Permanent magnet motor energy efficiency optimization and energy-saving control method and system based on artificial intelligence

Abstract

The invention provides an artificial intelligence-based permanent magnet motor energy efficiency optimization and energy-saving control method and system, and relates to the technical field of artificial intelligence, and the method comprises the steps: collecting motor operation data through employing a multi-dimensional sensor array, carrying out the feature extraction based on a mutual information entropy criterion, constructing a multi-physics field coupling model, and building a digital twinborn model; a double-layer deep reinforcement learning neural network is constructed based on the model, a global optimization target is generated through a high-layer strategy network, and a low-layer execution network adopts a deep deterministic strategy gradient algorithm to decompose sub-targets; training the network by adopting a priority experience playback mechanism to obtain an optimal control strategy; and inputting the optimized parameters into a vector controller to realize three-closed-loop control. The operation efficiency of the permanent magnet motor is improved, and the system stability and the dynamic response performance are enhanced.

Description

Technical Field

[0001] The present invention relates to artificial intelligence technology, and in particular to an artificial intelligence-based permanent magnet motor energy efficiency optimization and energy-saving control method and system. Background Art

[0002] Permanent magnet motors are widely used in industry and transportation due to their high efficiency, high power density and good dynamic performance. With the continuous improvement of energy efficiency and environmental protection requirements, higher requirements are placed on energy efficiency optimization and energy-saving control of permanent magnet motors. Traditional permanent magnet motor control methods mainly rely on fixed mathematical models and empirical parameters, which are difficult to adapt to complex and changing operating environments and load conditions.

[0003] In recent years, with the rapid development of artificial intelligence technology, intelligent control methods based on machine learning and deep learning have provided new ideas and means for optimizing the energy efficiency of permanent magnet motors. However, the existing intelligent control methods still have some limitations and shortcomings.

[0004] First, existing permanent magnet motor modeling methods often only consider a single physical field, which makes it difficult to fully reflect the complex electromagnetic, thermal and mechanical coupling effects of the motor during actual operation, resulting in limited control accuracy and reliability.

[0005] Secondly, traditional reinforcement learning algorithms often face problems of slow convergence and low sample utilization efficiency when dealing with complex control problems such as permanent magnet motors with continuous states and action spaces, making it difficult to achieve fast and accurate real-time control.

[0006] Finally, the existing permanent magnet motor vector control method usually adopts a fixed-gain PI controller, which is difficult to adapt to motor parameter changes and external disturbances, especially in the weak magnetic control area. It is difficult to achieve accurate tracking and control of the magnetic flux, affecting the overall performance and efficiency of the motor. Summary of the invention

[0007] The embodiments of the present invention provide a permanent magnet motor energy efficiency optimization and energy-saving control method and system based on artificial intelligence, which can solve the problems in the prior art.

[0008] According to a first aspect of the embodiments of the present invention,

[0009] Provides an AI-based permanent magnet motor energy efficiency optimization and energy-saving control method, including:

[0010] A multi-dimensional sensor array is used to collect real-time operation data of the permanent magnet motor during operation, and feature extraction and dimensionality reduction processing are performed on the real-time operation data based on the mutual information entropy criterion to obtain the motor characteristic parameters of the permanent magnet motor; a multi-physics field coupling model including electromagnetic field distribution, thermal field distribution, and stress field distribution is constructed; the motor characteristic parameters are dynamically mapped and matched with the multi-physics field coupling model to establish a permanent magnet motor digital twin model with adaptive feature fusion capability;

[0011] A double-layer deep reinforcement learning neural network is constructed based on the permanent magnet motor digital twin model, and the double-layer deep reinforcement learning neural network includes a high-level policy network and a low-level execution network; the high-level policy network adopts a long short-term memory network structure to generate a global optimization target of the permanent magnet motor; the low-level execution network adopts a deep deterministic policy gradient algorithm to decompose the global optimization target into multiple sub-targets; a reward function is constructed for each sub-target, and the reward function includes an energy efficiency index, a dynamic response index and a stability index; a priority experience replay mechanism is used to train the double-layer deep reinforcement learning neural network to obtain an optimal flux control strategy and an optimal torque control strategy; based on the optimal flux control strategy and the optimal torque control strategy, an adaptive particle swarm algorithm is used to perform multi-objective optimization on control parameters;

[0012] The optimized control parameters are input into a permanent magnet motor vector controller with a model prediction function; the permanent magnet motor vector controller is based on a sliding mode variable structure control principle to construct a three-closed-loop control structure including a torque loop, a flux loop and a current loop; the deviation between the flux observation value and the given value of the permanent magnet motor is calculated in real time, and an adaptive fuzzy neural network is used to compensate the d-axis current in real time.

[0013] Constructing a multi-physics field coupling model including electromagnetic field distribution, thermal field distribution, and stress field distribution; dynamically mapping and matching the motor characteristic parameters with the multi-physics field coupling model, and establishing a permanent magnet motor digital twin model with adaptive feature fusion capability, including:

[0014] The stator current and the rotor position in the characteristic parameters of the motor are input into the Maxwell equations, the calculation domain is divided into a plurality of grid units by using the finite element method, and the magnetic flux density of the grid units is solved by Newton-Raphson iteration; the eddy current loss and the hysteresis loss of each grid unit are calculated based on the magnetic flux density to obtain the electromagnetic loss distribution;

[0015] The electromagnetic loss distribution is combined with the temperature signal in the motor characteristic parameter to establish a heat conduction equation; the heat conduction equation is solved to obtain the instantaneous temperature distribution of each grid unit; based on the instantaneous temperature distribution, the thermal stress distribution of each grid unit is calculated by the thermal expansion coefficient;

[0016] The electromagnetic force generated by the magnetic flux density is superimposed on the thermal stress distribution, and a stress balance equation is established in combination with the vibration signal in the motor characteristic parameter; the stress balance equation is solved to obtain the stress distribution of each grid unit, wherein the stress distribution includes radial stress and tangential stress;

[0017] A second-order polynomial global trend function and a Gaussian kernel local deviation function are constructed, the motor characteristic parameters are taken as input variables, the magnetic flux density, the instantaneous temperature distribution and the stress distribution are taken as output variables, and a permanent magnet motor digital twin model with adaptive feature fusion capability is established through dynamic mapping and matching.

[0018] A two-layer deep reinforcement learning neural network is constructed based on the permanent magnet motor digital twin model, and the two-layer deep reinforcement learning neural network includes a high-level policy network and a low-level execution network; the high-level policy network adopts a long short-term memory network structure to generate a global optimization target of the permanent magnet motor; the low-level execution network adopts a deep deterministic policy gradient algorithm to decompose the global optimization target into multiple sub-targets including:

[0019] A two-layer deep reinforcement learning neural network is constructed based on the permanent magnet motor digital twin model, wherein the two-layer deep reinforcement learning neural network includes a high-level strategy network and a low-level execution network; the magnetic flux density, the instantaneous temperature distribution, and the stress distribution of the permanent magnet motor digital twin model are used as state input signals; the high-level strategy network and the low-level execution network are pre-trained based on the state input signals to obtain network initial parameters;

[0020] The high-level strategy network is constructed by using a long short-term memory network, and the high-level strategy network includes an input gate unit, a forget gate unit and an output gate unit; the input gate unit calculates the input importance based on the current state input signal and the hidden layer state at the previous moment, the forget gate unit calculates the historical information retention ratio based on the memory unit state at the previous moment, and the output gate unit calculates the output selection ratio based on the updated memory unit state; the calculated input importance, the historical information retention ratio and the output selection ratio are input into the value function to generate the global optimization target of the permanent magnet motor;

[0021] The low-level execution network is constructed using a deep deterministic policy gradient algorithm, and the global optimization objective is decomposed into electromagnetic optimization sub-objectives, thermal management sub-objectives, and vibration control sub-objectives.

