MRAS permanent magnet synchronous motor parameter online identification method based on BP neural network optimization

Abstract

The invention provides an MRAS permanent magnet synchronous motor parameter online identification method based on BP neural network optimization, and the method comprises the steps: constructing two BP neural network models, carrying out the offline training optimization of the two models, and obtaining an optimal neural network model which can accurately predict the flux linkage of a rotor and the inductance of a stator; based on an MRAS step-by-step identification strategy, two-stage motor parameter step-by-step identification units are built, the two optimal neural network models are embedded into the corresponding motor parameter step-by-step identification units, and based on the problem that the stator resistance identification effect of the BP neural network is poor, the two optimal neural network models are embedded into the corresponding motor parameter step-by-step identification units; according to the invention, the BP neural network and the adaptive rate unit are established in the second MRAS parameter identification unit, and the online dynamic identification of the stator resistance and the stator inductance is realized through the combination of the BP neural network and the adaptive rate unit. According to the method, an off-line training-on-line identification mode is adopted, the complex parameter setting process in a traditional method is avoided, the identification efficiency is remarkably improved, and the identification result is accurate.

Description

Technical Field

[0001] This invention belongs to the field of permanent magnet synchronous motor parameter identification technology, specifically involving an online parameter identification method for MRAS permanent magnet synchronous motors based on BP neural network optimization. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs) are widely used in industrial drives, home appliances, aerospace, and other fields due to their advantages such as small size, high power density, and light weight. However, high-performance control of these motors depends on accurate motor parameter identification, and the online identification of motor parameters faces significant challenges due to multi-parameter coupling and changes in operating conditions.

[0003] Currently, online parameter identification methods mainly include Least Squares (LS), Extended Kalman Filter (EKF), and Model Reference Adaptive System (MRAS). Least Squares is simple to implement but computationally intensive and memory-intensive. Improvements such as recursive least squares can reduce computational burden but are sensitive to noise and easily affected by old data in time-varying parameter systems, exhibiting slow dynamic response to new data. EKF effectively suppresses noise and improves identification accuracy, but involves numerous matrix operations during the identification process, placing high demands on controller performance. In contrast, MRAS, due to its simple structure, low computational cost, and good convergence, has become a commonly used method for online parameter identification of PMSMs.

[0004] However, traditional MRAS methods still suffer from two significant problems: 1. Difficulty in adjusting the adaptive rate: When identifying multiple parameters simultaneously, the initial values ​​of the parameters to be identified and the adaptive rate parameters need to be manually adjusted. Repeated trial and error is required to determine suitable adaptive rate parameters, lacking systematic optimization methods. 2. Underrank problem: Since the MRAS method model is based on the two-dimensional state equation of the motor's steady-state current, traditional schemes cannot simultaneously identify three key parameters (stator resistance, stator inductance, and rotor flux linkage), resulting in limited identification accuracy.

[0005] To address the aforementioned issues, existing research primarily focuses on improvements in two areas: intelligent optimization methods and step-by-step identification strategies. Regarding intelligent optimization: Liu ZH et al. proposed an adaptive PI scheme switching mechanism based on system parameter changes to identify stator resistance. The system selects a suitable PI scheme based on the amplitude of the current error variable, thereby improving the identification accuracy of the stator resistance. However, the PI controller parameters and switching threshold of this method need to be determined through simulation or empirical debugging, making it difficult to directly transfer to other motor models or application scenarios. Kakodia SK's team introduced a neural network controller into MRAS as its adaptive mechanism, reducing the dependence of traditional MRAS methods on adaptive rate parameter tuning, thus improving identification accuracy or reducing the difficulty of parameter tuning. However, these methods are currently limited to motor speed or single parameter identification and have not yet been extended to multi-parameter motor identification, resulting in limited generalization ability. Regarding step-by-step identification strategies: Huang S et al. proposed a multi-parameter identification method based on cascaded MRAS, solving the underrank problem through incremental equations. This method assumes an electric angular velocity... The parameters are kept constant within adjacent control cycles to eliminate unknowns; however, this assumption is difficult to satisfy under varying operating conditions. Zhang Yi et al. adopted a step-by-step strategy: first, fix the stator resistance and stator inductance according to the motor nameplate to identify the rotor flux linkage, and then use the identification results to deduce the stator resistance and stator inductance. However, when there is a large deviation between the initial fixed parameters and their actual values, it may lead to flux linkage estimation errors, thus affecting the accuracy of subsequent parameter identification.

[0006] In summary, although these methods can solve the underrank problem, the challenge of tuning the adaptive rate parameter remains because these methods are still based on the traditional MRAS framework. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing PMSM online parameter identification technologies, which cannot simultaneously solve the difficulties in adaptive rate tuning and the underrank problem, and to provide an online parameter identification method for MRAS permanent magnet synchronous motors based on BP neural network optimization. Based on the adaptive rate formula derived from MRAS theory, this invention constructs a BP neural network model reflecting the nonlinear mapping relationship between motor operating conditions and the parameters to be identified. Compared with existing MRAS step-by-step identification strategies, this method can simultaneously overcome the difficulties in adaptive rate tuning and the underrank problem of traditional PMSM multi-parameter online identification technology. It not only effectively reduces the dependence of the identification system on manual tuning of adaptive rate parameters, but also improves identification accuracy and efficiency, meeting the needs of high-performance motor control.

[0008] To achieve the above objectives, the technical solution provided by this invention is:

[0009] The online parameter identification method for MRAS permanent magnet synchronous motors based on BP neural network optimization includes the following steps:

[0010] Step 1: Collect data on the permanent magnet synchronous motor under different parameter combinations And motor operating status data under multiple working conditions, and generate a training dataset within the feasible domain of motor parameters; among which, , and These are the stator resistance, stator inductance, and rotor flux linkage, respectively.

[0011] Step 2: Construct a first BP neural network model and a second BP neural network model; wherein, the first BP neural network model is used to learn the dynamic mapping relationship between rotor flux and motor operating state to predict the output rotor flux; the second BP neural network model is used to learn the dynamic mapping relationship between stator inductance and motor operating state to predict the output stator inductance.

