructed based on the comprehensive optimization objective, with the population size set to 50, the initial step size to 0.3, and the number of parents to 25, to obtain a parameter optimization algorithm.

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

[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, and the current reference value configuration scheme is obtained;

[0093] According to the current reference value configuration scheme, a correction formula for the direct-axis current reference value and a correction formula for the quadrature-axis current reference value are established, which are respectively expressed as a function of the direct-axis current and the speed and a function of the quadrature-axis current and the torque demand, thereby obtaining a current reference value calculation model.

[0094] The current reference value calculation model is verified and tested, the current reference value is calculated under different working conditions, and boundary correction is performed based on the verification results to form the basic parameters of weak magnetic control.

[0095] Specifically, a deep neural network constructs a precise performance mapping model using key motor operating parameters as input. Input parameters include the current motor speed, load torque, and voltage constraints. These parameters undergo complex nonlinear feature extraction and mapping, transforming them into predicted values for four core performance indicators. The first stage of data processing involves the deep neural network's feature extraction mechanism. The network utilizes a five-layer hidden layer, each equipped with a funnel-shaped decreasing activation function, to achieve progressive abstraction and transformation of the input data. Specifically, the input layer receives the raw parameters, which are then processed hierarchically with 256, 192, 128, 96, and 64 neurons. Batch normalization and dropout strategies are used in each layer to enhance the model's generalization capabilities. This process is essentially a complex data dimensionality reduction and feature reconstruction technique that captures subtle nonlinear correlations between motor operating parameters. The multi-objective optimization objective is constructed as a comprehensive balance between system stability margin, dynamic response time, efficiency, and temperature rise. Through weighted single-objective function conversion, stability margin and efficiency each receive a weight of 0.3, while response time and temperature rise each receive a weight of 0.2. This weight design not only reflects the differentiated importance of various performance indicators in engineering practice, but also provides an accurate objective function for subsequent optimization.

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

[0097] In a 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, including the four variables of the weak magnetic field controller proportional gain adjustment, integral gain adjustment, direct-axis current reference value correction coefficient, and quadrature-axis current reference value correction coefficient, to obtain the control action vector;

[0100] The state observation vector and the basic parameters of magnetic weakening control are combined to construct the state-action mapping relationship and design the reward function to obtain the reinforcement learning environment model;

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

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

[0103] The exploration strategy is designed according to the network training rules. The initial exploration rate is set to 1.0, which decays to a minimum value of 0.01 according to the number of steps. The decay rate is 0.995, and the discount factor is set to 0.95, resulting in an exploration-exploitation balance strategy.

[0104] The Q-value estimation network and the exploration-exploitation balance strategy were combined and trained for 500,000 steps in a motor simulation environment to obtain a 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 amount of the controller's proportional gain and integral gain, and form dynamic control parameters.

[0106] Specifically, the state space is defined, consisting of seven key variables: motor speed represents the current operating speed of the motor; direct-axis current is the current component along the magnetic field in permanent magnet synchronous motor vector control, used for field weakening control; quadrature-axis current is the current component perpendicular to the magnetic field and primarily controls motor torque; direct-axis voltage and quadrature-axis voltage represent the two orthogonal components of the stator voltage in the rotating coordinate system; direct-axis voltage margin refers to the margin between the direct-axis voltage and the system's maximum allowable voltage, reflecting the system's tolerance to voltage fluctuations; and the system stability margin quantifies how close the system is to an unstable state under the current operating conditions. These seven variables constitute the state observation vector, which comprehensively describes the operating state of the motor's field weakening control system. Four core control variables were selected to define the action space: the field weakening controller proportional gain adjustment, which represents the real-time adjustment of the proportional controller gain; the integral gain adjustment, which represents the real-time adjustment of the integral controller gain; the direct-axis current reference correction factor, which modifies the base direct-axis current reference value for more precise field weakening control; and the quadrature-axis current reference correction factor, which modifies the base quadrature-axis current reference value to optimize torque output. These four variables together constitute the control action vector, which directly acts on the key parameter adjustment of the weak magnetic control system.