[0022] A reward function is constructed for each sub-goal, and the reward function includes an energy efficiency index, a dynamic response index, and a stability index; the dual-layer deep reinforcement learning neural network is trained using a priority experience replay mechanism to obtain the optimal flux control strategy and the optimal torque control strategy, including:

[0023] A multi-objective reward function is constructed for the electromagnetic optimization sub-objective, the thermal management sub-objective and the vibration control sub-objective; the multi-objective reward function includes an energy efficiency index reward item, a dynamic response index reward item and a stability index reward item; the energy efficiency index reward item is composed of a weighted sum of a motor efficiency parameter and a power factor parameter; the dynamic response index reward item is composed of a weighted sum of a ratio of a stator temperature rise parameter to a maximum allowable stator temperature rise, and a ratio of a rotor temperature rise parameter to a maximum allowable rotor temperature rise; the stability index reward item is composed of a weighted sum of a ratio of a torque pulsation coefficient to a reference torque pulsation coefficient, and a ratio of a radial force fluctuation coefficient to a reference radial force fluctuation coefficient;

[0024] Constructing an experience conversion tuple based on the reward value corresponding to the multi-objective reward function, wherein the experience conversion tuple includes a current state parameter, the reward value, and a next state parameter; storing the experience conversion tuple in an experience conversion tuple repository; calculating a time series difference error based on the current state parameter, the reward value, and the next state parameter; determining a priority parameter of the experience conversion tuple according to a sum of an absolute value of the time series difference error and a preset bias parameter;

[0025] Calculating the sampling probability of the experience conversion tuple according to the priority parameter, the sampling probability is proportional to the power term of the priority parameter; extracting training samples from the experience conversion tuple repository based on the sampling probability; calculating the importance sampling weight according to the sampling probability of the training sample in the experience conversion tuple repository, the importance sampling weight is inversely proportional to the power term of the sampling probability;

[0026] The training samples and the importance sampling weights are input into a two-layer deep reinforcement learning neural network; a high-level policy network of the two-layer deep reinforcement learning neural network generates a global optimization target based on the training samples; a state space model of magnetic flux and torque is constructed based on the global optimization target, the control parameters of the magnetic flux and the torque are optimized by a reinforcement learning algorithm to obtain optimized control parameters, and the global optimization target is decomposed into an optimal magnetic flux control strategy and an optimal torque control strategy based on the optimized control parameters.

[0027] Based on the optimal flux control strategy and the optimal torque control strategy, the adaptive particle swarm algorithm is used to perform multi-objective optimization of control parameters, including:

[0028] Receiving control parameters of the optimal flux control strategy and the optimal torque control strategy; constructing a four-dimensional target optimization function based on the control parameters, wherein the four-dimensional target optimization function includes a flux fluctuation coefficient, a torque pulsation coefficient, a loss coefficient, and a temperature rise rate coefficient;

[0029] The control parameters are encoded into particle position vectors, and a multi-objective optimization space is constructed based on the four-dimensional objective optimization function; initial sampling is performed on the multi-objective optimization space to obtain the position distribution of the initial particle swarm; the fitness value of the initial particle swarm is calculated, and an initial non-dominated solution set is determined based on the fitness value;

[0030] Calculating a population diversity index of the initial particle swarm, where the population diversity index is determined by the average value of the Euclidean distances from all particles to the mass center of the swarm; calculating a fitness change rate index of the initial particle swarm, where the fitness change rate index is determined by the ratio of the change in the optimal fitness value between two adjacent iterations to the current fitness value;

[0031] Inputting the population diversity index and the fitness change rate index into a fuzzy adaptive system; the fuzzy adaptive system dynamically adjusts the inertia weight coefficient and the learning factor coefficient of the particle swarm algorithm based on a preset fuzzy rule; substituting the adjusted inertia weight coefficient and the learning factor coefficient into a particle velocity update equation;

[0032] Calculate a new particle position based on the updated particle velocity; decode the new particle position into a control parameter; input the decoded control parameter into the optimal flux control strategy and the optimal torque control strategy; calculate a new four-dimensional objective optimization function value based on the output response of the optimal flux control strategy and the optimal torque control strategy;

[0033] Perform non-dominated sorting on all particles to determine the non-dominated level of each particle; calculate the crowding distance of particles within the same non-dominated level; the crowding distance is used to measure the distribution density of solutions in the space of the four-dimensional objective optimization function; update the elite archive set based on the non-dominated level and the crowding distance;

[0034] Determine whether the elite archive set exceeds a preset capacity; when the elite archive set exceeds the preset capacity, delete the solution of the most crowded area based on the crowding distance; retain the solution whose non-dominated level is higher than a preset level threshold and whose crowding distance is greater than a preset distance;

[0035] Repeat the fuzzy adaptive adjustment, particle swarm update and elite archive set maintenance process until a preset number of iterations or convergence conditions are reached; extract the Pareto optimal solution set from the elite archive set; apply the optimal control parameters in the Pareto optimal solution set to the optimal flux control strategy and the optimal torque control strategy to achieve multi-objective optimization of permanent magnet motor control parameters.

[0036] The optimized control parameters are input into a permanent magnet motor vector controller with a model prediction function; the permanent magnet motor vector controller is based on a sliding mode variable structure control principle to construct a three-loop control structure including a torque loop, a flux loop and a current loop; the deviation between the flux observation value and the given value of the permanent magnet motor is calculated in real time, and the d-axis current is compensated in real time using an adaptive fuzzy neural network, including:

[0037] The optimized control parameters are input into a permanent magnet motor vector controller with a model prediction function; the permanent magnet motor vector controller is based on a sliding mode variable structure control principle to construct a three-closed-loop control structure including a torque loop, a flux loop and a current loop; the three-closed-loop control structure achieves stable operation and dynamic response performance of the system through coordinated control; during the control process, the deviation between the flux observation value and the flux given value of the permanent magnet motor is calculated in real time, and an adaptive fuzzy neural network is used to compensate the d-axis current in real time to improve the control accuracy of the system;

[0038] The stator current and rotor position signals of the permanent magnet motor are collected in real time; the flux observation value is calculated according to the stator current and the rotor position signal; the flux observation value is compared with the flux given value to obtain the flux deviation; the flux deviation change rate is calculated based on the flux deviation; an adaptive fuzzy neural network with online learning capability is constructed, the input of the adaptive fuzzy neural network is the flux deviation and the flux deviation change rate; the adaptive fuzzy neural network can adaptively adjust network parameters according to the system operation status;

[0039] In the adaptive fuzzy neural network, a Gaussian function is used to perform fuzzy processing on the input variables to realize the membership calculation of the input variables; fuzzy rule matching is performed based on the Mandene inference mechanism to establish a fuzzy control rule base of the system; a d-axis current compensation amount is obtained through real-time matching of the fuzzy control rule base; the d-axis current compensation amount is superimposed on the original d-axis current given value to obtain the compensated d-axis current given value.

[0040] According to a second aspect of the embodiments of the present invention,

[0041] Provides AI-based permanent magnet motor energy efficiency optimization and energy-saving control system, including:

[0042] The first unit is used to collect real-time operation data of the permanent magnet motor during operation by using a multi-dimensional sensor array, perform feature extraction and dimensionality reduction processing on the real-time operation data based on the mutual information entropy criterion, and obtain motor characteristic parameters of the permanent magnet motor; construct a multi-physics field coupling model including electromagnetic field distribution, thermal field distribution, and stress field distribution; dynamically map and match the motor characteristic parameters with the multi-physics field coupling model, and establish a permanent magnet motor digital twin model with adaptive feature fusion capability;

[0043] The second unit is used to build a two-layer deep reinforcement learning neural network based on the permanent magnet motor digital twin model, and the two-layer deep reinforcement learning neural network includes a high-level strategy network and a low-level execution network; the high-level strategy network adopts a long short-term memory network structure to generate a global optimization target of the permanent magnet motor; the low-level execution network adopts a deep deterministic policy gradient algorithm to decompose the global optimization target into multiple sub-targets; a reward function is constructed for each sub-target, and the reward function includes an energy efficiency index, a dynamic response index, and a stability index; a priority experience replay mechanism is used to train the two-layer deep reinforcement learning neural network to obtain an optimal flux control strategy and an optimal torque control strategy; based on the optimal flux control strategy and the optimal torque control strategy, an adaptive particle swarm algorithm is used to perform multi-objective optimization on control parameters;

[0044] The third unit is used to input the optimized control parameters into a permanent magnet motor vector controller with a model prediction function; the permanent magnet motor vector controller is based on the sliding mode variable structure control principle, constructs a three-closed-loop control structure including a torque loop, a flux loop and a current loop; calculates the deviation between the flux observation value and the given value of the permanent magnet motor in real time, and uses an adaptive fuzzy neural network to compensate the d-axis current in real time.