[0012] Step 3: Set up a training optimization strategy and use the training dataset from Step 1 to train and optimize the first BP neural network model and the second BP neural network model respectively.

[0013] Step 4: Construct a step-by-step MRAS parameter identification module based on BPNN optimization, wherein the step-by-step MRAS parameter identification module based on BPNN optimization includes a permanent magnet synchronous motor reference unit, a first MRAS parameter identification unit, and a second MRAS parameter identification unit; the first MRAS parameter identification unit is used to identify the rotor flux linkage according to the motor operating conditions; the second MRAS parameter identification unit is used to identify the stator resistance and the stator inductance according to the identification results of the first MRAS parameter identification unit.

[0014] The permanent magnet synchronous motor reference unit includes a first motor reference model and a second motor reference model, and is configured to: enable the first motor reference model when the first MRAS parameter identification unit performs parameter identification, and enable the second motor reference model when the second MRAS parameter identification unit performs parameter identification.

[0015] The first MRAS parameter identification unit includes a first adjustable model of the motor and a trained first BP neural network model; the stator resistance and stator inductance in the first adjustable model of the motor are determined according to the motor parameter nameplate;

[0016] The second MRAS parameter identification unit includes a second adjustable motor model, a trained second BP neural network model, and an adaptive rate unit; the adaptive rate unit is used to calculate output parameters based on the output of the second adjustable motor model. and parameters And feed both back to the second adjustable model of the motor, while simultaneously sending the parameters It is passed to the second BP neural network model; where, It is the reciprocal of the estimated value of the stator inductance. , This is an estimated value for the stator resistance. This is an estimated value for the stator inductance;

[0017] Step 5: Receive the operating parameters of the permanent magnet synchronous motor to be identified online, and output the rotor flux linkage, stator resistance and stator inductance through the MRAS parameter step-by-step identification module based on BPNN optimization.

[0018] Further, step 1 includes the following sub-steps:

[0019] Within the feasible region of stator resistance, stator inductance, and rotor flux linkage, several uniformly distributed combinations of motor parameters are generated by Latin hypercube sampling.

[0020] Multiple motor operating conditions are set up, and dynamic simulations are performed on each set of motor parameters under different motor operating conditions. The motor state parameters and model estimation parameters required by the two BP neural network models are collected in real time, including: shaft voltage , shaft voltage Rotor electrical angle , Shaft current error , Shaft current error Rotor flux linkage estimation value The reciprocal of the estimated stator inductance This yields several sets of time series data;

[0021] The training set data is formed using the motor parameter combination and the motor operating state quantities contained in the time series data.

[0022] Furthermore, the feasible domain of motor parameters is: stator resistance stator inductor Rotor flux .

[0023] Furthermore, the motor operating conditions include the operating time and motor speed of the starting phase, medium-speed loading phase, medium-speed unloading phase, medium-speed stabilization phase, deceleration phase, and low-speed no-load phase. and load torque .

[0024] Furthermore, in step 2, both the first BP neural network model and the second BP neural network model constructed include an input layer, two hidden layers, and an output layer; wherein,

[0025] The first BP neural network model has an input layer consisting of three neurons, each used to receive signals. Shaft current error Rotor electrical angle Rotor flux linkage estimation value The first hidden layer contains 10 neurons, and the second hidden layer contains 8 neurons. Both hidden layers use the tanh function as the activation function to achieve nonlinear mapping. The output layer is used to output the rotor flux identification value. Furthermore, the neuron nodes of the input layer of the first BP neural network model are established based on the following equation:

[0026]

[0027] In the formula, Represents the Laplace variable; , For adaptive rate gain, and ; This is the estimated value of the rotor flux linkage. for True value of shaft current Compared with the estimated value The difference, The rotor electrical angle; The initial value of the stator inductance, This is the initial value of the rotor flux linkage. and All parameters are determined based on the motor nameplate.

[0028] The second BP neural network model has an input layer consisting of 7 neurons, each used to receive signals. shaft voltage , shaft voltage Rotor electrical angle , Shaft current error , Shaft current error Rotor flux identification value and the reciprocal of the estimated stator inductance The hidden layer and output layer have the same structure as the first BP neural network model, and its output layer is used to output the identification value of the stator inductor. Furthermore, the neuron nodes of the input layer of the second BP neural network model are established according to the following equation:

[0029]

[0030] In the formula, Represents the Laplace variable; and All are adaptive rate gains, and ; It is the reciprocal of the estimated value of the stator inductance. for True value of shaft current Compared with the estimated value The difference, for True value of shaft current Compared with the estimated value The difference, for shaft voltage, for Shaft voltage; The rotor flux linkage is the rotor flux linkage identification value output by the first BP neural network model. ; The initial value of the stator inductance The information is derived from the motor's nameplate.

[0031] Furthermore, in step 3, the first BP neural network model and the second BP neural network model are trained and optimized using the training dataset generated in step 1. During the training process, the gradient descent algorithm is used to dynamically adjust the connection weights between each layer to obtain the optimal BP neural network model that reflects the mapping relationship between the motor operating state and the parameters to be identified.

[0032] Furthermore, in step 4, the first motor reference model is:

[0033] In the formula, for shaft current, for shaft current, for shaft voltage, For rotor electrical angle, For rotor flux linkage; The initial value of the stator inductance, The initial values ​​of the stator resistance are determined based on the motor parameter nameplate.

[0034] The first adjustable model of the motor is:

[0035] In the formula, for Estimated value of shaft current, for shaft current, for shaft voltage, The rotor electrical angle; This is the rotor flux linkage estimate, which is the rotor flux linkage identification value output by the first BP neural network model at the previous time step. Obtained after delay by the delay module; The initial value of the stator inductance is derived from the motor's nameplate. The initial value of the stator resistance is derived from the motor parameter nameplate.