[0107] By combining the state observation vector and the basic parameters of magnetic-weakening control, a state-action mapping is constructed. This mapping represents the system's response to a control action taken under a specific state. A reward function is also designed, which comprehensively considers system stability, dynamic response speed, efficiency, and temperature rise. Positive rewards are given for control actions that improve stability, shorten response time, increase efficiency, and reduce temperature rise, while negative rewards are given for actions that do not. By defining 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 deep Q-network (Double DQN) with a dual network structure is used for parameter optimization. The deep Q-network is a type of reinforcement learning algorithm that uses a neural network to approximate the Q-value function to learn the optimal strategy. The dual network structure consists of two parts: an evaluation network and a target network. The two network structures are exactly the same, but the parameter update frequency is different. Both networks use a four-layer fully connected neural network. The input layer corresponds to the state space dimension, and the output layer corresponds to the action space dimension. There are three hidden layers in the middle, with the number of neurons being 128, 96, and 64 respectively, forming a layer-by-layer 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 to stabilize the training process.

[0109] An update mechanism is implemented for the Q-value estimation network. The evaluation network is updated immediately after each interaction with the environment, while the target network parameters are updated every 100 steps, copied from the evaluation network parameters. This asynchronous update mechanism effectively prevents oscillations and divergence during training. The experience replay buffer is a crucial component of deep reinforcement learning, storing historical data on the agent's interactions with the environment, including a four-tuple consisting of state, action, reward, and next state. The experience replay buffer is set to a capacity of 10,000, meaning that the most recent 10,000 steps of interaction data are stored. A batch of data is randomly sampled from this data for training during each network update, breaking the temporal correlation between the 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 and full exploration of the environment. As training steps increase, the exploration rate decreases at a decay rate of 0.995, ultimately reaching a minimum of 0.01, at which point the agent selects the currently estimated optimal action with 99% probability and randomly explores with 1% probability. A discount factor of 0.95, representing the rate at which future rewards are discounted, controls the algorithm's trade-off between short-term and long-term gains. This balanced exploration-exploitation strategy ensures that the algorithm fully explores the action space while gradually converging to the optimal strategy.

[0111] A Q-value estimation network combined with an exploration-exploitation balance strategy was trained in a motor simulation environment for a total of 500,000 interaction steps. In each interaction step, the agent observed the current state and selected an action based on the exploration strategy. The environment executed the action, transitioned to the new state, and provided a reward. The agent stored this experience in a replay buffer. A batch of experience data was then randomly sampled from the buffer to calculate the target Q-value and update the evaluation network parameters. In this way, the Q-network gradually learned the value of various control actions under 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. During each control cycle, the current motor state data, including speed, current, voltage, and other parameters, is first acquired to form a state observation vector. This vector is then fed 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 a correction factor for the current reference. Finally, the controller parameters and current reference are modified based on these adjustments, achieving 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 motor speed is detected to be 4500 rpm (1.5 times the rated speed), the direct-axis current is -40 A, the quadrature-axis current is 60 A, the direct-axis voltage is 150 V, the quadrature-axis voltage is 200 V, the direct-axis voltage margin is 50 V, 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 four control adjustment value estimates 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 correction factor of 1.05, and a quadrature-axis current reference correction factor of 0.98. This means that in the current state, the proportional gain should be appropriately increased, the integral gain should be decreased, the negative value of the direct-axis current should be slightly increased to enhance the field weakening effect, and the quadrature-axis current should be slightly reduced to optimize 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 increased to 0.35, and the dynamic response became faster, achieving the effect of intelligent adaptive control.

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

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

[0116] The comprehensive control parameters are configured into the motor flux weakening control system, and the controller proportional gain and integral gain are updated to obtain a parameter update control system;

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

[0118] Calculate system performance indicators using operational feedback data, including stability margin, dynamic response time, efficiency, and temperature rise, and obtain actual performance indicator values;

[0119] Compare the actual performance index value with the predicted value of the field weakening control performance prediction model, calculate the prediction error, and obtain the model accuracy evaluation data;

[0120] A data buffer pool is set up based on the model accuracy evaluation data, and the latest 1,000 sets of data are stored according to the first-in-first-out principle to obtain an incremental training data set;

[0121] Merge the incremental training data set with the original training data, filter and retain representative samples to obtain an updated training data set;

[0122] The updated training data set is used to perform incremental training on the flux weakening control performance prediction model. The learning rate is set to 0.0005 and the iteration cycle is 50, forming a closed-loop optimization control method for the flux weakening motor.

