Small data driven observer, multiphase motor predictive control system and method

Abstract

The application discloses a small-data-driven observer, a multiphase motor predictive control system and method, and the multiphase motor predictive control system comprises a sampling module and a small-data-driven observer, the observer comprises a neural network module and a neural network weight adjustment rhythm module; the neural network module comprises a plurality of neurons, and the neural network module calculates a current k time point current error compensation prediction value through the weight of each neuron according to the feedback value of each phase current of the multiphase motor at the current k time point; the neural network weight adjustment rhythm module calculates an optimal weight adjustment increment of the neural network module at k+1 time point according to the output value and the observation state of the observer at k time point, updates the weight matrix of the neural network module at k time point based on the optimal weight adjustment increment, and obtains the neural network weight matrix at k+1 time point. The application independently updates the network weight matrix, and improves the stability and observation accuracy of the observer at different working points of the multiphase motor.

Description

Technical Field

[0001] This invention relates to a multiphase motor control system and method, and particularly to a small data-driven observer, a multiphase motor predictive control system and method. Background Technology

[0002] Currently, the development of predictive control applications in power electronics is hindered by the over-reliance on the accuracy of predictive model parameters for control precision. The mathematical models of power electronic systems vary under different operating conditions; therefore, compensating for prediction errors caused by errors in predictive model parameters is a crucial technical challenge for predictive control applications in power electronics. Existing classic methods for addressing model parameter variations include lookup tables, online identification, and state observers. However, lookup tables require coverage of all operating points, making it difficult to obtain model change trajectories and necessitating significant storage space. Online identification requires injecting high-frequency signals into the system, and its coupling with the control action may affect the accuracy of the results. The effective operating range of state observers is affected by their parameter settings, requiring real-time parameter tuning, and cannot meet the requirements for fully adaptive model error compensation.

[0003] To achieve fully adaptive model error compensation, existing research has employed data-driven observers based on offline training and online adjustment of neural networks. This approach involves densely sampling voltage and current data within the operating range as a dataset to train an initial value network, which is then adjusted online at different operating points according to predefined weights. However, the requirement for large, intensive datasets severely limits its application in industrial scenarios where repeated testing is prohibited. Therefore, designing small-data-driven observers for adaptive model error compensation in the power electronics field is a crucial approach to overcoming model parameter dependencies in predictive control. Summary of the Invention

[0004] This invention provides a small data-driven observer, a multiphase motor predictive control system, and a method to solve the technical problems existing in the prior art.

[0005] The technical solution adopted by this invention to solve the technical problems existing in the prior art is: a small data-driven observer, which is used to predict the current error compensation value of a multiphase motor, and outputs the current error compensation value. This current error compensation value is superimposed with a current reference value and then input to the multiphase motor driver. The observer includes a neural network module and a neural network weight adjustment law module. The neural network module includes multiple neurons, and the neural network module adjusts the current... k The feedback values ​​of the current in each phase of the multiphase motor at any given time are calculated using the weights of each neuron to obtain the current value. k Current error compensation prediction value at time; neural network weight adjustment law module, which adjusts the value based on the observer's current... k The output value at time and the observed state are used to calculate the future.k The optimal weight adjustment increment of the neural network module at time +1 is based on the optimal weight adjustment increment for the current neural network module. k The weight matrix at each time step is updated to obtain the future... k The neural network weight matrix at time +1.

[0006] Furthermore, the mathematical model of this observer is as follows: ; in: ; ; ; ; ; ; In the formula: The time derivative of the observer's state variables; For observer state variables; For the dq axis voltage vector of a multiphase motor; This refers to the dq-axis current vector of a multiphase motor. The dq-axis current vector estimated by the observer; For model estimation error; The d-axis current component estimated by the observer; The q-axis current component estimated by the observer; G The state matrix; F The input matrix; Output matrix for the observer; This is the inherent disturbance quantity; For the d-axis resistance of a multiphase motor; For the q-axis resistance of a multiphase motor; For the d-axis inductance of a multiphase motor; For the q-axis inductance of a multiphase motor; This refers to the angular velocity of the multiphase motor. The controller sampling period; For permanent magnet flux linkage in multiphase motors; For observer gain; It is a two-dimensional identity matrix; Observer Gain K The values ​​satisfy the matrix ( G - KC o The real part of the eigenvalues ​​of ) is less than zero.