[0036] The second motor reference model is ;

[0037] in, For differential operators, , , ;

[0038] In the formula, for shaft current, for shaft current, for shaft voltage, for shaft voltage, For rotor electrical angle, For rotor flux linkage; , ,in This is the actual value of the stator resistance. This is the actual value of the stator inductance;

[0039] The second adjustable model of the motor is:

[0040] in, , ;

[0041] In the formula, for Estimated value of shaft current, for Estimated value of shaft current, for shaft voltage, for shaft voltage, For rotor electrical angle, The value is the rotor flux linkage, which is the rotor flux linkage identification value output by the first MRAS parameter identification unit. ; , ,in This is an estimated value for the stator resistance. This is an estimated value for the stator inductance;

[0042] The adaptive rate unit is:

[0043]

[0044] In the formula, Represents the Laplace variable; , For adaptive gain, and initial Set between 5 and 25, initially Set between 0.5 and 2; , This is an estimated value for the stator resistance. This is an estimated value for the stator inductance; It is the reciprocal of the estimated value of the stator inductance. for Estimated value of shaft current, for Estimated value of shaft current, for True value of shaft current Compared with the estimated value The difference, for True value of shaft current Compared with the estimated value The difference, for shaft voltage, for Shaft voltage; The value is the rotor flux linkage, which is the rotor flux linkage identification value output by the first MRAS parameter identification unit. ; , The initial value of the stator resistance with initial value of stator inductance The information is derived from the motor's nameplate.

[0045] Furthermore, in step 5, the BPNN-optimized MRAS parameter step-by-step identification module performs the following steps during the parameter identification process:

[0046] Step 5.1: Simultaneously acquire data from the first motor reference model and the first adjustable motor model during motor operation. shaft current , shaft current Rotor electrical angle as well as shaft voltage ;

[0047] The first adjustable model of the motor calculates the output. shaft current estimate ;

[0048] Will shaft current The output of the first adjustable model of the motor shaft current estimate Difference, output Shaft current error ;

[0049] Step 5.2: The first BP neural network receives the output from step 5.1. Shaft current error Simultaneously receive the rotor electrical angle of the motor. and rotor flux estimation value Predict and output the rotor flux identification value at the current moment. Among them, the rotor flux linkage estimate The rotor flux identification value predicted and output by the first BP neural network at the previous moment. Obtained after delay by the delay module;

[0050] Step 5.3: Simultaneously acquire data from the second motor reference model and the second adjustable motor model during motor operation. shaft voltage , shaft voltage , shaft current i d , shaft current and rotor electrical angle ;

[0051] Motor Second Adjustable Model Calculation Output shaft current estimate and shaft current estimate ;Will shaft current and shaft current The outputs of the second adjustable model of the motor are respectively shaft current estimate and shaft current estimate Difference, output Shaft current error and Shaft current error ;

[0052] Step 5.4: The adaptive rate unit acquires the rotor electrical angle during motor operation. , shaft voltage , shaft voltage , shaft current , shaft current , Shaft current error , Shaft current error and rotor flux identification value The parameters were calculated. and parameters and parameters and parameters The parameters are synchronously fed back to the second adjustable model of the motor, and... Transmitted to the second BP neural network model;

[0053] Simultaneously, the second BP neural network model obtains the rotor electrical angle during motor operation. , Shaft current error , Shaft current error , shaft voltage , shaft voltage Rotor flux identification value and parameters Predicted output stator inductance identification value ;

[0054] According to parameters and stator inductance identification value And according to the formula The stator resistance identification value was calculated. .

[0055] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements step 5 of the above-described online parameter identification method for MRAS permanent magnet synchronous motors based on BP neural network optimization.

[0056] The present invention also provides an electronic device for data processing, comprising: one or more processors; and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement step 5 of the above-described online identification method for MRAS permanent magnet synchronous motor parameters based on BP neural network optimization.

[0057] The advantages of this invention are:

[0058] 1. In the online parameter identification method for permanent magnet synchronous motors proposed in this invention, two BP neural network models are first constructed. Based on the designed model structure and training optimization strategy, the two neural network models are optimized to obtain BP neural network models that can reliably identify rotor flux linkage and stator resistance and stator inductance respectively. Then, a step-by-step identification module for motor parameters based on a two-level MRAS step-by-step identification strategy is built. The two BP neural network models are embedded into the two-level identification units respectively to realize the step-by-step identification of rotor flux linkage and stator resistance and stator inductance under dynamic motor conditions. This solves the problem of the parameter tuning link that needs to be manually trial and error in the traditional identification process, and improves the efficiency and accuracy of parameter identification.

[0059] 2. In this invention, simulations revealed that BP neural networks cannot effectively learn the complex relationship between stator resistance and motor operating parameters. They only demonstrate good identification of stator resistance under certain motor operating conditions, with significant fluctuations in the latter two conditions. Therefore, to ensure the convergence and real-time performance of the system identification, the applicant designed a second MRAS parameter identification unit comprising a second adjustable motor model, a trained second BP neural network model, and an adaptive rate unit. The adaptive rate unit uses a traditional PI adaptive law for the initial dynamic response and employs the added second BP neural network model as an error compensator, utilizing its nonlinear modeling capability to correct the PI output. Therefore, an initial... and Furthermore, in the application process after offline training of the second BP neural network, it is not necessary to... and Then perform manual parameter tuning, because initially... and The sole purpose is to ensure the convergence of the second BP neural network model, not its recognition performance. This design combines the convergence advantages of traditional PI with the recognition accuracy compensation capability of BP neural networks, effectively reducing the impact on... or Dependence on precise tuning.

[0060] 3. In the method of this invention, the prediction accuracy of the network is ensured by optimizing the two BP neural network models offline. Evaluation using the Matlab / Simulink simulation platform shows that the convergence time of the motor parameter identification process in this invention is controlled within 0.22s, and the maximum error is controlled within 3.5%. Compared with the traditional MRAS step-by-step identification strategy, the maximum errors of rotor flux linkage, stator resistance, and inductance are reduced by 75%, 22.2%, and 37.7%, respectively, and the identification time is shortened by 79.8%.