[0045] According to a third aspect of the embodiments of the present invention,

[0046] An electronic device is provided, comprising:

[0047] processor;

[0048] a memory for storing processor-executable instructions;

[0049] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0050] A fourth aspect of the embodiments of the present invention is:

[0051] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0052] The beneficial effects of this application are as follows:

[0053] By using a multi-dimensional sensor array to collect permanent magnet motor operating data, combining the mutual information entropy criterion for feature extraction and dimensionality reduction, building a multi-physics field coupling model, and establishing a digital twin model with adaptive feature fusion capabilities, this method can fully and accurately reflect the operating status of the permanent magnet motor, provide a reliable data foundation and model support for subsequent optimization control, and thus improve the accuracy and adaptability of the control strategy.

[0054] The double-layer deep reinforcement learning neural network, combined with the long short-term memory network and the deep deterministic policy gradient algorithm, realizes the generation and decomposition of the global optimization target. By constructing a reward function that includes energy efficiency, dynamic response and stability indicators, and using the priority experience replay mechanism for training, the optimal flux linkage and torque control strategy is obtained. This method can effectively balance multiple performance indicators of permanent magnet motors, achieve global optimal control, and significantly improve the overall performance of the motor.

[0055] The optimized control parameters are applied to a vector controller with model prediction function, a three-loop control structure and sliding mode variable structure control principle are adopted, and an adaptive fuzzy neural network is introduced to compensate the current in real time. This control method can effectively improve the control accuracy and dynamic response capability of the permanent magnet motor, while enhancing the robustness and anti-interference ability of the system, and ultimately achieve efficient and energy-saving operation of the permanent magnet motor. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A schematic diagram of a flow chart of a permanent magnet motor energy efficiency optimization and energy-saving control method based on artificial intelligence according to an embodiment of the present invention;

[0057] Figure 2 It is a structural schematic diagram of an artificial intelligence-based permanent magnet motor energy efficiency optimization and energy-saving control system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0060] Figure 1 FIG. 1 is a flow chart of a permanent magnet motor energy efficiency optimization and energy-saving control method based on artificial intelligence according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0061] S11. Use a multidimensional sensor array to collect real-time operation data during the operation of the permanent magnet motor, perform feature extraction and dimensionality reduction processing on the real-time operation data based on the mutual information entropy criterion, and obtain the motor characteristic parameters of the permanent magnet motor; construct a multi-physics field coupling model including electromagnetic field distribution, thermal field distribution, and stress field distribution; dynamically map and match the motor characteristic parameters with the multi-physics field coupling model, and establish a permanent magnet motor digital twin model with adaptive feature fusion capability;

[0062] S12. A double-layer deep reinforcement learning neural network is constructed based on the permanent magnet motor digital twin model, and the double-layer deep reinforcement learning neural network includes a high-level strategy network and a low-level execution network; the high-level strategy network adopts a long short-term memory network structure to generate the global optimization target of the permanent magnet motor; the low-level execution network adopts a deep deterministic policy gradient algorithm to decompose the global optimization target into multiple sub-targets; a reward function is constructed for each sub-target, and the reward function includes an energy efficiency index, a dynamic response index and a stability index; a priority experience replay mechanism is used to train the double-layer deep reinforcement learning neural network to obtain an optimal flux control strategy and an optimal torque control strategy; based on the optimal flux control strategy and the optimal torque control strategy, an adaptive particle swarm algorithm is used to perform multi-objective optimization on control parameters;

[0063] S13. Input the optimized control parameters into a permanent magnet motor vector controller with model prediction function; the permanent magnet motor vector controller is based on the sliding mode variable structure control principle, constructs a three-closed-loop control structure including a torque loop, a flux loop and a current loop; calculates the deviation between the observed flux value of the permanent magnet motor and the given value in real time, and uses an adaptive fuzzy neural network to compensate the d-axis current in real time.

[0064] In an optional implementation, constructing a multi-physics field coupling model including electromagnetic field distribution, thermal field distribution, and stress field distribution; dynamically mapping and matching the motor characteristic parameters with the multi-physics field coupling model, and establishing a permanent magnet motor digital twin model with adaptive feature fusion capability includes:

[0065] The stator current and the rotor position in the characteristic parameters of the motor are input into the Maxwell equations, the calculation domain is divided into a plurality of grid units by using the finite element method, and the magnetic flux density of the grid units is solved by Newton-Raphson iteration; the eddy current loss and the hysteresis loss of each grid unit are calculated based on the magnetic flux density to obtain the electromagnetic loss distribution;

[0066] The electromagnetic loss distribution is combined with the temperature signal in the motor characteristic parameter to establish a heat conduction equation; the heat conduction equation is solved to obtain the instantaneous temperature distribution of each grid unit; based on the instantaneous temperature distribution, the thermal stress distribution of each grid unit is calculated by the thermal expansion coefficient;

[0067] The electromagnetic force generated by the magnetic flux density is superimposed on the thermal stress distribution, and a stress balance equation is established in combination with the vibration signal in the motor characteristic parameter; the stress balance equation is solved to obtain the stress distribution of each grid unit, wherein the stress distribution includes radial stress and tangential stress;

[0068] A second-order polynomial global trend function and a Gaussian kernel local deviation function are constructed, the motor characteristic parameters are taken as input variables, the magnetic flux density, the instantaneous temperature distribution and the stress distribution are taken as output variables, and a permanent magnet motor digital twin model with adaptive feature fusion capability is established through dynamic mapping and matching.

[0069] This embodiment provides a method for constructing a digital twin model of a permanent magnet motor, which includes two main steps: constructing a multi-physical field coupling model and establishing a dynamic mapping matching model.

[0070] First, a multi-physics coupling model including electromagnetic field distribution, thermal field distribution and stress field distribution is constructed. This process involves three sub-steps: electromagnetic field analysis, thermal field analysis and stress field analysis.

[0071] In the electromagnetic field analysis, the stator current and rotor position in the motor characteristic parameters are input into the Maxwell equations. The finite element method is used to divide the calculation domain into multiple grid cells, and the size of each grid cell is usually between 0.1mm and 1mm, depending on the size of the motor and the required calculation accuracy. The magnetic flux density of each grid cell is solved using the Newton-Raphson iteration method. During the iteration process, the convergence condition is set to a relative error of less than 0.1% between two adjacent iteration results. Based on the calculated magnetic flux density, the eddy current loss and hysteresis loss of each grid cell are calculated. The eddy current loss is proportional to the frequency and the square of the magnetic flux density, while the hysteresis loss is proportional to the frequency and the 1.6th power of the magnetic flux density. By accumulating the losses of each grid cell, the electromagnetic loss distribution of the entire motor is obtained.

[0072] In the thermal field analysis, the electromagnetic loss distribution obtained in the previous step is combined with the temperature signal in the motor characteristic parameters to establish the heat conduction equation. The temperature signal can be obtained through temperature sensors installed at key locations of the motor, such as stator windings, rotor surface and bearings. The heat conduction equation takes into account the thermal physical parameters such as thermal conductivity, specific heat capacity and density of each component of the motor. The heat conduction equation is solved by the finite difference method or the finite element method to obtain the instantaneous temperature distribution of each grid unit. The calculation time step is usually set to 0.1 seconds to 1 second to capture the dynamic characteristics of temperature changes. Based on the instantaneous temperature distribution obtained, the thermal stress distribution of each grid unit is calculated by the thermal expansion coefficient of each material. The thermal expansion coefficient changes with temperature, so its nonlinear characteristics need to be considered.

[0073] In the stress field analysis, the electromagnetic force generated by the magnetic flux density obtained from the electromagnetic field analysis is superimposed on the thermal stress distribution obtained from the thermal field analysis. The electromagnetic force includes radial force and tangential force, which can be calculated by the Maxwell stress tensor. At the same time, the stress balance equation is established in combination with the vibration signal in the characteristic parameters of the motor. The vibration signal can be measured by an acceleration sensor, which is usually installed on the motor housing or bearing seat. The stress balance equation takes into account the mechanical parameters such as the elastic modulus and Poisson's ratio of each component of the motor. The stress balance equation is solved by the finite element method to obtain the stress distribution of each grid unit, including radial stress and tangential stress.