[0123] Specifically, the dynamic control parameters and the basic parameters of the weak magnetic control are not simply superimposed linearly, but are deeply coupled through a complex weight distribution algorithm based on information entropy. Each parameter is regarded as a multidimensional feature vector, and its weight is determined by the instantaneous state of the motor operation, historical performance data, and the confidence 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 regarded as orthogonal basic characteristics of the motor operating state. Each sampling point is not just a numerical value, but also a multidimensional information carrier that contains rich dynamic characteristics of the motor operation.

[0124] Calculating performance metrics involves 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; rather, it is achieved through a series of highly nonlinear signal transformation and feature mapping algorithms. This algorithm must simultaneously consider instantaneous sampled data, historical operating data, and the motor's physical characteristic model to construct a multidimensional performance evaluation hyperspace. Model accuracy assessment is a complex comparative analysis process based on information theory. Error analysis between actual measured metrics and model predictions goes beyond simple numerical comparisons; it uses advanced statistical metrics such as information divergence and entropy change in the multidimensional error space for a comprehensive assessment. Each error is not just a numerical value; it is a quantitative indicator of the model's cognitive capabilities. The design of the data buffer pool embodies a dynamic learning strategy based on information value. The first-in-first-out mechanism is more than just simple data storage; it involves a complex data value assessment and screening process. Using multi-dimensional metrics such as information entropy, data sparsity, and representativeness, the most cognitively valuable samples are extracted from 1,000 data sets. This screening mechanism ensures that the model continuously learns the most critical and information-dense operating data. Incremental model training 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, self-organizing learning process in which the model continuously adjusts its cognitive boundaries through tiny parameter perturbations.

[0125] For example, under specific motor operating conditions, the algorithm captures a multi-dimensional set of operating data. This data includes not only 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 converted into predicted values for performance indicators. When there is a slight deviation between the predicted value and the actual measured value, this deviation itself becomes a key source of information for further model learning and optimization.

[0126] The above describes the weak magnetic motor optimization method based on deep learning in the embodiment of the present application. The following describes the weak magnetic motor optimization system based on deep learning in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of a weak magnetic motor optimization system based on deep learning includes:

[0127] The acquisition module is used to collect the stator current, rotor position angle, voltage and temperature data of the permanent magnet motor at different speeds to obtain the original data of the weak magnetic operation;

[0128] A processing module is used to perform denoising and maximum and minimum value normalization on the raw data of the weak magnetic operation to obtain standardized weak magnetic control characteristic data;

[0129] A training module, configured to input the standardized magnetic weakening control characteristic data into a five-hidden-layer deep neural network for training to obtain a magnetic weakening control performance prediction model;

[0130] an adjustment module, configured to perform parameter adjustment calculation on a direct-axis current reference value and a quadrature-axis current reference value according to an output result of the flux-weakening control performance prediction model, so as to obtain basic flux-weakening control parameters;

[0131] A calculation module is used to calculate the real-time adjustment amount of the controller proportional gain and integral gain by using the basic parameters of the weak magnetic control in combination with the motor operating state data to obtain dynamic control parameters;

[0132] The optimization module is used to apply the dynamic control parameters to the motor weakening control system and optimize the weakening control performance prediction model through feedback data loop to form a weakening motor closed-loop optimization control method.

[0133] Through the collaborative efforts of these components, a high-frequency (10kHz) data acquisition strategy is 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-hidden layer structure of the deep neural network, with a decreasing number of neurons at each layer, enables multi-scale feature extraction of motor operating data, effectively capturing the motor's complex nonlinear dynamic characteristics. Artificial intelligence algorithms play a key role in this solution, especially the introduction of deep learning and reinforcement learning models, which transcend the static parameter limitations of traditional control methods and build 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 for stability margin and efficiency, 0.2 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 in a complex high-dimensional parameter space, effectively avoiding falling into local optimal solutions. The design of the deep reinforcement learning model has taken controller adaptive adjustment to a whole new level. By defining a detailed state space and action space, an intelligent system capable of learning and adjusting control strategies in real time has been constructed. The attention mechanism combined with the state prediction model of the long short-term memory network (LSTM) further enhances the system's predictive control capabilities and can identify potential unstable working conditions in advance. The performance indicators of the motor have been significantly improved, and high-performance and high-efficiency operation of the motor over a wide speed range has been achieved.