[0061] 4. In the method of the present invention, the offline training-online identification approach is adopted, which reduces the average execution time of a single motor parameter identification from 972.6s to 196.1s, significantly reducing the real-time computing burden. Attached Figure Description

[0062] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0063] Figure 1 This is a block diagram of the online parameter identification method for MRAS permanent magnet synchronous motor based on BP neural network of the present invention;

[0064] Figure 2 This is a structural diagram of the BP neural network model constructed in this invention;

[0065] Figure 3 This is a schematic diagram of the prediction results using the first BP neural network model constructed in this invention;

[0066] Figure 4 This is a schematic diagram of the prediction results using the second BP neural network model constructed in this invention;

[0067] Figure 5 This is a Simulink simulation structure block diagram of the MRAS parameter identification unit in this invention;

[0068] Figure 6 This is a schematic diagram of a PMSM parameter identification system including the BPNN-optimized stepwise MRAS parameter identification module of this invention.

[0069] Figure 7 This is a comparative schematic diagram of rotor flux linkage identification results based on the method of the present invention and the conventional method. In the figure, 7a is the rotor flux linkage identification result of the conventional method, and 7b is the rotor flux linkage identification result of the method of the present invention.

[0070] Figure 8 This is a comparative diagram of stator resistance identification results based on the method of the present invention and the conventional method. In the figure, 8a is the stator resistance identification result of the conventional method and 8b is the stator resistance identification result of the method of the present invention.

[0071] Figure 9 This is a schematic diagram comparing the stator inductance identification results based on the method of the present invention and the conventional method. In the figure, 9a is the stator inductance identification result of the conventional method, and 9b is the stator inductance identification result of the method of the present invention. Detailed Implementation

[0072] The embodiments of the present invention are described in detail below. These embodiments are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0073] To address the difficulties in adaptive rate tuning and underrank issues during multi-parameter identification of PMSMs, this invention provides an online parameter identification method for MRAS permanent magnet synchronous motors based on a BP neural network, referring to... Figure 1 This includes the following steps:

[0074] Step 1: Collect data on the permanent magnet synchronous motor under different parameter combinations It also includes motor operation data under multiple operating conditions, and generates a training dataset within the feasible domain of motor parameters.

[0075] Step 1.1: Simulation parameters and operating conditions design.

[0076] The parameters of the permanent magnet synchronous motor (PMSM) are shown in Table 1. To verify the adaptability of the MRAS step-by-step identification algorithm based on BPNN optimization under multiple operating conditions, the embodiments of this invention designed operating conditions including dynamic excitation and steady-state testing as shown in Table 2. The system verifies the identification effect of the algorithm on motor parameters under different operating conditions. Among them, the starting stage is the motor acceleration process with a load of 1N·m, and the speed rapidly increases from 0 r / min to 1333 r / min (2 / 3 of the rated speed); the medium-speed loading stage and the medium-speed unloading stage are used to verify the sensitivity of the identification algorithm to the load. Under the premise of maintaining a medium speed of 1333 r / min, the load torque is suddenly increased from 1N·m to 3N·m in 0.2s and unloaded back to 1N·m in 1.5s; in the medium-speed steady-state stage, the motor maintains a constant speed and a stable 1N·m load; the deceleration stage and the low-speed no-load stage are used to verify the parameter identification effect under deceleration conditions and low-speed steady-state conditions.

[0077] Table 1

[0078]

[0079] Table 2

[0080]

[0081] Step 1.2: Training dataset generation.

[0082] This invention generates a training dataset within the feasible region of motor parameters through Latin Hypercube Sampling (LHS), specifically implemented as follows:

[0083] In parameter space , , Within the specified range, 490 uniformly distributed parameter combinations were generated using LHS. Each parameter combination was dynamically simulated under the multi-condition working conditions shown in Table 2. During the simulation, dynamic parameters required by the adaptive rate modules of the two MRAS were collected. Each simulation collected 500 sets of time-series data, ultimately forming a training dataset containing 245,000 samples. The training dataset was divided into a training set, a test set, and a cross-validation set, with the training set accounting for 70%, and the test set and cross-validation set each accounting for 15%. The data was standardized using the Z-score method to form standardized training data for subsequent neural network model training and optimization.

[0084] Step 2: Construct the first BP neural network model and the second BP neural network model.

[0085] The first BP neural network model is used to learn the dynamic mapping relationship between rotor flux linkage and motor operating state, so as to predict the output rotor flux linkage based on the input motor operating state data. The second BP neural network model is used to learn the dynamic mapping relationship between stator resistance, stator inductance and motor operating state, so as to predict the output stator inductance.

[0086] The first and second BP neural network models constructed in this invention have the same network architecture, both including an input layer, two hidden layers, and an output layer, such as... Figure 2 As shown.

[0087] The specific configuration of the first BP neural network model is as follows: The input layer is designed based on equation (1) with 3 neuron nodes, respectively... Shaft current error Rotor electrical angle Rotor flux linkage estimation value As input features. The first hidden layer contains 10 neurons, that is... The second hidden layer contains 8 neurons, that is... Both hidden layers use the tanh function as the activation function to achieve nonlinear mapping. The output layer is used to output the true value of the rotor flux linkage. By building , , and The nonlinear mapping relationship between them achieves an accurate approximation of equation (1). Equation (1) is:

[0088] (1)

[0089] In the formula, Represents the Laplace variable; and For adaptive rate gain, where For integral gain, The gain is proportional; all are greater than 0. This is the estimated value of the rotor flux linkage. for True value of shaft current Compared with the estimated value The difference, The rotor electrical angle; The initial value of the stator inductance is determined based on the motor parameter nameplate. This is the initial value of the rotor flux linkage, determined according to the motor parameter nameplate.

[0090] The specific configuration of the second BP neural network model is as follows: Maintaining the same network architecture as the first BP neural network model, the first hidden layer contains 10 neurons, i.e. The second hidden layer contains 8 neurons, that is... Both hidden layers use the tanh function as the activation function to achieve nonlinear mapping. The input layer is designed with 7 neuron nodes based on equation (2), each with a specific number of neurons. shaft voltage , shaft voltage Rotor electrical angle , Shaft current error , Shaft current error Rotor flux identification value and the reciprocal of the estimated stator inductance As input features, the output layer is used to output the true value of the stator inductance. Equation (2) is:

[0091] (2)

[0092] In the formula, Represents the Laplace variable; and These are adaptive rate gains, and all are greater than 0; , This is an estimated value for the stator resistance. This is an estimated value for the stator inductance; It is the reciprocal of the estimated value of the stator inductance, i.e. ; for Estimated value of shaft current, for Estimated value of shaft current, for True value of shaft current Compared with the estimated value The difference, for True value of shaft current Compared with the estimated value The difference, for shaft voltage, for Shaft voltage; For rotor flux linkage, the rotor flux linkage identification value output by the first BP neural network model is taken. ; , The initial value of the stator resistance with initial value of stator inductance Determine based on the motor's nameplate parameters.