[0074] After completing the construction of the multi-physics field coupling model, the next step is to establish a dynamic mapping and matching model to realize a permanent magnet motor digital twin model with adaptive feature fusion capabilities.

[0075] First, a second-order polynomial global trend function is constructed. This function is used to describe the overall relationship between the motor characteristic parameters and the multi-physics field distribution. The second-order polynomial can capture the nonlinear relationship between the parameters and improve the accuracy of the model. The coefficients of the global trend function are obtained by least squares fitting.

[0076] Next, a Gaussian kernel local deviation function is constructed. This function is used to describe local changes that cannot be accurately expressed by the global trend function. The hyperparameters of the Gaussian kernel function, such as length scale and signal variance, are determined by maximum likelihood estimation or cross-validation methods.

[0077] The motor characteristic parameters are used as input variables, including stator current, rotor position, temperature signal, vibration signal, etc. The magnetic flux density, instantaneous temperature distribution and stress distribution obtained by the multi-physics field coupling model are used as output variables. The relationship between the input variables and the output variables is established through dynamic mapping and matching.

[0078] The dynamic mapping matching process uses online learning algorithms such as recursive least squares or sliding window least squares. These algorithms can update model parameters in real time to adapt to changes in the motor's operating state. The update frequency is set from 1 to 10 times per second to balance computational efficiency and model accuracy.

[0079] In order to improve the adaptability of the model, a feature fusion mechanism is introduced. The characteristic parameters of different types of motors are weighted and combined, and the weight coefficients are dynamically adjusted through an adaptive algorithm. For example, an adaptive Kalman filter or a particle filter can be used to optimize the weight coefficients.

[0080] Finally, the model performance was evaluated by cross-validation and error analysis. The root mean square error (RMSE) and coefficient of determination (R 2 ) and other indicators to quantify the prediction accuracy of the model. If the prediction error exceeds the preset threshold (such as RMSE greater than 5% or R2 Less than 0.95), the model retraining mechanism is triggered to ensure that the digital twin model always maintains high accuracy.

[0081] Through the above steps, a permanent magnet motor digital twin model with adaptive feature fusion capability was established. This model can reflect the multi-physical field distribution of the motor in real time, providing strong support for motor performance optimization and fault diagnosis.

[0082] The solution of this application can:

[0083] By constructing a multi-physics coupling model and a dynamic mapping matching model, high-precision modeling of the digital twin model of the permanent magnet motor is achieved. The multi-physics coupling model takes into account the interaction between the electromagnetic field, thermal field and stress field, and fully reflects the working state of the motor. The dynamic mapping matching model adopts an online learning algorithm and a feature fusion mechanism to make the digital twin model adaptive and able to track the changes in motor performance in real time. The finite element method and iterative algorithm are used to solve complex physical field equations, which improves the computational efficiency and numerical stability. By reasonably setting the grid size, time step and iterative convergence conditions, a good balance is achieved between computational accuracy and efficiency. At the same time, the second-order polynomial global trend function and the Gaussian kernel local deviation function are introduced to improve the nonlinear fitting ability and local adaptability of the model. The established digital twin model has broad application prospects. It can be used for motor performance optimization, and by analyzing the distribution of multi-physics fields, the key factors affecting the efficiency and reliability of the motor can be found. In terms of fault diagnosis, the model can monitor the stress distribution and temperature distribution of the motor in real time, and detect potential fault hazards as early as possible. In addition, the model can also be used for motor life prediction and predictive maintenance, reducing maintenance costs and increasing the service life of the motor.

[0084] In an optional embodiment, a two-layer deep reinforcement learning neural network is constructed based on the permanent magnet motor digital twin model, and the two-layer deep reinforcement learning neural network includes a high-level policy network and a low-level execution network; the high-level policy network adopts a long short-term memory network structure to generate a global optimization target of the permanent magnet motor; the low-level execution network adopts a deep deterministic policy gradient algorithm to decompose the global optimization target into multiple sub-targets including:

[0085] A two-layer deep reinforcement learning neural network is constructed based on the permanent magnet motor digital twin model, wherein the two-layer deep reinforcement learning neural network includes a high-level strategy network and a low-level execution network; the magnetic flux density, the instantaneous temperature distribution, and the stress distribution of the permanent magnet motor digital twin model are used as state input signals; the high-level strategy network and the low-level execution network are pre-trained based on the state input signals to obtain network initial parameters;

[0086] The high-level strategy network is constructed by using a long short-term memory network, and the high-level strategy network includes an input gate unit, a forget gate unit and an output gate unit; the input gate unit calculates the input importance based on the current state input signal and the hidden layer state at the previous moment, the forget gate unit calculates the historical information retention ratio based on the memory unit state at the previous moment, and the output gate unit calculates the output selection ratio based on the updated memory unit state; the calculated input importance, the historical information retention ratio and the output selection ratio are input into the value function to generate the global optimization target of the permanent magnet motor;

[0087] The low-level execution network is constructed using a deep deterministic policy gradient algorithm, and the global optimization objective is decomposed into electromagnetic optimization sub-objectives, thermal management sub-objectives, and vibration control sub-objectives.

[0088] First, a digital twin model of the permanent magnet motor is constructed. The model includes an electromagnetic field simulation module, a thermal field simulation module, and a stress field simulation module. The electromagnetic field simulation module uses the finite element method to perform real-time calculations on the magnetic flux density distribution of the permanent magnet motor; the thermal field simulation module dynamically estimates the temperature of each motor component based on the thermal network method; and the stress field simulation module analyzes the vibration characteristics of the motor through elastic mechanics modeling.

[0089] Then, a two-layer deep reinforcement learning neural network architecture is constructed. The high-level policy network adopts a long short-term memory network structure, which includes an input layer, a hidden layer, and an output layer. The input layer receives the magnetic flux density, temperature distribution, and stress distribution data output by the digital twin model; the hidden layer consists of multiple memory units, each of which contains an input gate, a forget gate, and an output gate; the output layer generates a global optimization target. The low-level execution network uses a deep deterministic policy gradient algorithm to decompose the global target into specific control instructions.

[0090] In the network training phase, the digital twin model is first simulated in large quantities and state data is collected for pre-training. The pre-training adopts a supervised learning method and uses expert experience data as labels. After the pre-training is completed, the network parameters are continuously optimized through online reinforcement learning.

[0091] The training process of the high-level policy network is as follows: the input gate unit receives the current state signal and the hidden state of the previous moment, and calculates the importance of the input data; the forget gate unit determines the proportion of historical information that needs to be retained based on the historical memory state; the output gate unit calculates the final output selection ratio based on the updated memory state. The outputs of these three gate units act together on the value function to generate the global optimization target of the permanent magnet motor.

[0092] After receiving the global optimization goal, the low-level execution network decomposes it into three sub-goals: the electromagnetic optimization sub-goal focuses on the torque ripple and efficiency of the permanent magnet motor; the thermal management sub-goal optimizes the temperature distribution of each component; and the vibration control sub-goal focuses on reducing motor vibration and noise. The execution network uses a deep deterministic policy gradient algorithm to convert these sub-goals into specific control instructions.

[0093] In actual application cases, under rated working conditions, the peak flux density of a certain type of permanent magnet motor is 1.8 Tesla, the maximum temperature of the stator winding is 120 degrees Celsius, and the peak vibration acceleration is 2.5 meters per square second. After optimization of the control system, the torque pulsation is reduced by 25%, the maximum temperature is reduced to 105 degrees Celsius, and the vibration acceleration is reduced by 30%.

[0094] The solution of this application can:

[0095] This technical solution achieves all-round optimization of permanent magnet motor performance by building a control system that combines a digital twin model with deep reinforcement learning. The digital twin model can accurately reflect the real-time operating status of the motor and provide a reliable basis for control decisions. The design of the two-layer neural network architecture embodies the idea of ​​hierarchical control. The high-level network is responsible for global strategy formulation, and the low-level network performs specific control tasks, which not only ensures the integrity of the control, but also ensures the accuracy of execution. The long and short-term memory network structure can effectively utilize historical operating data and improve the accuracy of control decisions. This solution unifies electromagnetic property optimization, thermal management, and vibration control into one framework, achieving multi-objective collaborative optimization. Through the deep deterministic policy gradient algorithm, the continuity and stability of the control instructions are guaranteed, and the stability and reliability of the system operation are improved.