[0135] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may 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 of the computer design is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, the above method is implemented.

[0136] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0137] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which implements the above-described method when executed by a processor. 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 the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for optimizing a weak magnetic motor based on deep learning, characterized in that: The deep learning-based flux weakening motor optimization method includes: The stator current, rotor position angle, voltage and temperature data of the permanent magnet motor at different speeds are collected to obtain the original data of weak magnetic operation; De-noising and maximum and minimum value normalization processing are performed on the raw magnetic weakening operation data to obtain standardized magnetic weakening control characteristic data; Inputting the standardized magnetic weakening control characteristic data into a five-hidden-layer deep neural network for training to obtain a magnetic weakening control performance prediction model; According to the output result of the field weakening control performance prediction model, parameter adjustment calculation is performed on the direct-axis current reference value and the quadrature-axis current reference value to obtain the field weakening control basic parameters; The basic parameters of the weak magnetic control are combined with the motor operating status data to calculate the real-time adjustment amount of the controller proportional gain and integral gain to obtain the dynamic control parameters; The dynamic control parameters are applied to the motor flux weakening control system, and the flux weakening control performance prediction model is optimized through feedback data loop to form a flux weakening motor closed-loop optimization control method.

2. The method for optimizing a magnetic field-weakening motor based on deep learning according to claim 1, characterized in that: The stator current, rotor position angle, voltage and temperature data of the permanent magnet motor at different speeds are collected to obtain the original data of the weakening magnetic operation, including: Multiple test points are set within the range of 0.2 to 3 times the rated speed of the motor. Data including stator current, rotor position angle, motor speed, stator voltage, flux-weakening current component, torque current component, direct-axis current, quadrature-axis current, power factor, and efficiency index are collected at each test point to obtain raw data under multiple operating conditions. The raw data of the multiple working conditions are classified according to the working area of the motor, and divided into three types of data: constant torque area, weak magnetic area and constant power area, to obtain a regional data set; The original features of the data in the constant torque region are maintained, a weak magnetic depth index is added to the data in the weak magnetic region, and a power fluctuation index is added to the data in the constant power region to obtain an enhanced feature data set; Calculating the stable working boundary of each working area based on the enhanced feature data set, determining the critical point and unstable operating area of the weakening magnetic area, and obtaining the operating boundary data of the weakening magnetic area; The original collected data is screened using the weak magnetic region operation boundary data to remove abnormal data points that do not meet stable operation conditions, thereby obtaining preliminary purified data; Applying a sliding window mean filtering algorithm to the preliminary cleaned data, setting the window length to 5 sampling points, eliminating high-frequency noise interference, and obtaining denoised data; Normalizing the denoised data according to the maximum and minimum values of each parameter, performing maximum and minimum normalization conversion on each parameter to obtain normalized data; The normalized data is randomly divided in a ratio of 8:2, 80% of the data is divided into training data, and 20% of the data is divided into verification data, to form the original data of weak magnetic operation.

3. The method for optimizing a magnetic field-weakening motor based on deep learning according to claim 1, characterized in that: The de-noising and maximum and minimum value normalization processing is performed on the raw magnetic weakening operation data to obtain standardized magnetic weakening control characteristic data, including: Performing outlier detection on the raw magnetic weakening operation data, identifying data points that exceed the normal range using the triple standard deviation principle, and obtaining outlier marked data; The raw data of the weak magnetic operation are screened according to the abnormal value marking data, and the data points marked as abnormal are replaced with the average value of the adjacent normal data points to obtain preliminary cleaned data; The preliminary cleaned data is processed using a median filter algorithm, and a seven-point window length is selected to filter each signal channel separately to obtain noise-suppressed data; Applying wavelet transform decomposition to the noise suppression data, selecting five-layer Debecy wavelet to perform multi-scale analysis, extracting signal features and reconstructing to obtain fine denoised data; Calculating the maximum and minimum values of each parameter based on the refined denoised data, establishing a normalized conversion parameter table, and obtaining normalized benchmark data; Using the normalized reference data, performing a maximum-minimum normalization transformation on each parameter in the refined denoised data, adjusting the value range to an interval of zero to one, and obtaining normalized feature data; Performing data balancing processing on the normalized characteristic data, focusing on sampling the weak magnetic area data, ensuring that the data proportions of each working area are balanced, and obtaining balanced characteristic data; The balance characteristic data is divided into a plurality of data segments according to a time series, each segment contains state parameters of twenty consecutive time points, and standardized magnetic weakening control characteristic data is formed.

