Motor dead-beat prediction control method and system based on real-time neural network identification

Abstract

The method is suitable for PMSM speed regulation control in the field of artillery follow-up systems. The invention provides a motor dead-beat prediction control method and system based on a real-time neural network identifier aiming at the requirements of an artillery follow-up system on quick response, high steady-state precision and good anti-interference capability of a permanent magnet synchronous motor. Dead-beat prediction control and neural network control are combined, a dead-beat prediction controller can ensure that a motor tracks an expected target with the maximum output capacity, and an improved neural network identifier can compensate lumped disturbance of a rotating speed ring of the permanent magnet synchronous motor in real time through an adaptive gradient descent algorithm. The disturbance resistance performance of the dead-beat prediction controller is improved, the steady-state control precision of the motor and the rotating speed is improved, and meanwhile the complex global parameter identification process is avoided. According to the motor dead-beat prediction control method based on the real-time neural network identifier, the rotating speed tracking performance and the robust performance of the permanent magnet synchronous motor can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of artillery servo system control, and more particularly to motor speed control. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs) are widely used in weaponry due to their high power density, high efficiency, and excellent dynamic performance. In high-performance artillery servo control systems, PMSM speed control systems typically require fast response, high steady-state accuracy, and good disturbance rejection capabilities. Deadbeat predictive control (DMC) is a high-performance control algorithm widely used in motor speed control. Its main characteristic is that it directly calculates the control input through a mathematical model, eliminating the need for complex gain adjustments and enabling rapid convergence to the desired value within one control cycle, ensuring the system achieves accurate target tracking with maximum output capability. However, DMC relies on an accurate motor mathematical model, and during operation, the PMSM is subject to load disturbances and torque pulsations, making it difficult to guarantee the system's steady-state control accuracy.

[0003] To mitigate the impact of disturbances and model uncertainties on deadbeat predictive control, numerous methods have been proposed. One approach is to employ integral deadbeat predictive control to eliminate disturbances; however, this method can cause oscillations in the system response. Another more common method involves using observers. To improve system robustness, different types of observers have been proposed, such as generalized proportional-integral observers, extended state observers, and Kalman filter observers. Neural networks have gained widespread attention for their powerful nonlinear approximation and adaptive learning capabilities in solving complex system modeling and control problems. In the speed control of permanent magnet synchronous motors, neural networks can effectively identify the dynamic characteristics of the motor.

[0004] However, traditional neural network identifiers require large amounts of data and converge slowly. The speed control system of PMSM needs to have high computational efficiency and real-time performance, which places higher demands on the control algorithm. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a motor deadbeat predictive control method and system based on a real-time neural network identifier that combines neural network technology and deadbeat predictive control method to compensate for lumped disturbances in the speed loop of permanent magnet synchronous motor in real time, thereby improving the dynamic performance and control accuracy of the artillery servo system.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a motor deadbeat predictive control method based on a real-time neural network identifier, comprising the following steps: Step 1: Taking the speed loop of the permanent magnet synchronous motor control system as the controlled object, based on the lumped disturbance of the controlled object, establish a hyperlocal model of the permanent magnet synchronous motor with the desired set current of the q axis in the two-phase rotating dq coordinate system as the control input and the angular velocity of the motor rotor as the output. Step 2: Discretize the hyperlocal model using the first-order forward Euler method to construct a discrete prediction model of the motor speed, and obtain the deadbeat predictive control law of the deadbeat predictive controller of the speed loop based on the deadbeat principle. Step 3: Establish a real-time BP neural network identifier with the motor speed at the current sampling time as input and the lumped disturbance estimate of the speed loop at the next time as output. This is used to identify the lumped disturbance of the speed loop online and to use the obtained lumped disturbance estimate to perform feedforward compensation on the deadbeat predictive control law, generating the final motor speed control law, which is then applied to the motor control system to achieve fast, accurate, and robust control of the motor speed.

[0007] Furthermore, step one specifically includes: In the two-phase rotating dq coordinate system, the motion equation of the permanent magnet synchronous motor is expressed as: , In the formula, This represents the moment of inertia of the motor rotor. Indicates the angular velocity of the motor rotor; Indicates the number of pole pairs of the motor; Indicates permanent magnet flux linkage; and These represent the rotating dq-axis currents of the two phases, respectively. and These represent the two-phase rotating dq-axis inductances, respectively. Indicates load torque; Indicates torque pulsation; Indicates the coefficient of viscous friction; Then, with the angular velocity of the motor rotor As output, the desired set current for the q-axis. As the control input, the hyperlocal model of the permanent magnet synchronous motor is obtained as follows: , In the formula, The lumped disturbance representing the velocity loop, and Represented as: .