[0007] Furthermore, the neural network weight adjustment law module includes an observer error evaluation model, which is used to evaluate the fitting error of the neural network module. The mathematical model of the observer error evaluation model is as follows: ; ; in: ; ; ; In the formula: ε This represents the fitting error of the neural network module; These are sampled values ​​of the observer's state variables; The error in estimating the observer's state variables; The time derivative of the observer state quantity estimation error; The observer estimation error for the dq-axis current vector of a multiphase motor; This is the input vector for the neural network module; is the activation function for the neural network module; This is the input matrix for the neural network module; This is the weight matrix of the neural network module; This represents the optimal value of the weight matrix for the neural network module; This represents the optimal increment of the weight matrix for the neural network module.

[0008] Furthermore, the neural network weight adjustment module also includes a neural network weight adjustment module; the mathematical model of the neural network weight adjustment module is as follows: ; in: = P It is a positive definite matrix and satisfies , Q It is also a positive definite matrix; In the formula: For the first k The model error estimated by the neural network at the sampling time; For the first k Sampled values ​​of the dq-axis current vector of the multiphase motor at the sampling time; For the first k Predicted values ​​of the dq-axis current vector of the multiphase motor at the sampling time; For the first k The weight matrix correction value of the neural network module at sampling time +1; For the first k The weight matrix of the neural network module at the sampling time; This is the input matrix for the neural network module; For the first k The input vector of the neural network module at the sampling time; To correct the inertia coefficient of the neural network weight matrix; To correct the dynamic coefficients of the neural network weight matrix.

[0009] The present invention also provides a multiphase motor predictive control system, which includes a sampling module and the aforementioned small data-driven observer. The sampling module is used to sample the phase current and phase voltage of each phase winding of the multiphase motor at each time moment, and to sample the electrical angle and electrical angular velocity of the multiphase motor rotor relative to the initial position at each time moment. 。

[0010] Furthermore, the system also includes a predictive control module; the predictive control module takes into account the current error compensation value from the observer and the sampled actual current value, and outputs the optimal control signal to the multiphase motor driver; the predictive control module includes a predictive current trajectory unit and a control signal optimization unit. The predictive current trajectory unit, based on the mathematical model of the multiphase motor, generates the predicted current value for the next moment from the sampled signal at the current moment, and further forms the predicted current trajectory; the control signal optimization unit is used to select... kThe optimal output at time +1 makes the corresponding value in the predicted current trajectory... k The error between the predicted current value at time +1 and the current reference value at that time is the smallest.

[0011] Furthermore, the mathematical model of the control signal optimization unit is as follows: ; in: ; ; ; In the formula: In the first k The role of sampling time i voltage vectors along the dq axes The corresponding cost of control error; In the first k The role of sampling time i voltage vectors The corresponding number k Predicted value of dq-axis current vector at sampling time +1; For the first k +1 sampling time control reference value of dq axis current; In the first k The sampling time allows the action of the first i One dq-axis voltage vector; In the first k The optimal dq-axis voltage vector selected at the sampling time; For the first k The angular velocity of the multiphase motor at the sampling time; For the first k The model error estimated by the neural network at the sampling time; In the first k The set of dq-axis voltage vectors that are allowed to operate at the sampling time.

[0012] The present invention also provides a multiphase motor predictive control method utilizing the above-described multiphase motor predictive control system, the method comprising the following steps: Step 1: Select an operating point of the multiphase motor and identify the electrical parameters of the dq axis of the multiphase motor at this operating point, including the inductance matrix, resistance matrix, and permanent magnet flux linkage; Based on the dq axis electrical parameters, construct a mathematical model of the relationship between the reference value of the dq axis current at the next moment of the multiphase motor at this operating point and the actual dq axis current and actual dq axis voltage at the current moment, as the mathematical model of the multiphase motor in the dq coordinate system; Step 2: Simultaneously collect the phase current and phase voltage of each phase winding of the multiphase motor at multiple times using the sampling module, as well as the electrical angle and electrical angular velocity of the multiphase motor rotor relative to the initial position. Step 3: Based on the sampled phase currents and phase voltages of each phase winding of the multiphase motor, obtain the actual dq-axis current and actual dq-axis voltage at the corresponding time moments using a coordinate transformation algorithm; use the actual dq-axis current and actual dq-axis voltage at each time moment as observation state data to construct an observation state time series matrix. Z Based on the actual dq-axis current and voltage at each moment, the reference dq-axis current for the next moment is obtained according to the mathematical model of the multiphase motor in the dq coordinate system. The difference between the actual dq-axis current and the reference dq-axis current at each moment is taken as the observation error for that moment, and an observation error time series matrix is ​​constructed. ; Step 4: Construct a neural network module, where the activation function is the ReLU function; based on the backpropagation principle, the observation state time series matrix is ​​used. Z and observation error time series matrix Initialize the weight matrix of the training neural network; Step 5: Preset the optimal weight matrix of the neural network module, construct the neural network weight adjustment law module based on Lyapunov's theorem and the observer's error evaluation model, and update the weight matrix according to the neural network weight adjustment law module; Step 6: Select the set of control variables and predict the current trajectory; Step 7, Select the future k The optimization objective is to minimize the tracking error of the predicted current trajectory at time +1 relative to the current reference value. An optimization objective function is then constructed. The optimal control quantity is obtained by comparing the corresponding optimization objective function values. Finally, the optimal control quantity is sent to the multiphase motor driver.