[0096] In an optional implementation, a reward function is constructed for each sub-goal, the reward function including an energy efficiency index, a dynamic response index and a stability index; the dual-layer deep reinforcement learning neural network is trained using a priority experience replay mechanism to obtain an optimal flux control strategy and an optimal torque control strategy including:

[0097] A multi-objective reward function is constructed for the electromagnetic optimization sub-objective, the thermal management sub-objective and the vibration control sub-objective; the multi-objective reward function includes an energy efficiency index reward item, a dynamic response index reward item and a stability index reward item; the energy efficiency index reward item is composed of a weighted sum of a motor efficiency parameter and a power factor parameter; the dynamic response index reward item is composed of a weighted sum of a ratio of a stator temperature rise parameter to a maximum allowable stator temperature rise, and a ratio of a rotor temperature rise parameter to a maximum allowable rotor temperature rise; the stability index reward item is composed of a weighted sum of a ratio of a torque pulsation coefficient to a reference torque pulsation coefficient, and a ratio of a radial force fluctuation coefficient to a reference radial force fluctuation coefficient;

[0098] Constructing an experience conversion tuple based on the reward value corresponding to the multi-objective reward function, wherein the experience conversion tuple includes a current state parameter, the reward value, and a next state parameter; storing the experience conversion tuple in an experience conversion tuple repository; calculating a time series difference error based on the current state parameter, the reward value, and the next state parameter; determining a priority parameter of the experience conversion tuple according to a sum of an absolute value of the time series difference error and a preset bias parameter;

[0099] Calculating the sampling probability of the experience conversion tuple according to the priority parameter, the sampling probability is proportional to the power term of the priority parameter; extracting training samples from the experience conversion tuple repository based on the sampling probability; calculating the importance sampling weight according to the sampling probability of the training sample in the experience conversion tuple repository, the importance sampling weight is inversely proportional to the power term of the sampling probability;

[0100] The training samples and the importance sampling weights are input into a two-layer deep reinforcement learning neural network; a high-level policy network of the two-layer deep reinforcement learning neural network generates a global optimization target based on the training samples; a state space model of magnetic flux and torque is constructed based on the global optimization target, the control parameters of the magnetic flux and the torque are optimized by a reinforcement learning algorithm to obtain optimized control parameters, and the global optimization target is decomposed into an optimal magnetic flux control strategy and an optimal torque control strategy based on the optimized control parameters.

[0101] Aiming at the multi-objective optimization problem of permanent magnet synchronous motor control system, a multi-objective reward function including energy efficiency index, dynamic response index and stability index is first constructed. The energy efficiency index reward item in the reward function is weighted by the motor efficiency parameter and the power factor parameter. The motor efficiency parameter ranges from 0.8 to 0.95, and the power factor parameter ranges from 0.85 to 0.98, with weighting coefficients of 0.6 and 0.4 respectively. The dynamic response index reward item is composed of the stator temperature rise ratio and the rotor temperature rise ratio. The maximum allowable temperature rise of the stator is 120 degrees Celsius, and the maximum allowable temperature rise of the rotor is 140 degrees Celsius. The weighting coefficients are both 0.5. The stability index reward item includes the torque pulsation ratio and the radial force fluctuation ratio. The reference torque pulsation coefficient is set to 5%, and the reference radial force fluctuation coefficient is set to 10%. The weighting coefficients are 0.55 and 0.45 respectively.

[0102] When constructing the experience conversion tuple, the current state parameters include operating parameters such as speed, torque, stator current and rotor position, and the next state parameters are obtained through system simulation. The reward value is obtained based on the conversion from the current state to the next state, and the experience conversion tuple is stored in a storage repository with a capacity of 10,000. The temporal difference error calculation uses a preset discount factor of 0.9, and the preset bias parameter is set to 0.01. The priority parameter is obtained by summing the absolute value of the temporal difference error and the bias parameter, which is used for subsequent importance sampling.

[0103] The sampling probability is proportional to the 0.6 power of the priority parameter, and a training batch of size 128 is randomly extracted from the repository for each training. The importance sampling weight is proportional to the negative 0.4 power of the sampling probability, which is used to balance the over-learning of high-priority samples. In the two-layer deep reinforcement learning network, the high-level policy network adopts a four-layer fully connected structure, with the number of hidden layer neurons being 256, 128, 64, and 32, respectively, and using the ReLU activation function.

[0104] After the high-level policy network generates the global optimization target based on the training samples, a state space model including flux and torque is constructed. The control parameters are iteratively optimized through the deep deterministic policy gradient algorithm, with the learning rate set to 0.001 and updated 200 times per iteration. The optimal flux control strategy and optimal torque control strategy finally obtained correspond to the optimal control sequences of flux amplitude and torque angle, respectively.

[0105] The solution of this application can:

[0106] By constructing a multi-objective reward function, the unified optimization of multiple sub-objectives such as electromagnetic optimization, thermal management and vibration control is achieved, and the comprehensive performance of the permanent magnet synchronous motor system is improved. The priority experience replay mechanism is adopted to improve the utilization efficiency of training samples with important value, accelerate the convergence speed of the deep reinforcement learning network, and enhance the stability of the training process. The hierarchical control structure based on the double-layer deep reinforcement learning network realizes the effective decomposition from the global optimization goal to the specific control strategy, and improves the dynamic performance and robustness of the permanent magnet synchronous motor control system.

[0107] In an optional implementation, based on the optimal flux control strategy and the optimal torque control strategy, using an adaptive particle swarm algorithm to perform multi-objective optimization on control parameters includes:

[0108] Receiving control parameters of the optimal flux control strategy and the optimal torque control strategy; constructing a four-dimensional target optimization function based on the control parameters, wherein the four-dimensional target optimization function includes a flux fluctuation coefficient, a torque pulsation coefficient, a loss coefficient, and a temperature rise rate coefficient;

[0109] The control parameters are encoded into particle position vectors, and a multi-objective optimization space is constructed based on the four-dimensional objective optimization function; initial sampling is performed on the multi-objective optimization space to obtain the position distribution of the initial particle swarm; the fitness value of the initial particle swarm is calculated, and an initial non-dominated solution set is determined based on the fitness value;

[0110] Calculating a population diversity index of the initial particle swarm, where the population diversity index is determined by the average value of the Euclidean distances from all particles to the mass center of the swarm; calculating a fitness change rate index of the initial particle swarm, where the fitness change rate index is determined by the ratio of the change in the optimal fitness value between two adjacent iterations to the current fitness value;

[0111] Inputting the population diversity index and the fitness change rate index into a fuzzy adaptive system; the fuzzy adaptive system dynamically adjusts the inertia weight coefficient and the learning factor coefficient of the particle swarm algorithm based on a preset fuzzy rule; substituting the adjusted inertia weight coefficient and the learning factor coefficient into a particle velocity update equation;

[0112] Calculate a new particle position based on the updated particle velocity; decode the new particle position into a control parameter; input the decoded control parameter into the optimal flux control strategy and the optimal torque control strategy; calculate a new four-dimensional objective optimization function value based on the output response of the optimal flux control strategy and the optimal torque control strategy;

[0113] Perform non-dominated sorting on all particles to determine the non-dominated level of each particle; calculate the crowding distance of particles within the same non-dominated level; the crowding distance is used to measure the distribution density of solutions in the space of the four-dimensional objective optimization function; update the elite archive set based on the non-dominated level and the crowding distance;

[0114] Determine whether the elite archive set exceeds a preset capacity; when the elite archive set exceeds the preset capacity, delete the solution of the most crowded area based on the crowding distance; retain the solution whose non-dominated level is higher than a preset level threshold and whose crowding distance is greater than a preset distance;

[0115] Repeat the fuzzy adaptive adjustment, particle swarm update and elite archive set maintenance process until a preset number of iterations or convergence conditions are reached; extract the Pareto optimal solution set from the elite archive set; apply the optimal control parameters in the Pareto optimal solution set to the optimal flux control strategy and the optimal torque control strategy to achieve multi-objective optimization of permanent magnet motor control parameters.