[0008] Furthermore, step two specifically includes: The hyperlocal model is discretized using the first-order forward Euler method, resulting in the discrete prediction model for motor speed: , In the formula, The sampling period of the velocity loop. and They represent the first k The and the first k +1 motor speed value at sampling time, This is the desired setpoint for the motor speed. The desired set current for the q-axis. Indicates permanent magnet flux linkage. Indicates the number of pole pairs of the motor. This represents the moment of inertia of the motor rotor. The lumped disturbance representing the velocity loop; Based on the principle of no beat time, the discrete prediction model is... use Substituting the desired setpoint for the motor speed, the control law of the deadbeat predictive controller for the speed loop is obtained as follows: .

[0009] Furthermore, in step three, the established real-time BP neural network identifier is as follows: , In the formula, for k The estimated value of the motor speed at time +1. Indicates the first k Motor speed at each sampling time, The desired set current for the q-axis. Indicates permanent magnet flux linkage. Indicates the number of pole pairs of the motor. This represents the moment of inertia of the motor rotor. The lumped disturbance estimate for the rotational speed loop is obtained by a real-time BP neural network with a single input, single output, and a single hidden layer.

[0010] Furthermore, the real-time BP neural network includes an input layer, a hidden layer, and an output layer, with the input layer using... k Motor speed at any given time The input layer to the hidden layer uses weight coefficients of... The layer has a 1×N weight matrix and a bipolar sigmoid activation function, with multiple nonlinear nodes N in the hidden layer; the weight coefficients from the hidden layer to the output layer are... The N×1 weight vector and linear activation function, the output layer node L, and the output estimate. k Lumped disturbance of the velocity loop at any moment ; in, , , These represent the number of nodes in the hidden layer, input layer, and output layer, respectively. At the same time, the number of neurons in both the input layer and the output layer is 1; The number of neurons in the hidden layer is determined during training. The activation function between the hidden layer and the output layer is a linear function, and the activation function between the input layer and the hidden layer is a bipolar sigmoid function. , And the cost function is: , The lumped disturbance of the speed loop is identified online, where, and They are respectively k The estimated and actual values ​​of the motor speed at time +1 express k The rotational speed identification error at time +1, in practical applications The desired setpoint for the motor speed is used instead, and the cost function is gradually reduced during the identification process to ensure the estimation... k Lumped disturbance of the velocity loop at any moment Always approaching k Moment-time velocity lumped disturbance.

[0011] Furthermore, it also includes the real-time online identification of lumped disturbances in the speed loop using an adaptive gradient descent algorithm, specifically including: , , In the formula, and They represent k The changes in weight coefficients from the input layer to the hidden layer and from the hidden layer to the output layer at any given time. For hidden layers in k Output at any moment; This is the input to the hidden layer at time k; For learning rate, ; The inertia coefficient, ; This is the rate adjustment coefficient. , , , These represent the number of nodes in the hidden layer, input layer, and output layer of the real-time BP neural network identifier, respectively. Used when the cost function value is too large. The weight learning rate will be proportional to the deviation of the option value from the previous week, and the learning rate will automatically increase to shorten the dynamic response time. When the cost function approaches zero, The learning rate tends to This slows down the convergence speed compared to the original gradient descent algorithm, meaning the learning rate is automatically reduced to suppress jitter and ensure asymptotic stability.

[0012] Furthermore, the lumped disturbance estimate of the speed loop is obtained using the aforementioned real-time BP neural network identifier. By performing feedforward correction on the deadbeat predictive speed controller, the control law for the motor speed is obtained as follows: , In the formula, The desired set current for the q-axis. This represents the moment of inertia of the motor rotor. Indicates the number of pole pairs of the motor. Indicates permanent magnet flux linkage. The sampling period of the velocity loop. This is the desired setpoint for the motor speed. This indicates the motor speed value. This is the estimated value of the lumped disturbance in the rotational speed loop.