[0013] Furthermore, step 4 includes the following method steps: First, select m Observation status at any moment z ( m As input to the neural network module, the matrix W i and The elements are randomly selected between (0,1); the prediction error of the neural network module is then calculated using the following formula: ; Define a neural network for input z ( m The correction error is: ; Based on the gradient descent learning method, the correction is made The elements are used to obtain the initial values ​​of the weight matrix. ; In the formula: For the first m The model error between the predicted and actual values ​​of the dq-axis current at the sampling time; For the first m Model estimation error at sampling time; For the first m The input vector of the neural network module at the sampling time; Preset values ​​for the weight matrix of the neural network module; This is the input matrix for the neural network module; These are the initial values ​​for the weight matrix of the neural network module; is the activation function for the neural network module; This is the error correction vector for the neural network module.

[0014] The advantages and positive effects of this invention are as follows: This invention solves the problem that training the initial weight matrix of a neural network observer requires a large, intensive dataset, leading to a heavy data acquisition workload. By employing an adaptive neural network weight adjustment module, this invention reduces the dependence of the data-driven observer on the initial values ​​of the neural network module's weight matrix. The adaptive neural network weight adjustment module autonomously updates the network weight matrix by referencing the time-series trajectory of the prediction model's error compensation values, thereby enhancing the observer's stability and observation accuracy at different operating points of a multiphase motor driver. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the framework of a multiphase motor predictive control system according to the present invention.

[0016] Figure 2 This is a schematic diagram of the structure and principle of a small data-driven observer according to the present invention.

[0017] Figure 3 This is a diagram showing the effect of applying a small data-driven observer when the model inductance is twice that of the actual inductance.

[0018] Figure 4 This is a diagram showing the effect of applying a small data-driven observer when the model flux linkage is twice that of the actual one.

[0019] In the picture: For the first i The timing sequence of the switching signals for each phase of a multiphase motor.

[0020] For the first k Predicted value of the dq axis current vector of the multiphase motor at sampling time +1.

[0021] For the first k +1 sampling time control reference value of dq axis current.

[0022] In the first k The optimal dq-axis voltage vector selected at the sampling time.

[0023] For the first k The model error estimated by the neural network at the sampling time.

[0024] For the first k Neural network estimation at sampling time d Shaft model error.

[0025] For the first k Neural network estimation at sampling time q Shaft model error.

[0026] For the first k The sampling value of the dq axis current vector of the multiphase motor at the sampling time.

[0027] i d ( k ) is the first k The sampling value of the d-axis current vector of the multiphase motor at the sampling time.

[0028] i q ( k ) is the first k The sampling value of the q-axis current vector of the multiphase motor at the sampling time.

[0029] i d This represents the d-axis current vector value of a multiphase motor.

[0030] i q This represents the q-axis current vector value of a multiphase motor.

[0031] In the first k The dq-axis voltage vector that is allowed to operate at the sampling time.

[0032] u d ( k ) for the first k The d-axis voltage vector that is allowed to operate at the sampling time.

[0033] u q ( k ) for the first k The q-axis voltage vector that is allowed to operate at the sampling time.

[0034] These are the correction values ​​for the weight matrix of the neural network module.

[0035] These are the sampled values ​​of the six-phase current from the current sensor.

[0036] These are the sampled values ​​of the six-phase voltage from the voltage sensor.

[0037] For multiphase motors in the first k The rotor position electrical angle at the sampling time.

[0038] ~ This represents the weights stored in the four neurons of the input layer of the neural network.

[0039] ~ This represents the weights stored in the ten neurons of the hidden layer of a neural network.

[0040] This is the activation function for the neural network module. Detailed Implementation

[0041] To further understand the invention's content, features, and effects, the following embodiments are provided, along with detailed descriptions in conjunction with the accompanying drawings: Please see Figures 1 to 4 A small data-driven observer is disclosed, which predicts the current error compensation value of a multiphase motor and outputs the current error compensation value. This current error compensation value is superimposed with a current reference value and input to the multiphase motor driver. The observer includes a neural network module and a neural network weight adjustment law module. The neural network module includes multiple neurons, and the neural network module adjusts the current data based on the current data... k The feedback values ​​of the current in each phase of the multiphase motor at any given time are calculated using the weights of each neuron to obtain the current value. kCurrent error compensation prediction value at time; neural network weight adjustment law module, which adjusts the value based on the observer's current... k The output value at time and the observed state are used to calculate the future. k The optimal weight adjustment increment of the neural network module at time +1 is based on the optimal weight adjustment increment for the current neural network module. k The weight matrix at each time step is updated to obtain the future... k The neural network weight matrix at time +1.