[0093] Step 3: Set up training optimization strategies and use the training dataset from Step 1 to train and optimize the first BP neural network model and the second BP neural network model respectively.

[0094] The training optimization strategy is as follows: In this embodiment, the mean squared error (MSE) is used as the loss function, the learning step size is set to 0.001, and the gradient descent algorithm is adaptively optimized through the Adam optimizer to dynamically adjust the connection weights between each layer. At the same time, an early stopping mechanism is introduced to prevent overfitting, so as to obtain the optimal BP neural network model that reflects the mapping relationship between the motor operating state and the parameters to be identified.

[0095] Based on the aforementioned training optimization strategy, the first BP neural network model and the second BP neural network model are trained and optimized offline using the training dataset divided in step 1.

[0096] Training and optimization of the first BP neural network model: After 10,000 training rounds, a first BP neural network model with optimal weights was obtained. After training, the partitioned test set was fed into the trained first BP neural network model to predict the output rotor flux linkage. The prediction results are as follows: Figure 3 As shown in the figure, the predicted and actual values ​​exhibit a high degree of consistency. The predictive performance of the first BP neural network model was quantitatively evaluated using three metrics: Mean Squared Error (MSE), Mean Absolute Error (MAE), and Coefficient of Determination (R²). The MSE was 0.0075, the MAE was as low as 0.0018, and the R² reached 0.9920, close to the theoretical optimal value of 1. This indicates that the trained and optimized first BP neural network possesses excellent prediction accuracy and generalization ability, and can accurately predict the output rotor flux linkage based on the actual operating conditions of the motor. .

[0097] Training and optimization of the second BP neural network model: After 10,000 training rounds, a second BP neural network model with optimal weights is obtained. After training, the partitioned test set is fed into the trained second BP neural network model to predict the output stator inductance. The prediction results are as follows: Figure 4 As shown in the figure, the predicted values ​​and actual values ​​exhibit a high degree of consistency. Mean Squared Error (MSE), Mean Absolute Error (MAE), and coefficient of determination were used. These three metrics quantitatively evaluate the predictive performance of the second BP neural network model. The mean squared error (MSE) is 0.0023, the mean absolute error (MAE) is as low as 0.00003, and the coefficient of determination is... The result of 0.9978 indicates that the second BP neural network model has excellent prediction accuracy and generalization ability, and can accurately predict the output stator inductance based on the actual operating conditions of the motor.

[0098] Step 4: Construct a step-by-step identification module for MRAS parameters based on BP neural network optimization.

[0099] Based on the step-by-step identification strategy for MRAS parameters, a BP neural network-optimized MRAS parameter step-by-step identification module is constructed. This module includes a permanent magnet synchronous motor reference unit, a first MRAS parameter identification unit, and a second MRAS parameter identification unit, all connected in series. The first MRAS parameter identification unit identifies the rotor flux linkage based on the motor's operating conditions. The second MRAS parameter identification unit identifies the stator resistance and stator inductance based on the identification results of the first MRAS parameter identification unit and other motor operating parameters. The permanent magnet synchronous motor reference unit includes a first motor reference model and a second motor reference model, configured to activate the first motor reference model during parameter identification by the first MRAS parameter identification unit and the second motor reference model during parameter identification by the second MRAS parameter identification unit.

[0100] The first motor reference model is:

[0101] (3)

[0102] In the formula, for True value of shaft current for True value of shaft current for shaft voltage, The rotor electrical angle; For rotor flux linkage; The initial value of the stator inductance is determined based on the motor parameter nameplate. The initial value of the stator resistance is determined based on the motor parameter nameplate.

[0103] The second motor reference model is

[0104] (4)

[0105] in, For differential operators, , ,

[0106] (5)

[0107] In the formula, for shaft current, for True value of shaft current for shaft voltage, for shaft voltage, For rotor electrical angle, For rotor flux linkage; , ,in This is the actual value of the stator resistance. This is the actual value of the stator inductance.

[0108] The first MRAS parameter identification unit includes a first adjustable model of the motor and a first BP neural network model that has been trained and predicted, connected in series. The stator resistance and stator inductance in the first adjustable model of the motor are determined according to the motor parameter nameplate. The first adjustable model of the motor is as follows:

[0109] (6)

[0110] In the formula, for Estimated value of shaft current, for shaft current, for shaft voltage, For rotor electrical angle, This is the estimated value of the rotor flux linkage; The initial value of the stator inductance is derived from the motor parameter nameplate; The initial value of the stator resistance is derived from the motor parameter nameplate.

[0111] It should be noted that during the construction of the second MRAS parameter identification unit, the applicant's initial improvements included a second adjustable model of the motor and a trained BP neural network model. The BP neural network model was used to simultaneously predict the stator resistance and stator inductance, thereby eliminating the need for manual parameter tuning in the traditional method. or The process is complex. However, during simulation, the applicant discovered that when the second MRAS parameter identification unit fully utilizes the BP neural network model for identification, it leads to instability in the feedback system, especially a significant decrease in the identification performance for stator resistance. Analysis revealed that this is due to a closed-loop coupling problem in the identification process. Specifically, in the second MRAS parameter identification unit, the output of the BP neural network affects the input of the second adjustable motor model, and vice versa, resulting in mutual dependence and influence between input and output. Furthermore, the BP neural network cannot effectively learn the complex relationship between stator resistance and motor operating parameters, achieving good identification results only under the first four motor operating conditions shown in Table 2, while the identification results fluctuate significantly under the latter two conditions. Therefore, to ensure the convergence and real-time performance of the system identification, the applicant designed the second MRAS parameter identification unit, which includes a second adjustable motor model, a trained second BP neural network model, and an adaptive rate unit. The adaptive rate unit retains the traditional PI adaptive law for the initial dynamic response and uses the added second BP neural network model as an error compensator, utilizing its nonlinear modeling capability to correct the PI output. Therefore, the initial... and Furthermore, in the application process after offline training of the second BP neural network, it is not necessary to further train it. and Then perform manual parameter tuning, because initially... and The sole purpose of this optimization is to ensure the convergence of the second BP neural network model, not its recognition performance. This optimization combines the convergence advantage of the traditional PI model with the recognition accuracy compensation capability of the BP neural network, effectively reducing the impact on... and The dependence on precise tuning. Simulation verification shows that even without optimizing the PI parameters (i.e., only manual initial setting is required), and The second MRAS parameter identification unit can still achieve high parameter identification accuracy through the compensation effect of the second BP neural network, achieving the goal of "no manual parameter tuning required". In this invention, based on experience, the initial... Choose between 0.5 and 2, initially. Choosing a value between 5 and 25 allows the second MRAS parameter identification unit to perform stable identification operations without drastic output fluctuations, thus enabling offline training of the second BP neural network model.