4. The method for optimizing a weak magnetic motor based on deep learning according to claim 1, characterized in that: The standardized magnetic weakening control characteristic data is input into a five-layer hidden layer deep neural network for training to obtain a magnetic weakening control performance prediction model, including: Construct a deep neural network structure consisting of an input layer, five hidden layers, and an output layer. The number of neurons in the input layer is the same as the feature dimension, the number of neurons in the five hidden layers is 256, 192, 128, 96, and 64 respectively, and the number of neurons in the output layer is 4, thus obtaining the basic architecture of the neural network. Setting a funnel-shaped structure for the hidden layers of the neural network infrastructure, and adding a batch normalization layer and a dropout layer after each layer, with a dropout rate set to 20%, to obtain an anti-overfitting network structure; A funnel-shaped decreasing activation function is configured for hidden layer neurons in the anti-overfitting network structure, and a funnel-shaped decreasing activation function with a parameter of 0.01 is used to obtain a nonlinear mapping unit; According to the characteristic distribution of the standardized magnetic weakening control characteristic data, a linear activation function is set for the output layer, and a weighted mean square error is defined as a loss function to obtain a training objective function; Setting weights for each performance indicator in the training objective function: the system stability margin weight is 0.3, the dynamic response time weight is 0.2, the efficiency weight is 0.3, and the temperature rise weight is 0.2, to obtain a 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 a network training strategy; The neural network is trained using the network training strategy, with a training cycle of 300 and an early stopping strategy. Training is stopped when the validation set loss does not improve for 15 consecutive cycles to obtain an initial training model. The performance of the initial training model is tested, the prediction error on the test set is calculated, and the network structure is fine-tuned through Bayesian hyperparameter optimization to form a weak magnetic control performance prediction model.

5. The method for optimizing a weak magnetic motor based on deep learning according to claim 1, characterized in that: The method of performing parameter adjustment calculation on the direct-axis current reference value and the quadrature-axis current reference value according to the output result of the field weakening control performance prediction model to obtain the field weakening control basic parameters includes: Inputting the current speed, load torque and voltage limit conditions of the motor into the magnetic field weakening control performance prediction model, obtaining the predicted values of four performance indicators of system stability margin, dynamic response time, efficiency and temperature rise, and obtaining working condition performance characteristic data; Setting a multi-objective optimization function for the operating performance characteristic data, taking maximizing stability margin, minimizing response time, maximizing efficiency, and minimizing temperature rise as optimization objectives, and obtaining an objective function expression; The objective function expression is converted into a weighted single objective function, and the weight coefficients of the four indicators are set to 0.3, 0.2, 0.3 and 0.2 respectively to obtain a comprehensive optimization target; Based on the comprehensive optimization objective, an improved covariance matrix adaptive evolutionary strategy algorithm is constructed, and the population size, initial step size, and number of parents are set to 50, 0.3, and 25, respectively, to obtain a parameter optimization algorithm. The parameter optimization algorithm is used to iteratively optimize the direct-axis current reference value and the quadrature-axis current reference value, with the number of iterations set to 500, to obtain a Pareto optimal solution set; Analyzing the Pareto optimal solution set, extracting the optimal configuration of the direct-axis current reference value and the quadrature-axis current reference value, and obtaining a current reference value configuration scheme; According to the current reference value configuration scheme, a direct-axis current reference value correction formula and a quadrature-axis current reference value correction formula are established, which are respectively expressed as a function of the direct-axis current and the speed and a function of the quadrature-axis current and the torque demand, to obtain a current reference value calculation model; The current reference value calculation model is verified and tested, the current reference value is calculated under different working conditions, and boundary correction is performed according to the verification results to form basic parameters for weak magnetic control.