[0042] Preferably, the mathematical model of the observer can be: ; in: ; ; ; ; ; ; In the formula: The time derivative of the observer's state variables; For observer state variables; For the dq axis voltage vector of a multiphase motor; This refers to the dq-axis current vector of a multiphase motor. The dq-axis current vector estimated by the observer; For model estimation error; The d-axis current component estimated by the observer; The q-axis current component estimated by the observer; G The state matrix; F The input matrix; Output matrix for the observer; This is the inherent disturbance quantity; For the d-axis resistance of a multiphase motor; For the q-axis resistance of a multiphase motor; For the d-axis inductance of a multiphase motor; For the q-axis inductance of a multiphase motor; This refers to the angular velocity of the multiphase motor. The controller sampling period; For permanent magnet flux linkage in multiphase motors; For observer gain; It is a two-dimensional identity matrix; Observer Gain K The values ​​satisfy the matrix ( G - KC o The real part of the eigenvalues ​​of ) is less than zero.

[0043] Preferably, the neural network weight adjustment law module may include an observer error evaluation model, which can be used to evaluate the fitting error of the neural network module. The mathematical model of the observer error evaluation model may be: ; ; in: ; ; ; In the formula: ε This represents the fitting error of the neural network module; These are sampled values ​​of the observer's state variables; The error in estimating the observer's state variables; The time derivative of the observer state quantity estimation error; The observer estimation error for the dq-axis current vector of a multiphase motor; This is the input vector for the neural network module; is the activation function for the neural network module; This is the input matrix for the neural network module; This is the weight matrix of the neural network module; This represents the optimal value of the weight matrix for the neural network module; This represents the optimal increment of the weight matrix for the neural network module.

[0044] Preferably, the neural network weight adjustment module may further include a neural network weight adjustment module; the mathematical model of the neural network weight adjustment module may be: ; in: = P It is a positive definite matrix and satisfies , Q It is also a positive definite matrix; In the formula: For the first k The model error estimated by the neural network at the sampling time; For the first k Sampled values ​​of the dq-axis current vector of the multiphase motor at the sampling time; For the first k Predicted values ​​of the dq-axis current vector of the multiphase motor at the sampling time; For the first k The weight matrix correction value of the neural network module at sampling time +1; For the first k The weight matrix of the neural network module at the sampling time; This is the input matrix for the neural network module; For the first k The input vector of the neural network module at the sampling time; To correct the inertia coefficient of the neural network weight matrix; To correct the dynamic coefficients of the neural network weight matrix.

[0045] The present invention also provides a multiphase motor predictive control system, which includes a sampling module and the aforementioned small data-driven observer. The sampling module is used to sample the phase current and phase voltage of each phase winding of the multiphase motor at each time moment, and to sample the electrical angle and electrical angular velocity of the multiphase motor rotor relative to the initial position at each time moment. 。

[0046] Preferably, the system may further include a predictive control module; the predictive control module can input the current error compensation value from the observer and the sampled actual current value, and output the optimal control signal to the multiphase motor driver; the predictive control module may include a predictive current trajectory unit and a control signal optimization unit, the predictive current trajectory unit can generate the current prediction value for the next moment from the sampled signal at the current moment based on the mathematical model of the multiphase motor, and further form the predicted current trajectory; the control signal optimization unit can be used to select... k The optimal output at time +1 makes the corresponding value in the predicted current trajectory... k The error between the predicted current value at time +1 and the current reference value at that time is the smallest.

[0047] Preferably, the mathematical model of the control signal optimization unit can be: ; in: ; ; ; In the formula: In the first k The role of sampling time i voltage vectors along the dq axes The corresponding cost of control error; In the first k The role of sampling time i voltage vectors The corresponding number k Predicted value of dq-axis current vector at sampling time +1; For the first k +1 sampling time control reference value of dq axis current; In the first k The sampling time allows the action of the first i One dq-axis voltage vector; In the first k The optimal dq-axis voltage vector selected at the sampling time; For the first k The angular velocity of the multiphase motor at the sampling time; For the first k The model error estimated by the neural network at the sampling time; In the first k The set of dq-axis voltage vectors that are allowed to operate at the sampling time.