[0112] The second adjustable model of the motor is as follows:

[0113] (7)

[0114] In the formula, , ; It is a differential operator.

[0115] In the formula, for shaft current estimate for Estimated value of shaft current, for shaft voltage, for shaft voltage, For rotor electrical angle, The value is the rotor flux linkage, which is the rotor flux linkage identification value output by the first MRAS parameter identification unit. ; , ,in This is an estimated value for the stator resistance; This is the estimated value for the stator inductance.

[0116] In this invention, the structural block diagrams of the first MRAS parameter identification unit and the second MRAS parameter identification unit are as follows: Figure 5 As shown.

[0117] Step 5: Connect the constructed MRAS parameter step-by-step identification module based on BP neural network optimization to the PMSM parameter identification system, such as... Figure 6 As shown, the operating parameters of the permanent magnet synchronous motor to be identified are obtained online, and the rotor flux identification value is output by the MRAS parameter step-by-step identification module based on BP neural network optimization. Stator resistance identification value and stator inductance identification value .

[0118] The specific execution process of the step-by-step identification module for MRAS parameters based on the BP neural network of this invention to identify and output motor parameters is as follows:

[0119] Step 5.1: Simultaneously acquire data from the first motor reference model and the first adjustable motor model during motor operation. shaft current , shaft current Rotor electrical angle as well as shaft voltage ;

[0120] The first adjustable model of the motor calculates the output. shaft current estimate ;

[0121] Will shaft current The output of the first adjustable model of the motor shaft current estimate Difference, output Shaft current error .

[0122] Step 5.2: The first BP neural network receives the output from step 5.1. Shaft current error Simultaneously receive the rotor electrical angle of the motor. and rotor flux estimation value Predict and output the rotor flux identification value at the current moment. Among them, the rotor flux linkage estimate The rotor flux identification value predicted and output by the first BP neural network at the previous moment. It is obtained after being delayed by the delay module.

[0123] Step 5.3: Simultaneously acquire the d-axis voltage during motor operation using both the second motor reference model and the second adjustable motor model. , shaft voltage , shaft current , shaft current and rotor electrical angle .

[0124] Motor Second Adjustable Model Calculation Output shaft current estimate and shaft current estimate ; will shaft current and shaft current The outputs of the second adjustable model of the motor are respectively shaft current estimate and shaft current estimate Difference, output Shaft current error and Shaft current error .

[0125] Step 5.4: The adaptive rate unit acquires the rotor electrical angle during motor operation. , shaft voltage , shaft voltage , shaft current , shaft current , Shaft current error , Shaft current error and rotor flux identification value The parameters were calculated. and parameters and parameters and parameters The parameters are synchronously fed back to the second adjustable model of the motor, and... Transmitted to the second BP neural network model;

[0126] Meanwhile, the second BP neural network model obtains the rotor electrical angle during motor operation. , Shaft current error , Shaft current error , shaft voltage , shaft voltage Rotor flux identification value and parameters Predicted output stator inductance identification value .

[0127] because Based on the obtained parameters and stator inductance identification value According to the formula The stator resistance identification value was calculated. .

[0128] To illustrate the identification effect of the online parameter identification method for permanent magnet synchronous motors based on BP neural network optimization of the present invention, this embodiment uses the traditional MRAS identification method and the method of the present invention to identify the rotor flux linkage, stator resistance and stator inductance of the permanent magnet synchronous motor (PMSM) designed based on the parameters in Tables 1 and 2 and the operating conditions, respectively. The identification results are compared by simulation, and then the identification effect of the two methods is evaluated using the maximum error and root mean square error (RMSE).

[0129] Figure 7 The rotor flux linkage identification performance of the two methods is compared and shown. Among them, 7a is the identification result of the traditional MRAS identification method. As can be seen from the figure, there is a significant overshoot phenomenon in the system during the motor start-up stage. The identification result converges to the vicinity of the true value within 0.05s, with a maximum error of 2.00% and a calculated root mean square error (RMSE) of 0.008. Figure 7 b represents the identification result of the identification method of this invention. By adjusting the constructed BP neural network, the system completely eliminates the start-up overshoot. Although the convergence time increases slightly to 0.06s, the maximum error is reduced to 0.5%, and the calculated root mean square error (RMSE) is 0.007, thus improving the identification accuracy.

[0130] Figure 8 The stator resistance identification performance of the two methods is compared and demonstrated. Figure 8 Figure 'a' shows the identification result of the traditional MRAS identification method. As can be seen from the figure, the identification result converges to the vicinity of the true value within 0.21s, with a maximum error of 4.50%. Figure 8 b represents the identification result of the identification method of the present invention. As can be seen from the figure, although the convergence time has increased slightly to 0.22s, the maximum error has been reduced to 3.5%.

[0131] Figure 9 The comparison demonstrates the identification performance of the two methods for stator inductors. Figure 9a shows the identification results of the traditional MRAS method. As can be seen from the figure, the identification results converge to near the true value within 0.5s, with a maximum error of 4.64%, and the calculated root mean square error (RMSE) is 0.0307. Figure 9 b represents the identification result of the identification method of the present invention, with its convergence time shortened to 0.21s, maximum error reduced to 2.89%, and root mean square error (RMSE) reduced to 0.0185.