[0116] The multi-objective optimization method for permanent magnet motor control parameters based on the optimal flux control strategy and the optimal torque control strategy first needs to receive the initial control parameters of the control strategy, including the proportional coefficient and integral coefficient of the flux control, and the proportional coefficient and integral coefficient of the torque control. These parameters will serve as the basic input for optimization.

[0117] When constructing the four-dimensional objective optimization function, the flux fluctuation coefficient is calculated by the sum of the squares of the deviations between the actual flux value and the given value within the sampling period, and the typical value range is 0.01-0.05; the torque pulsation coefficient is determined by the root mean square of the deviation between the actual torque value and the given value, and the typical value range is 0.02-0.08; the loss coefficient considers the combined effects of copper loss, iron loss and switching loss, and the typical value range is 0.1-0.3; the temperature rise rate coefficient is calculated by the temperature change rate of the stator winding, and the typical value range is 0.05-0.15.

[0118] In the particle swarm encoding phase, the control parameters are mapped to a four-dimensional search space, and each particle represents a set of candidate control parameter solutions. The initial particle swarm size is set to 50 and randomly distributed in the search space. The Monte Carlo method is used for initial sampling to ensure the uniformity of the initial particle distribution. For each initial particle, four objective function values ​​are calculated, and the initial non-dominated solution set is determined based on the non-dominated relationship.

[0119] When calculating the population diversity index, first determine the position of the group's centroid, then calculate the Euclidean distance from all particles to the centroid, and take the average value as the diversity measure. The fitness change rate index is obtained by comparing the ratio of the change in the optimal fitness value of two adjacent iterations to the current fitness value. These two indicators are used as inputs to the fuzzy adaptive system.

[0120] The fuzzy adaptive system uses a triangular membership function, and the input variables and output variables are divided into three fuzzy subsets: low, medium, and high. A fuzzy rule base is built based on expert experience, which contains 9 rules. The centroid method is used to defuzzify and obtain the exact values ​​of the inertia weight coefficient and the learning factor coefficient. The typical inertia weight coefficient ranges from 0.4 to 0.9, and the learning factor coefficient ranges from 1.5 to 2.5.

[0121] In the particle position update process, the new particle velocity is first calculated based on the velocity update equation, and then the particle position is updated. The new position is checked for boundaries to ensure that it is within the valid search space. The updated position is decoded into control parameters, input into the control strategy for simulation verification, and the system response is obtained and the new objective function value is calculated.

[0122] Non-dominated sorting uses a fast non-dominated sorting algorithm to divide the population into different non-dominated layers. When calculating the crowding distance, each objective function is normalized to ensure the consistency of different objective dimensions. The capacity of the elite archive set is set to 100. When the capacity is exceeded, solutions with a non-dominated level higher than 0.8 and a crowding distance greater than 0.5 are preferentially retained.

[0123] The algorithm iteration termination condition is set to reach the maximum number of iterations of 500 or the improvement of the optimal solution is less than 0.001 for 50 consecutive iterations. The Pareto optimal solution set is extracted from the final elite archive set, and the optimal control parameter scheme is selected through the decision maker's preference information.

[0124] The solution of this application can:

[0125] By introducing the fuzzy adaptive mechanism, the dynamic adjustment of the particle swarm algorithm parameters is realized, the convergence performance and search efficiency of the algorithm are improved, and the problem of falling into the local optimum is avoided. A multi-objective optimization method is adopted, and multiple performance indicators such as flux fluctuation, torque pulsation, loss and temperature rise are considered at the same time, and a set of optimal solutions with trade-offs is obtained, which provides greater flexibility for the selection of controller parameters. Through the elite archive strategy and congestion distance calculation, the diversity and uniform distribution of the optimal solution set are guaranteed, the quality and selectivity of the solution are improved, and the needs of practical engineering applications are met.

[0126] In an optional implementation, the optimized control parameters are input into a permanent magnet motor vector controller with a model prediction function; the permanent magnet motor vector controller is based on a sliding mode variable structure control principle to construct a three-loop control structure including a torque loop, a flux loop and a current loop; the deviation between the flux observation value of the permanent magnet motor and the given value is calculated in real time, and the d-axis current is compensated in real time using an adaptive fuzzy neural network, including:

[0127] The optimized control parameters are input into a permanent magnet motor vector controller with a model prediction function; the permanent magnet motor vector controller is based on a sliding mode variable structure control principle to construct a three-closed-loop control structure including a torque loop, a flux loop and a current loop; the three-closed-loop control structure achieves stable operation and dynamic response performance of the system through coordinated control; during the control process, the deviation between the flux observation value and the flux given value of the permanent magnet motor is calculated in real time, and an adaptive fuzzy neural network is used to compensate the d-axis current in real time to improve the control accuracy of the system;

[0128] The stator current and rotor position signals of the permanent magnet motor are collected in real time; the flux observation value is calculated according to the stator current and the rotor position signal; the flux observation value is compared with the flux given value to obtain the flux deviation; the flux deviation change rate is calculated based on the flux deviation; an adaptive fuzzy neural network with online learning capability is constructed, the input of the adaptive fuzzy neural network is the flux deviation and the flux deviation change rate; the adaptive fuzzy neural network can adaptively adjust network parameters according to the system operation status;

[0129] In the adaptive fuzzy neural network, a Gaussian function is used to perform fuzzy processing on the input variables to realize the membership calculation of the input variables; fuzzy rule matching is performed based on the Mandene inference mechanism to establish a fuzzy control rule base of the system; a d-axis current compensation amount is obtained through real-time matching of the fuzzy control rule base; the d-axis current compensation amount is superimposed on the original d-axis current given value to obtain the compensated d-axis current given value.

[0130] The permanent magnet motor vector control system first needs to establish a complete control architecture. The control system adopts a three-closed-loop control structure based on the sliding mode variable structure control principle, including the outermost torque control link, the middle link flux control link, and the innermost current control link. When the system is running, the controller collects the stator current signal and the rotor position signal in real time, and pre-processes them through the signal conditioning circuit to ensure the accuracy of the collected signals.

[0131] In the flux observation link, the system calculates the real-time flux observation value through the permanent magnet motor mathematical model based on the collected stator current and rotor position signals. In specific implementation, the extended Kalman filter algorithm can be used to estimate the flux in real time. The state variables of the filter include d-axis current, q-axis current, rotor angular velocity and rotor position angle. The sampling period of the filter is set to 100 microseconds, which can effectively suppress the influence of measurement noise.

[0132] In the calculation of flux deviation, the observed flux value is compared with the given value in real time. The system sets the given value of flux to 0.3 Weber. When the observed flux value deviates from the given value in actual operation, the controller calculates the flux deviation value. At the same time, the controller calculates the rate of change of the flux deviation to reflect the dynamic characteristics of the system.

[0133] The construction of the adaptive fuzzy neural network adopts a five-layer network structure. The input layer receives two input quantities, namely, the flux linkage deviation and the deviation change rate. The fuzzification layer uses the Gaussian membership function to convert the input variables into fuzzy quantities. For each input variable, seven language values ​​are set: negative large, negative medium, negative small, zero, positive small, positive medium, and positive large. The fuzzy rule layer builds a complete rule base based on expert experience, which contains forty-nine control rules.

[0134] During the online learning process, the network adjusts the membership function parameters and rule layer weights in real time through the back propagation algorithm. The learning rate is set to 0.01 and the momentum factor is set to 0.9 to ensure the network convergence performance. The network output layer uses the centroid method to perform defuzzification operations to obtain the compensation amount of the d-axis current.

[0135] Finally, the controller superimposes the calculated d-axis current compensation with the original given value. The original d-axis current given value is usually set to zero. The compensated d-axis current given value can effectively compensate for the flux fluctuation during the operation of the permanent magnet motor. The system sampling period is set to 100 microseconds, and the controller uses a 32-bit floating-point DSP as the core processor, and the calculation accuracy meets the control requirements.