[0048] The present invention also provides a multiphase motor predictive control method utilizing the above-described multiphase motor predictive control system, the method comprising the following steps: Step 1: Select an operating point of the multiphase motor and identify the electrical parameters of the dq axis of the multiphase motor at this operating point, including the inductance matrix, resistance matrix, and permanent magnet flux linkage; Based on the dq axis electrical parameters, construct a mathematical model of the relationship between the reference value of the dq axis current at the next moment of the multiphase motor at this operating point and the actual dq axis current and actual dq axis voltage at the current moment, as the mathematical model of the multiphase motor in the dq coordinate system; Step 2: Simultaneously collect the phase current and phase voltage of each phase winding of the multiphase motor at multiple times using the sampling module, as well as the electrical angle and electrical angular velocity of the multiphase motor rotor relative to the initial position. Step 3: Based on the sampled phase currents and phase voltages of each phase winding of the multiphase motor, obtain the actual dq-axis current and actual dq-axis voltage at the corresponding time moments using a coordinate transformation algorithm; use the actual dq-axis current and actual dq-axis voltage at each time moment as observation state data to construct an observation state time series matrix. Z Based on the actual dq-axis current and voltage at each moment, the reference dq-axis current for the next moment is obtained according to the mathematical model of the multiphase motor in the dq coordinate system. The difference between the actual dq-axis current and the reference dq-axis current at each moment is taken as the observation error for that moment, and an observation error time series matrix is ​​constructed. ; Step 4: Construct a neural network module, where the activation function is the ReLU function; based on the backpropagation principle, the observation state time series matrix is ​​used. Z and observation error time series matrix Initialize the weight matrix of the training neural network; Step 5: Preset the optimal weight matrix of the neural network module, construct the neural network weight adjustment law module based on Lyapunov's theorem and the observer's error evaluation model, and update the weight matrix according to the neural network weight adjustment law module; Step 6: Select the set of control variables and predict the current trajectory; Step 7, Select the future k The optimization objective is to minimize the tracking error of the predicted current trajectory at time +1 relative to the current reference value. An optimization objective function is then constructed. The optimal control quantity is obtained by comparing the corresponding optimization objective function values. Finally, the optimal control quantity is sent to the multiphase motor driver.

[0049] Preferably, step 4 may include the following method steps: First, select k Observation status at any moment z ( m As input to the neural network module, the matrixW i and The elements are randomly selected between (0,1); the prediction error of the neural network module is then calculated using the following formula: ; Define a neural network for input z ( m The correction error is: ; Based on the gradient descent learning method, the correction is made The elements are used to obtain the initial values ​​of the weight matrix. ; In the formula: For the first m The model error between the predicted and actual values ​​of the dq-axis current at the sampling time; For the first m Model estimation error at sampling time; For the first m The input vector of the neural network module at the sampling time; Preset values ​​for the weight matrix of the neural network module; This is the input matrix for the neural network module; These are the initial values ​​for the weight matrix of the neural network module; is the activation function for the neural network module; This is the error correction vector for the neural network module.

[0050] The workflow and working principle of the present invention will be further described below with reference to a preferred embodiment: Figure 1 This is a schematic diagram of the framework of a multiphase motor predictive control system according to the present invention, including: a small data-driven observer, a predictive control module, a sampling module, and a multiphase motor driver; the multiphase motor driver is used to drive a multiphase permanent magnet synchronous motor. The small data-driven observer includes a neural network module and a neural network weight adjustment law module; the multiphase motor driver includes a two-level multiphase inverter circuit; the sampling module includes current and voltage sensors and a signal sampling processor; the predictive control module includes a predictive current trajectory unit and a control signal optimization unit.

[0051] The present invention provides a multiphase motor predictive control method, employing the aforementioned multiphase motor predictive control system, which includes the following steps: Step I: In the current kAt any given time, the current in each phase winding of the multiphase permanent magnet synchronous motor is obtained through current and voltage sensors and signal sampling. i s ( k )=[ i a ( k ) i b ( k ) i c ( k ) i u ( k ) i v ( k ) i w ( k )] T and the voltage at both ends u s ( k )=[ u a ( k ) u b ( k ) u c ( k ) u u ( k ) u v ( k ) u w ( k )] T .