[0132] Meanwhile, Tables 3 and 4 present the evaluation results of the identification results of the traditional MRAS step-by-step identification method and the identification method of this invention, respectively. The given value for the rotor flux linkage is 0.2 Wb, and the given value for the stator resistance is 1 Wb. The given value for the stator inductance is 5.6 mH. Simulation results show that the BPNN-MRAS identification strategy has better performance than the existing MRAS stepwise identification strategy: the fluctuation range of parameter identification values ​​is reduced, and the system stability is improved; the maximum identification errors of rotor flux linkage, stator resistance, and stator inductance are all controlled within 3.5%, and the identification accuracy is improved.

[0133] Table 3

[0134]

[0135] Table 4

[0136]

[0137] To quantify the computational complexity of the two methods, this embodiment of the invention uses the MATLAB `tic / toc` command to measure the time taken for a single complete simulation of both methods. To eliminate the influence of random fluctuations, each method is run independently 5 times, and the average value is taken as the final result. The traditional MRAS step-by-step identification method has an average time of 972.6 seconds, while the method of this invention, due to the use of an offline-trained BPNN model, reduces the computation time to 196.1 seconds, significantly lowering the computational complexity.

[0138] In summary, the MRAS parameter identification strategy based on BPNN optimization proposed in this invention exhibits good accuracy, speed, and robustness under various operating conditions.

[0139] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in the present invention, and such modifications or substitutions should all be covered within the scope of protection of the present invention.

Claims

1. A method for online parameter identification of MRAS permanent magnet synchronous motors based on BP neural network optimization, characterized in that, Includes the following steps: Step 1: Collect data on the permanent magnet synchronous motor under different parameter combinations And motor operating status data under multiple working conditions, and generate a training dataset within the feasible domain of motor parameters; among which, , and These are the stator resistance, stator inductance, and rotor flux linkage, respectively. Step 2: Construct a first BP neural network model and a second BP neural network model; wherein, the first BP neural network model is used to learn the dynamic mapping relationship between rotor flux and motor operating state to predict the output rotor flux; the second BP neural network model is used to learn the dynamic mapping relationship between stator inductance and motor operating state to predict the output stator inductance. Step 3: Set up a training optimization strategy, and use the training dataset from Step 1 to train and optimize the first BP neural network model and the second BP neural network model respectively; Step 4: Construct a step-by-step MRAS parameter identification module based on BPNN optimization, wherein the step-by-step MRAS parameter identification module based on BPNN optimization includes a permanent magnet synchronous motor reference unit, a first MRAS parameter identification unit, and a second MRAS parameter identification unit; the first MRAS parameter identification unit is used to identify the rotor flux linkage according to the motor operating conditions; the second MRAS parameter identification unit is used to identify the stator resistance and the stator inductance according to the identification results of the first MRAS parameter identification unit. The permanent magnet synchronous motor reference unit includes a first motor reference model and a second motor reference model, and is configured to: enable the first motor reference model when the first MRAS parameter identification unit performs parameter identification, and enable the second motor reference model when the second MRAS parameter identification unit performs parameter identification. The first MRAS parameter identification unit includes a first adjustable model of the motor and a trained first BP neural network model; the stator resistance and stator inductance in the first adjustable model of the motor are determined according to the motor parameter nameplate; The second MRAS parameter identification unit includes a second adjustable motor model, a trained second BP neural network model, and an adaptive rate unit; the adaptive rate unit is used to calculate output parameters based on the output of the second adjustable motor model. and parameters And feed both back to the second adjustable model of the motor, while simultaneously sending the parameters It is passed to the second BP neural network model; where, It is the reciprocal of the estimated value of the stator inductance. , This is an estimated value for the stator resistance. This is an estimated value for the stator inductance; Step 5: Receive the operating parameters of the permanent magnet synchronous motor to be identified online, and output the rotor flux linkage, stator resistance and stator inductance through the MRAS parameter step-by-step identification module based on BPNN optimization.

2. The online parameter identification method for MRAS permanent magnet synchronous motors based on BP neural network optimization according to claim 1, characterized in that, Step 1 includes the following sub-steps: Within the feasible region of stator resistance, stator inductance, and rotor flux linkage, several sets of uniformly distributed motor parameter combinations are generated by Latin hypercube sampling. Multiple motor operating conditions are set up, and dynamic simulations are performed on each set of motor parameters under different motor operating conditions. The motor state parameters and model estimation parameters required by the two BP neural network models are collected in real time, including: shaft voltage , shaft voltage Rotor electrical angle , Shaft current error , Shaft current error Rotor flux linkage estimation value The reciprocal of the estimated stator inductance This yields several sets of time series data; Training set data is formed using the motor parameter combination and the motor operating state quantities contained in the time series data.

3. The online parameter identification method for MRAS permanent magnet synchronous motors based on BP neural network optimization according to claim 2, characterized in that, The feasible range of motor parameters is: stator resistance. stator inductor Rotor flux .

4. The online parameter identification method for MRAS permanent magnet synchronous motors based on BP neural network optimization according to claim 3, characterized in that, The motor operating conditions include the running time and motor speed during the starting phase, medium-speed loading phase, medium-speed unloading phase, medium-speed stabilization phase, deceleration phase, and low-speed no-load phase. and load torque .