6. The method for optimizing a weak magnetic motor based on deep learning according to claim 1, characterized in that: The method of calculating the real-time adjustment amount of the controller proportional gain and integral gain by using the basic parameters of the weak magnetic control in combination with the motor operating status data to obtain the 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, including the four variables of the weak magnetic field controller proportional gain adjustment, integral gain adjustment, direct-axis current reference value correction coefficient, and quadrature-axis current reference value correction coefficient, to obtain the control action vector; Combining the state observation vector and the basic magnetic weakening control parameters, constructing a state-action mapping relationship, and designing a reward function 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 a Q value estimation network. An update mechanism is set for the Q-value estimation network, where the evaluation network is updated once per step, the target network is updated once every 100 steps, and the capacity of the experience replay buffer is set to 10,000, to obtain the network training rules; An exploration strategy was designed based on the network training rules, with an initial exploration rate set to 1.0, decayed to a minimum value of 0.01 according to the number of steps, a decay rate of 0.995, and a discount factor set to 0.95, resulting in an exploration-exploitation balance strategy; Combining the Q-value estimation network with the exploration-exploitation balance strategy, performing 500,000 training iterations in a motor simulation environment, and obtaining a 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 amount of the controller proportional gain and integral gain, and form dynamic control parameters.

7. The method for optimizing a weak magnetic motor based on deep learning according to claim 1, characterized in that: The method of applying the dynamic control parameters to the motor flux weakening control system and optimizing the flux weakening control performance prediction model through feedback data loop to form a flux weakening motor closed-loop optimization control method includes: The dynamic control parameters are integrated with the basic parameters of the field-weakening control, and the final control parameters are calculated using a weighted average method to obtain the comprehensive control parameters. Configuring the comprehensive control parameters into the motor flux weakening control system, updating the controller proportional gain and integral gain, and obtaining a parameter update control system; The operating status of the parameter update control system is monitored in real time, with a sampling frequency of 10 kHz, and stator current, voltage, speed and temperature are collected to obtain operating feedback data; Calculating system performance indicators using the operational feedback data, including stability margin, dynamic response time, efficiency, and temperature rise, to obtain actual performance indicator values; Comparing the actual performance index value with the predicted value of the field weakening control performance prediction model, calculating the prediction error, and obtaining model accuracy evaluation data; A data buffer pool is set up according to the model accuracy evaluation data, and the latest 1000 sets of data are stored according to the first-in-first-out principle to obtain an incremental training data set; Merging the incremental training data set with the original training data, screening and retaining representative samples, and obtaining an updated training data set; The updated training data set is used to perform incremental training on the flux-weakening control performance prediction model, with a learning rate set to 0.0005 and an iteration cycle of 50, to form a flux-weakening motor closed-loop optimization control method.

8. A deep learning-based flux weakening motor optimization system, used to implement the deep learning-based flux weakening motor optimization method according to any one of claims 1 to 7, characterized in that: The deep learning-based flux weakening motor optimization system includes: The acquisition module is used to collect the stator current, rotor position angle, voltage and temperature data of the permanent magnet motor at different speeds to obtain the original data of the weak magnetic operation; a processing module, configured to perform denoising and maximum and minimum value normalization processing on the raw magnetic weakening operation data to obtain standardized magnetic weakening control characteristic data; A training module, configured to input the standardized magnetic weakening control characteristic data into a five-hidden-layer deep neural network for training to obtain a magnetic weakening control performance prediction model; an adjustment module, configured to perform parameter adjustment calculation on a direct-axis current reference value and a quadrature-axis current reference value according to an output result of the flux-weakening control performance prediction model, so as to obtain basic flux-weakening control parameters; A calculation module is used to calculate the real-time adjustment amount of the controller proportional gain and integral gain by using the basic parameters of the weak magnetic control in combination with the motor operating state data to obtain dynamic control parameters; The optimization module is used to apply the dynamic control parameters to the motor weakening control system and optimize the weakening control performance prediction model through feedback data loop to form a weakening motor closed-loop optimization control method.

9. A computer device, characterized in that: It includes a memory and a processor, the memory stores a computer program that can be run on the processor, and the processor implements the deep learning-based weak magnetic motor optimization method described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that A computer program is stored thereon, and when the computer program is executed by a processor, the processor is caused to execute the deep learning-based flux weakening motor optimization method according to any one of claims 1 to 7.

Images