[0052] In the formula: i s ( k ) is the first k The sampling values ​​of the current sensor for the six-phase current at the sampling time; i a ( k ) is the first k The current sensor samples the current of phase A at the sampling time; i b ( k ) is the first k The sampling value of the current sensor for phase B current at the sampling time; i c ( k ) is the firstk The sampling value of the current sensor for the C-phase current at the sampling time; i u ( k ) is the first k The sampling value of the current sensor for the U-phase current at the sampling time; i v ( k ) is the first k The sampling value of the current sensor for the V-phase current at the sampling time; i w ( k ) is the first k The sampling value of the current sensor for the W-phase current at the sampling time; u s ( k ) is the first k The voltage sensor samples the six-phase voltage at the sampling time. u a ( k ) is the first k The sampling value of the current sensor for the voltage of phase A at the sampling time; u b ( k ) is the first k The sampling value of the current sensor for the B-phase voltage at the sampling time; u c ( k ) is the first k The current sensor samples the C-phase voltage at the sampling time. u u ( k ) is the first k The sampling value of the current sensor for the U-phase voltage at the sampling time; u v ( k ) is the first k The sampling value of the current sensor for the V-phase voltage at the sampling time; u w ( k ) is the first k The sampling value of the current sensor for the voltage of phase W at the sampling time; The rotor position electrical angle is obtained through an encoder. θ e and position electric angular velocity ω eThe coordinate transformation is obtained through coordinate transformation in the coordinate transformation processor. k dq axis current at time i dq ( k ) and dq axis voltage u dq ( k Its transformation matrix is: (1) Step II: Select a specific operating point for the multiphase motor driver and simultaneously collect current and voltage feedback values ​​at multiple time points. i dq =[ i d i q ] T and u dq =[ u d u q ] T And construct information about the observation state z =[ i dq ; u dq The time series matrix z Identify the electrical parameters of the dq model of the multiphase permanent magnet synchronous motor at this operating point, including the inductance matrix. L =diag{ L d , L q}, resistance matrix R =diag{ R d , R q} and permanent magnet flux Ψ PM A mathematical model of the multiphase permanent magnet synchronous motor in the dq coordinate system at this operating point is constructed based on the electrical parameters: , (2a); In the formula, the state matrix A Input matrix B Inherent disturbance d It can be represented as , , (2b); In the formula: In the first k Voltage vector applied at sampling time The corresponding number k Predicted dq-axis current vector at sampling time +1; For the first k The sampled values ​​of the dq-axis current vector of the multiphase motor at the sampling time; For the first k The sampled values ​​of the dq-axis current vector of the multiphase motor at the sampling time; Step III: Utilize the information about the observation status from Step II z The time sequence matrix is ​​used to construct the neural network weight matrix. The initial value of the training set is determined. m and m Observation status at time +1 z ( m )=[ i dq ( m ); u dq ( m )], z ( m +1)=[ i dq ( m +1); u dq ( m +1)]; In the formula: i dq ( m ) is the first m The sampled value of the dq-axis current vector at the sampling time; i dq ( m +1) is the first m The sampled value of the dq-axis current vector at sampling time +1; u dq ( m ) is the first m The sampled value of the dq-axis voltage vector at the sampling time; u dq ( m +1) is the first m The sampled value of the dq-axis voltage vector at sampling time +1; According to equation (2a), we can obtain m Current state predicted by the reference mathematical model at time +1 ;but m Model prediction error at time 1 It can be defined according to formula (3): , (3); Therefore, we can establish information about model prediction errors. Time series matrix .

[0053] Step IV: Construct as follows Figure 2 The neural network module uses the ReLU activation function; based on the backpropagation principle, a time sequence matrix is ​​employed. z and Initial values ​​for the weight matrix of the training neural network: First, select... m Observation status at any moment z ( m As input to the neural network module, the matrix W i and The elements are randomly selected between (0,1); then the model prediction error observed by the neural network module is calculated using equation (4). ( m ): (4); Define a neural network for input z ( m The correction error is Based on gradient descent learning, correction The elements are used to obtain the initial values ​​of the weight matrix. .

[0054] Step V: Construct a small data-driven observer model: (5); In the formula, the observer state quantity Observer output matrix , The prediction error is caused by model bias. The observer output and the observer gain are the values ​​of the two variables. K The values ​​satisfy the matrix ( G - KC o The real part of the eigenvalues ​​of the given network is less than zero; the optimal weight matrix of the preset neural network module is... Then the optimal value of the observation error is: (6); Considering prediction error Weight matrix and activation function output The boundedness of the observer allows us to construct an error evaluation model: (7); in, , , , ε This represents the fitting error of the neural network module.

[0055] Step VI: Based on Lyapunov's theorem, design the neural network weight adjustment law: First, select the Lyapunov function in equation (8): (8); Among them, positive definite matrix P satisfy , Q It is also a positive definite matrix; considering the boundedness of matrix values, according to The neural network weight adjustment law in discrete time can be obtained in equation (9).

[0056] (9); The weight matrix is ​​updated based on the neural network weight adjustment rate.

[0057] Step VII: Select the set of control variables and predict the current trajectory: First, based on the virtual vector principle of the multiphase motor driver, determine the set of control variables for the predictive control module. ,in U The elements of the vector are all composed of 0s and 1s; then, based on the principle of coordinate transformation, the current vector is obtained. k The set of control variables of the dq model at time 1 Finally, predict the future based on formula (10). k At time +1, each Corresponding current value .