5. The online parameter identification method for MRAS permanent magnet synchronous motor based on BP neural network according to claim 1, characterized in that, In step 2, both the first BP neural network model and the second BP neural network model constructed include an input layer, two hidden layers, and an output layer; wherein, The first BP neural network model has an input layer consisting of three neurons, each used to receive signals. Shaft current error Rotor electrical angle Rotor flux linkage estimation value The first hidden layer contains 10 neurons, and the second hidden layer contains 8 neurons. Both hidden layers use the tanh function as the activation function to achieve nonlinear mapping. The output layer is used to output the rotor flux linkage identification value. Furthermore, the neuron nodes of the input layer of the first BP neural network model are established based on the following equation: In the formula, Represents the Laplace variable; , For adaptive rate gain, and ; This is the estimated value of the rotor flux linkage. for True value of shaft current Compared with the estimated value The difference, The rotor electrical angle; The initial value of the stator inductance, This is the initial value of the rotor flux linkage. and All parameters are determined based on the motor nameplate. The second BP neural network model has an input layer consisting of 7 neurons, each used to receive signals. shaft voltage , shaft voltage Rotor electrical angle , Shaft current error , Shaft current error Rotor flux identification value and the reciprocal of the estimated stator inductance The hidden layer and output layer have the same structure as the first BP neural network model, and its output layer is used to output the identification value of the stator inductor. Furthermore, the neuron nodes of the input layer of the second BP neural network model are established according to the following equation: In the formula, Represents the Laplace variable; and All are adaptive rate gains, and ; It is the reciprocal of the estimated value of the stator inductance. for True value of shaft current Compared with the estimated value The difference, for True value of shaft current Compared with the estimated value The difference, for shaft voltage, for Shaft voltage; The rotor flux linkage is the rotor flux linkage identification value output by the first BP neural network model. ; The initial value of the stator inductance The information is derived from the motor's nameplate.

6. The online parameter identification method for MRAS permanent magnet synchronous motors based on BP neural network optimization according to claim 5, characterized in that, In step 3, the first BP neural network model and the second BP neural network model are trained and optimized using the training dataset generated in step 1. During the training process, the gradient descent algorithm is used to dynamically adjust the connection weights between each layer to obtain the optimal BP neural network model that reflects the mapping relationship between the motor operating state and the parameters to be identified.

7. The online parameter identification method for MRAS permanent magnet synchronous motors based on BP neural network optimization according to claim 1, characterized in that, In step 4 The first motor reference model is: In the formula, for shaft current, for shaft current, for shaft voltage, For rotor electrical angle, For rotor flux linkage; The initial value of the stator inductance, The initial values ​​of the stator resistance are determined based on the motor parameter nameplate. The first adjustable model of the motor is: In the formula, for Estimated value of shaft current, for shaft current, for shaft voltage, The rotor electrical angle; This is the rotor flux linkage estimate, which is the rotor flux linkage identification value output by the first BP neural network model at the previous time step. Obtained after delay by the delay module; The initial value of the stator inductance is derived from the motor's nameplate. The initial value of the stator resistance is derived from the motor parameter nameplate. The second motor reference model is ; in, For differential operators, , , ; In the formula, for shaft current, for shaft current, for shaft voltage, for shaft voltage, For rotor electrical angle, For rotor flux linkage; , ,in This is the actual value of the stator resistance. This is the actual value of the stator inductance; The second adjustable model of the motor is: in, , ; In the formula, for Estimated value of shaft current, for Estimated value of shaft current, for shaft voltage, for shaft voltage, For rotor electrical angle, The value is the rotor flux linkage, which is the rotor flux linkage identification value output by the first MRAS parameter identification unit. ; , ,in This is an estimated value for the stator resistance. This is an estimated value for the stator inductance; The adaptive rate unit is: In the formula, Represents the Laplace variable; , For adaptive gain, and initial Set between 5 and 25, initially Set between 0.5 and 2; , This is an estimated value for the stator resistance. This is an estimated value for the stator inductance; It is the reciprocal of the estimated value of the stator inductance. for Estimated value of shaft current, for Estimated value of shaft current, for True value of shaft current Compared with the estimated value The difference, for True value of shaft current Compared with the estimated value The difference, for shaft voltage, for Shaft voltage; The value is the rotor flux linkage, which is the rotor flux linkage identification value output by the first MRAS parameter identification unit. ; , The initial value of the stator resistance with initial value of stator inductance The information is derived from the motor's nameplate.

8. The online parameter identification method for MRAS permanent magnet synchronous motors based on BP neural network optimization according to claim 1, characterized in that, In step 5, the BPNN-optimized MRAS parameter stepwise identification module performs the following steps during the parameter identification process: Step 5.1: Simultaneously acquire data from the first motor reference model and the first adjustable motor model during motor operation. shaft current , shaft current Rotor electrical angle as well as shaft voltage ; The first adjustable model of the motor calculates the output. shaft current estimate ; Will shaft current The output of the first adjustable model of the motor shaft current estimate Difference, output Shaft current error ; Step 5.2: The first BP neural network receives the output from step 5.

1. Shaft current error Simultaneously receive the rotor electrical angle of the motor. and rotor flux estimation value Predict and output the rotor flux identification value at the current moment. Among them, the rotor flux linkage estimate The rotor flux identification value predicted and output by the first BP neural network at the previous moment. Obtained after delay by the delay module; Step 5.3: Simultaneously acquire data from the second motor reference model and the second adjustable motor model during motor operation. shaft voltage , shaft voltage , shaft current i d , shaft current and rotor electrical angle ; Motor Second Adjustable Model Calculation Output shaft current estimate and shaft current estimate ;Will shaft current and shaft current The outputs of the second adjustable model of the motor are respectively shaft current estimate and shaft current estimate Difference, output Shaft current error and Shaft current error ; Step 5.4: The adaptive rate unit acquires the rotor electrical angle during motor operation. , shaft voltage , shaft voltage , shaft current , shaft current , Shaft current error , Shaft current error and rotor flux identification value The parameters were calculated. and parameters and parameters and parameters The parameters are synchronously fed back to the second adjustable model of the motor, and... Transmitted to the second BP neural network model; Simultaneously, the second BP neural network model obtains the rotor electrical angle during motor operation. , Shaft current error , Shaft current error , shaft voltage , shaft voltage Rotor flux identification value and parameters Predicted output stator inductance identification value ; According to parameters and stator inductance identification value And according to the formula The stator resistance identification value was calculated. .

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements step 5 of the online parameter identification method for MRAS permanent magnet synchronous motor based on BP neural network optimization as described in any one of claims 1-8.

10. An electronic device for data processing, comprising: One or more processors; And the computer-readable storage medium of claim 9, for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement step 5 of the online parameter identification method for MRAS permanent magnet synchronous motor based on BP neural network optimization of any one of claims 1-8.

Images