[0136] The solution of this application can:

[0137] The three-closed-loop control structure significantly improves the dynamic response performance of the system. Through the coordinated control of the torque loop, flux loop and current loop, the system achieves rapid dynamic response and stable operation, effectively reduces the system overshoot and shortens the adjustment time. The adaptive fuzzy neural network is introduced to compensate the d-axis current in real time, which improves the robustness and anti-interference ability of the system. The network has an online learning function and can adaptively adjust the control parameters according to the system operation status. It has strong adaptability and high control accuracy. The system adopts a high-precision flux observation algorithm and real-time compensation strategy to effectively suppress the flux fluctuation during the operation of the permanent magnet motor, improve the stability and reliability of the system operation, reduce the loss of the motor, and extend the service life of the equipment.

[0138] Figure 2 FIG. 1 is a schematic diagram of a permanent magnet motor energy efficiency optimization and energy-saving control system based on artificial intelligence according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0139] The first unit is used to collect real-time operation data of the permanent magnet motor during operation by using a multi-dimensional sensor array, perform feature extraction and dimensionality reduction processing on the real-time operation data based on the mutual information entropy criterion, and obtain motor characteristic parameters of the permanent magnet motor; construct a multi-physics field coupling model including electromagnetic field distribution, thermal field distribution, and stress field distribution; dynamically map and match the motor characteristic parameters with the multi-physics field coupling model, and establish a permanent magnet motor digital twin model with adaptive feature fusion capability;

[0140] The second unit is used to build a two-layer deep reinforcement learning neural network based on the permanent magnet motor digital twin model, and the two-layer deep reinforcement learning neural network includes a high-level strategy network and a low-level execution network; the high-level strategy network adopts a long short-term memory network structure to generate a global optimization target of the permanent magnet motor; the low-level execution network adopts a deep deterministic policy gradient algorithm to decompose the global optimization target into multiple sub-targets; a reward function is constructed for each sub-target, and the reward function includes an energy efficiency index, a dynamic response index, and a stability index; a priority experience replay mechanism is used to train the two-layer deep reinforcement learning neural network to obtain an optimal flux control strategy and an optimal torque control strategy; based on the optimal flux control strategy and the optimal torque control strategy, an adaptive particle swarm algorithm is used to perform multi-objective optimization on control parameters;

[0141] The third unit is used to input the optimized control parameters into a permanent magnet motor vector controller with a model prediction function; the permanent magnet motor vector controller is based on the sliding mode variable structure control principle, constructs a three-closed-loop control structure including a torque loop, a flux loop and a current loop; calculates the deviation between the flux observation value and the given value of the permanent magnet motor in real time, and uses an adaptive fuzzy neural network to compensate the d-axis current in real time.

[0142] According to a third aspect of the embodiments of the present invention,

[0143] An electronic device is provided, comprising:

[0144] processor;

[0145] a memory for storing processor-executable instructions;

[0146] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0147] A fourth aspect of the embodiments of the present invention is:

[0148] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0149] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention 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 replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A permanent magnet motor energy efficiency optimization and energy-saving control method based on artificial intelligence, characterized in that: include: A multi-dimensional sensor array is used to collect real-time operation data of the permanent magnet motor during operation, and feature extraction and dimensionality reduction processing are performed on the real-time operation data based on the mutual information entropy criterion to obtain the motor characteristic parameters of the permanent magnet motor; a multi-physics field coupling model including electromagnetic field distribution, thermal field distribution, and stress field distribution is constructed; the motor characteristic parameters are dynamically mapped and matched with the multi-physics field coupling model to establish a permanent magnet motor digital twin model with adaptive feature fusion capability; A two-layer deep reinforcement learning neural network is constructed based on the permanent magnet motor digital twin model, wherein the two-layer deep reinforcement learning neural network includes a high-level strategy network and a low-level execution network; the high-level strategy network adopts a long short-term memory network structure to generate a global optimization target for the permanent magnet motor; The low-level execution network uses a deep deterministic policy gradient algorithm to decompose the global optimization goal into multiple sub-goals; a reward function is constructed for each sub-goal, and the reward function includes an energy efficiency index, a dynamic response index, and a stability index; a priority experience replay mechanism is used to train the two-layer deep reinforcement learning neural network to obtain an optimal flux control strategy and an optimal torque control strategy; based on the optimal flux control strategy and the optimal torque control strategy, an adaptive particle swarm algorithm is used to perform multi-objective optimization of control parameters; The optimized control parameters are input into a permanent magnet motor vector controller with a model prediction function; the permanent magnet motor vector controller is based on a sliding mode variable structure control principle to construct a three-closed-loop control structure including a torque loop, a flux loop and a current loop; the deviation between the flux observation value and the given value of the permanent magnet motor is calculated in real time, and an adaptive fuzzy neural network is used to compensate the d-axis current in real time.

2. The method according to claim 1, characterized in that Constructing a multi-physics field coupling model including electromagnetic field distribution, thermal field distribution, and stress field distribution; dynamically mapping and matching the motor characteristic parameters with the multi-physics field coupling model, and establishing a permanent magnet motor digital twin model with adaptive feature fusion capability, including: The stator current and the rotor position in the characteristic parameters of the motor are input into the Maxwell equations, the calculation domain is divided into a plurality of grid units by using the finite element method, and the magnetic flux density of the grid units is solved by Newton-Raphson iteration; the eddy current loss and the hysteresis loss of each grid unit are calculated based on the magnetic flux density to obtain the electromagnetic loss distribution; The electromagnetic loss distribution is combined with the temperature signal in the motor characteristic parameter to establish a heat conduction equation; the heat conduction equation is solved to obtain the instantaneous temperature distribution of each grid unit; based on the instantaneous temperature distribution, the thermal stress distribution of each grid unit is calculated by the thermal expansion coefficient; The electromagnetic force generated by the magnetic flux density is superimposed on the thermal stress distribution, and a stress balance equation is established in combination with the vibration signal in the motor characteristic parameter; the stress balance equation is solved to obtain the stress distribution of each grid unit, wherein the stress distribution includes radial stress and tangential stress; A second-order polynomial global trend function and a Gaussian kernel local deviation function are constructed, the motor characteristic parameters are taken as input variables, the magnetic flux density, the instantaneous temperature distribution and the stress distribution are taken as output variables, and a permanent magnet motor digital twin model with adaptive feature fusion capability is established through dynamic mapping and matching.

3. The method according to claim 1, characterized in that A two-layer deep reinforcement learning neural network is constructed based on the permanent magnet motor digital twin model, wherein the two-layer deep reinforcement learning neural network includes a high-level strategy network and a low-level execution network; the high-level strategy network adopts a long short-term memory network structure to generate a global optimization target for the permanent magnet motor; The low-level execution network uses a deep deterministic policy gradient algorithm to decompose the global optimization goal into multiple sub-goals including: A two-layer deep reinforcement learning neural network is constructed based on the permanent magnet motor digital twin model, wherein the two-layer deep reinforcement learning neural network includes a high-level strategy network and a low-level execution network; the magnetic flux density, the instantaneous temperature distribution, and the stress distribution of the permanent magnet motor digital twin model are used as state input signals; the high-level strategy network and the low-level execution network are pre-trained based on the state input signals to obtain network initial parameters; The high-level strategy network is constructed by using a long short-term memory network, and the high-level strategy network includes an input gate unit, a forget gate unit and an output gate unit; the input gate unit calculates the input importance based on the current state input signal and the hidden layer state at the previous moment, the forget gate unit calculates the historical information retention ratio based on the memory unit state at the previous moment, and the output gate unit calculates the output selection ratio based on the updated memory unit state; the calculated input importance, the historical information retention ratio and the output selection ratio are input into the value function to generate the global optimization target of the permanent magnet motor; The low-level execution network is constructed using a deep deterministic policy gradient algorithm, and the global optimization objective is decomposed into electromagnetic optimization sub-objectives, thermal management sub-objectives, and vibration control sub-objectives.