[0058] (10); Step VIII: Select the future k The current trajectory at time +1 For current reference value Minimizing the tracking error is the optimization objective, and the optimization objective function in equation (11) is constructed accordingly. (11); Combining current trajectory prediction and control quantity set, the optimization problem in equation (12) is obtained. By comparing various The corresponding optimization objective function value Solving for the optimal control quantity The optimal control signal is then transmitted to the multiphase motor driver to complete the process. k The primary closed-loop control function of a multiphase motor predictive control system within a given time period.

[0059] .

[0060] The aforementioned neural network module, observer error evaluation model, sampling module, multiphase motor driver and other components or functional modules can all be components or functional modules in the prior art, or components or functional modules in the prior art can be constructed using conventional technical means.

[0061] The embodiments described above are only used to illustrate the technical ideas and features of the present invention. Their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The patent scope of the present invention should not be limited by these embodiments. That is, any equivalent changes or modifications made in accordance with the spirit disclosed in the present invention still fall within the patent scope of the present invention.

Claims

1. A small data-driven observer, characterized in that, This observer is used to predict the current error compensation value of a multiphase motor. It outputs the current error compensation value, which is then superimposed with a current reference value and input to the multiphase motor driver. The observer includes a neural network module and a neural network weight adjustment law module. The neural network module contains multiple neurons, and the neural network module adjusts the current... k The feedback values ​​of the current in each phase of the multiphase motor at any given time are calculated using the weights of each neuron to obtain the current value. k Current error compensation prediction value at time; neural network weight adjustment law module, which adjusts the value based on the observer's current... k The output value at time and the observed state are used to calculate the future. k The optimal weight adjustment increment of the neural network module at time +1 is based on the optimal weight adjustment increment for the current neural network module. k The weight matrix at each time step is updated to obtain the future... k The neural network weight matrix at time +1; The mathematical model for this observer is: ； in: ； ； ； ； ； ； In the formula: The time derivative of the observer's state variables; For observer state variables; For the dq axis voltage vector of a multiphase motor; This refers to the dq-axis current vector of a multiphase motor. The dq-axis current vector estimated by the observer; For model estimation error; The d-axis current component estimated by the observer; The q-axis current component estimated by the observer; G The state matrix; F The input matrix; Output matrix for the observer; This is the inherent disturbance quantity; For the d-axis resistance of a multiphase motor; For the q-axis resistance of a multiphase motor; For the d-axis inductance of a multiphase motor; For the q-axis inductance of a multiphase motor; This refers to the angular velocity of the multiphase motor. The controller sampling period; For permanent magnet flux linkage in multiphase motors; For observer gain; It is a two-dimensional identity matrix; Observer Gain K The values ​​satisfy the matrix ( G - KC o The real part of the eigenvalues ​​of ) is less than zero; The neural network weight adjustment module also includes a neural network weight adjustment module; the mathematical model of the neural network weight adjustment module is: ； in: = P It is a positive definite matrix and satisfies , Q It is also a positive definite matrix; In the formula: For the first k The model error estimated by the neural network at the sampling time; For the first k Sampled values ​​of the dq-axis current vector of the multiphase motor at the sampling time; For the first k Predicted values ​​of the dq-axis current vector of the multiphase motor at the sampling time; For the first k The weight matrix correction value of the neural network module at sampling time +1; For the first k The weight matrix of the neural network module at the sampling time; This is the input matrix for the neural network module; For the first k The input vector of the neural network module at the sampling time; To correct the inertia coefficient of the neural network weight matrix; To correct the dynamic coefficients of the neural network weight matrix.

2. The small data-driven observer according to claim 1, characterized in that, The neural network weight adjustment law module includes an observer error evaluation model, which is used to evaluate the fitting error of the neural network module. The mathematical model of the observer error evaluation model is as follows: ； ； in: ； ； ； In the formula: ε This represents the fitting error of the neural network module; These are sampled values ​​of the observer's state variables; The error in estimating the observer's state variables; The time derivative of the observer state quantity estimation error; The observer estimation error for the dq-axis current vector of a multiphase motor; This is the input vector for the neural network module; is the activation function for the neural network module; This is the input matrix for the neural network module; This is the weight matrix of the neural network module; This represents the optimal value of the weight matrix for the neural network module; This represents the optimal increment of the weight matrix for the neural network module.