4. The method according to claim 3, characterized in that A reward function is constructed for each sub-goal, and the reward function includes an energy efficiency index, a dynamic response index, and a stability index; the dual-layer deep reinforcement learning neural network is trained using a priority experience replay mechanism to obtain the optimal flux control strategy and the optimal torque control strategy, including: A multi-objective reward function is constructed for the electromagnetic optimization sub-objective, the thermal management sub-objective and the vibration control sub-objective; the multi-objective reward function includes an energy efficiency index reward item, a dynamic response index reward item and a stability index reward item; the energy efficiency index reward item is composed of a weighted sum of a motor efficiency parameter and a power factor parameter; the dynamic response index reward item is composed of a weighted sum of a ratio of a stator temperature rise parameter to a maximum allowable stator temperature rise, and a ratio of a rotor temperature rise parameter to a maximum allowable rotor temperature rise; the stability index reward item is composed of a weighted sum of a ratio of a torque pulsation coefficient to a reference torque pulsation coefficient, and a ratio of a radial force fluctuation coefficient to a reference radial force fluctuation coefficient; Constructing an experience conversion tuple based on the reward value corresponding to the multi-objective reward function, wherein the experience conversion tuple includes a current state parameter, the reward value, and a next state parameter; storing the experience conversion tuple in an experience conversion tuple repository; calculating a time series difference error based on the current state parameter, the reward value, and the next state parameter; determining a priority parameter of the experience conversion tuple according to a sum of an absolute value of the time series difference error and a preset bias parameter; Calculating the sampling probability of the experience conversion tuple according to the priority parameter, the sampling probability is proportional to the power term of the priority parameter; extracting training samples from the experience conversion tuple repository based on the sampling probability; calculating the importance sampling weight according to the sampling probability of the training sample in the experience conversion tuple repository, the importance sampling weight is inversely proportional to the power term of the sampling probability; The training samples and the importance sampling weights are input into a two-layer deep reinforcement learning neural network; a high-level policy network of the two-layer deep reinforcement learning neural network generates a global optimization target based on the training samples; a state space model of magnetic flux and torque is constructed based on the global optimization target, the control parameters of the magnetic flux and the torque are optimized by a reinforcement learning algorithm to obtain optimized control parameters, and the global optimization target is decomposed into an optimal magnetic flux control strategy and an optimal torque control strategy based on the optimized control parameters.

5. The method according to claim 1, characterized in that Based on the optimal flux control strategy and the optimal torque control strategy, the adaptive particle swarm algorithm is used to perform multi-objective optimization of control parameters, including: Receiving control parameters of the optimal flux control strategy and the optimal torque control strategy; constructing a four-dimensional target optimization function based on the control parameters, wherein the four-dimensional target optimization function includes a flux fluctuation coefficient, a torque pulsation coefficient, a loss coefficient, and a temperature rise rate coefficient; The control parameters are encoded into particle position vectors, and a multi-objective optimization space is constructed based on the four-dimensional objective optimization function; initial sampling is performed on the multi-objective optimization space to obtain the position distribution of the initial particle swarm; the fitness value of the initial particle swarm is calculated, and an initial non-dominated solution set is determined based on the fitness value; Calculating a population diversity index of the initial particle swarm, where the population diversity index is determined by the average value of the Euclidean distances from all particles to the mass center of the swarm; calculating a fitness change rate index of the initial particle swarm, where the fitness change rate index is determined by the ratio of the change in the optimal fitness value between two adjacent iterations to the current fitness value; Inputting the population diversity index and the fitness change rate index into a fuzzy adaptive system; the fuzzy adaptive system dynamically adjusts the inertia weight coefficient and the learning factor coefficient of the particle swarm algorithm based on a preset fuzzy rule; substituting the adjusted inertia weight coefficient and the learning factor coefficient into a particle velocity update equation; Calculate a new particle position based on the updated particle velocity; decode the new particle position into a control parameter; input the decoded control parameter into the optimal flux control strategy and the optimal torque control strategy; calculate a new four-dimensional objective optimization function value based on the output response of the optimal flux control strategy and the optimal torque control strategy; Perform non-dominated sorting on all particles to determine the non-dominated level of each particle; calculate the crowding distance of particles within the same non-dominated level; the crowding distance is used to measure the distribution density of solutions in the space of the four-dimensional objective optimization function; update the elite archive set based on the non-dominated level and the crowding distance; Determine whether the elite archive set exceeds a preset capacity; when the elite archive set exceeds the preset capacity, delete the solution of the most crowded area based on the crowding distance; retain the solution whose non-dominated level is higher than a preset level threshold and whose crowding distance is greater than a preset distance; Repeat the fuzzy adaptive adjustment, particle swarm update and elite archive set maintenance process until a preset number of iterations or convergence conditions are reached; extract the Pareto optimal solution set from the elite archive set; apply the optimal control parameters in the Pareto optimal solution set to the optimal flux control strategy and the optimal torque control strategy to achieve multi-objective optimization of permanent magnet motor control parameters.

6. The method according to claim 1, characterized in that The optimized control parameters are input into a permanent magnet motor vector controller with a model prediction function; the permanent magnet motor vector controller is based on a sliding mode variable structure control principle and constructs a three-loop control structure including a torque loop, a flux loop and a current loop; Real-time calculation of the deviation between the observed value and the given value of the permanent magnet motor flux linkage, and real-time compensation of the d-axis current using an adaptive fuzzy neural network include: The optimized control parameters are input into a permanent magnet motor vector controller with a model prediction function; the permanent magnet motor vector controller is based on a sliding mode variable structure control principle to construct a three-closed-loop control structure including a torque loop, a flux loop and a current loop; the three-closed-loop control structure achieves stable operation and dynamic response performance of the system through coordinated control; during the control process, the deviation between the flux observation value and the flux given value of the permanent magnet motor is calculated in real time, and an adaptive fuzzy neural network is used to compensate the d-axis current in real time to improve the control accuracy of the system; The stator current and rotor position signals of the permanent magnet motor are collected in real time; the flux observation value is calculated according to the stator current and the rotor position signal; the flux observation value is compared with the flux given value to obtain the flux deviation; the flux deviation change rate is calculated based on the flux deviation; an adaptive fuzzy neural network with online learning capability is constructed, the input of the adaptive fuzzy neural network is the flux deviation and the flux deviation change rate; the adaptive fuzzy neural network can adaptively adjust network parameters according to the system operation status; In the adaptive fuzzy neural network, a Gaussian function is used to perform fuzzy processing on the input variables to realize the membership calculation of the input variables; fuzzy rule matching is performed based on the Mandene inference mechanism to establish a fuzzy control rule base of the system; a d-axis current compensation amount is obtained through real-time matching of the fuzzy control rule base; the d-axis current compensation amount is superimposed on the original d-axis current given value to obtain the compensated d-axis current given value.

7. A permanent magnet motor energy efficiency optimization and energy-saving control system based on artificial intelligence, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to collect real-time operation data of the permanent magnet motor during operation by using a multi-dimensional sensor array, perform feature extraction and dimensionality reduction processing on the real-time operation data based on the mutual information entropy criterion, and obtain motor characteristic parameters of the permanent magnet motor; construct a multi-physics field coupling model including electromagnetic field distribution, thermal field distribution, and stress field distribution; dynamically map and match the motor characteristic parameters with the multi-physics field coupling model, and establish a permanent magnet motor digital twin model with adaptive feature fusion capability; The second unit is used to build a two-layer deep reinforcement learning neural network based on the permanent magnet motor digital twin model, and the two-layer deep reinforcement learning neural network includes a high-level strategy network and a low-level execution network; the high-level strategy network adopts a long short-term memory network structure to generate a global optimization target of the permanent magnet motor; The low-level execution network uses a deep deterministic policy gradient algorithm to decompose the global optimization goal into multiple sub-goals; a reward function is constructed for each sub-goal, and the reward function includes an energy efficiency index, a dynamic response index, and a stability index; a priority experience replay mechanism is used to train the two-layer deep reinforcement learning neural network to obtain an optimal flux control strategy and an optimal torque control strategy; based on the optimal flux control strategy and the optimal torque control strategy, an adaptive particle swarm algorithm is used to perform multi-objective optimization of control parameters; The third unit is used to input the optimized control parameters into a permanent magnet motor vector controller with a model prediction function; the permanent magnet motor vector controller is based on the sliding mode variable structure control principle, constructs a three-closed-loop control structure including a torque loop, a flux loop and a current loop; calculates the deviation between the flux observation value and the given value of the permanent magnet motor in real time, and uses an adaptive fuzzy neural network to compensate the d-axis current in real time.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images