3. A multiphase motor predictive control system, characterized in that, The system includes a sampling module and a small data-driven observer as described in any one of claims 1 to 2. The sampling module is used to sample the phase current and phase voltage of each phase winding of the multiphase motor at each moment, and to sample the electrical angle and electrical angular velocity of the multiphase motor rotor relative to its initial position at each moment. 。 4. The multiphase motor predictive control system according to claim 3, characterized in that, The system also includes a predictive control module; the predictive control module takes into input the current error compensation value from the observer and the sampled actual current value, and outputs the optimal control signal to the multiphase motor driver; the predictive control module includes a predictive current trajectory unit and a control signal optimization unit. The predictive current trajectory unit, based on the mathematical model of the multiphase motor, generates the predicted current value for the next moment from the sampled signal at the current moment, and further forms the predicted current trajectory; the control signal optimization unit is used to select... k The optimal output at time +1 makes the corresponding value in the predicted current trajectory... k The error between the predicted current value at time +1 and the current reference value at that time is the smallest.

5. The multiphase motor predictive control system according to claim 4, characterized in that, The mathematical model of the control signal optimization unit is as follows: ； in: ； ； ； In the formula: In the first k The role of sampling time i voltage vectors along the dq axes The corresponding cost of control error; In the first k The role of sampling time i voltage vectors The corresponding number k Predicted value of dq-axis current vector at sampling time +1; For the first k +1 sampling time control reference value of dq axis current; In the first k The sampling time allows the action of the first i One dq-axis voltage vector; In the first k The optimal dq-axis voltage vector selected at the sampling time; For the first k The angular velocity of the multiphase motor at the sampling time; For the first k The model error estimated by the neural network at the sampling time; In the first k The set of dq-axis voltage vectors that are allowed to operate at the sampling time.

6. A multiphase motor predictive control method utilizing the multiphase motor predictive control system according to any one of claims 3 to 5, characterized in that, The method includes the following steps: Step 1: Select an operating point of the multiphase motor and identify the electrical parameters of the dq axis of the multiphase motor at this operating point, including the inductance matrix, resistance matrix, and permanent magnet flux linkage; Based on the dq axis electrical parameters, construct a mathematical model of the relationship between the reference value of the dq axis current at the next moment of the multiphase motor at this operating point and the actual dq axis current and actual dq axis voltage at the current moment, as the mathematical model of the multiphase motor in the dq coordinate system; Step 2: Simultaneously collect the phase current and phase voltage of each phase winding of the multiphase motor at multiple times using the sampling module, as well as the electrical angle and electrical angular velocity of the multiphase motor rotor relative to the initial position. Step 3: Based on the sampled phase currents and phase voltages of each phase winding of the multiphase motor, obtain the actual dq-axis current and actual dq-axis voltage at the corresponding time moments using a coordinate transformation algorithm; use the actual dq-axis current and actual dq-axis voltage at each time moment as observation state data to construct an observation state time series matrix. Z Based on the actual dq-axis current and voltage at each moment, the reference dq-axis current for the next moment is obtained according to the mathematical model of the multiphase motor in the dq coordinate system. The difference between the actual dq-axis current and the reference dq-axis current at each moment is taken as the observation error for that moment, and an observation error time series matrix is ​​constructed. ; Step 4: Construct a neural network module, where the activation function is the ReLU function; based on the backpropagation principle, the observation state time series matrix is ​​used. Z and observation error time series matrix Initialize the weight matrix of the training neural network; Step 5: Preset the optimal weight matrix of the neural network module, construct the neural network weight adjustment law module based on Lyapunov's theorem and the observer's error evaluation model, and update the weight matrix according to the neural network weight adjustment law module; Step 6: Select the set of control variables and predict the current trajectory; Step 7, Select the future k The optimization objective is to minimize the tracking error of the predicted current trajectory at time +1 relative to the current reference value. An optimization objective function is then constructed. The optimal control quantity is obtained by comparing the corresponding optimization objective function values. Finally, the optimal control quantity is sent to the multiphase motor driver.

7. The multiphase motor predictive control method according to claim 6, characterized in that, Step 4 includes the following steps: First, select... m Observation status at any moment z ( m As input to the neural network module, the matrix W i and The elements are randomly selected between (0,1); the prediction error of the neural network module is then calculated using the following formula: ； Define a neural network for input z ( m The correction error is: ； Based on the gradient descent learning method, the correction is made The elements are used to obtain the initial values ​​of the weight matrix. ; In the formula: For the first m The model error between the predicted and actual values ​​of the dq-axis current at the sampling time; For the first m Model estimation error at sampling time; For the first m The input vector of the neural network module at the sampling time; Preset values ​​for the weight matrix of the neural network module; This is the input matrix for the neural network module; These are the initial values ​​for the weight matrix of the neural network module; is the activation function for the neural network module; This is the error correction vector for the neural network module.

Images