[0013] The present invention also provides a motor deadbeat predictive control system based on a real-time neural network identifier, comprising a deadbeat predictive controller for executing the motor deadbeat predictive control method based on a real-time neural network identifier, a permanent magnet synchronous motor, a three-phase inverter, a current loop controller, Clark and Park coordinate transformation, and a space vector pulse width modulation module. The three-phase current output by the permanent magnet synchronous motor is transformed into the current in the α-β stationary coordinate system in the Clark coordinate transformation module, and then transformed into the feedback current value in the dq rotating coordinate system by the Park coordinate transformation module. After setting the desired motor operating speed in the deadbeat predictive controller, the real-time BP neural network in the deadbeat predictive controller collects the actual motor speed and current at the current moment, and estimates the estimated motor speed at the next moment based on the actual motor speed and current. Then, the deadbeat predictive controller uses the estimated motor speed and the desired motor operating speed to calculate the cost function, and minimizes the cost function through the adaptive gradient descent algorithm to obtain the lumped disturbance estimate of the current loop. Then, the lumped disturbance estimate is used to feedforward compensate the control law of the deadbeat predictive controller to generate the motor speed compensation control law. The motor speed control law and the lumped disturbance estimate are input to the current loop controller. The resulting control voltage is transformed to the voltage in the α-β stationary coordinate system by the inverse Park coordinate transformation module, and then sent to the space vector pulse width modulation module to generate a PWM control switch signal. The PWM control switch signal drives the inverter to output three-phase voltage and apply it to the permanent magnet synchronous motor again, realizing that the motor speed compensation control law acts on the motor, thereby completing the power stage closed loop.

[0014] Furthermore, it also includes position and velocity detection sensors and external loads; The position and speed detection sensors extract the rotor position and speed of the permanent magnet synchronous motor from the shaft end. The rotor position is used for the real-time coordinate rotation of the Park coordinate transformation module and the inverse Park coordinate transformation module, while the speed is used to provide current for the speed outer loop. The torque of the external load acts directly on the shaft of the permanent magnet synchronous motor. The disturbance is quickly suppressed by the current loop through the motor current, so that the entire control system forms a high-performance drive link that includes power, signal and parameter identification triple closed loops.

[0015] Furthermore, the system is a control system in the artillery servo system, used to maintain sub-millisecond response, zero steady-state error, and steady-state speed fluctuation within ±1 rpm under conditions of sudden load changes, parameter drift, and torque pulsation.

[0016] The beneficial effects of this invention are as follows: By combining improved neural network technology and deadbeat predictive control method, this invention improves the dynamic performance and control accuracy of artillery servo systems. This method combines the advantages of deadbeat predictive controller and neural network identifier. The deadbeat predictive controller can ensure that the motor tracks the desired target with maximum output capability. The improved neural network identifier only needs the signal input at the current sampling time each time, and through the adaptive gradient descent algorithm, it can compensate for the lumped disturbance of the permanent magnet synchronous motor speed loop in real time, improve the disturbance resistance performance of the deadbeat predictive controller, improve the steady-state control accuracy of the motor speed, and at the same time avoid the complex global parameter identification process. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the real-time BP neural network structure of the hidden layer of the present invention; Figure 2 This is a predictive control block diagram of specific embodiment 1 of the present invention; Figure 3 This is a structural diagram of the predictive controller of the present invention; Figure 4 This is a flowchart of a specific example 1 of the present invention; Figure 5 This invention enables both traditional PI controllers and the algorithm of this invention to track the same step signal. A comparison of the step responses of ( ); Figure 6 The traditional PI controller of this invention and the algorithm of this invention are in contrast. Suddenly applied load torque ( Speed ​​response comparison chart; Figure 7 The conventional PI controller and the algorithm of this invention apply torque pulsation under no-load conditions. Comparison of steady-state speed fluctuations; Figure 8 The results are the fast Fourier transform analysis of the no-load steady-state speed of the traditional PI controller and the algorithm of this invention. Figure 9 This is a schematic diagram of the device structure of a specific embodiment 2 of the present invention. Detailed Implementation

[0018] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0019] This invention first takes the speed loop of the permanent magnet synchronous motor control system as the controlled object. Based on the lumped disturbance of the controlled object, a hyperlocal model of the permanent magnet synchronous motor is established with the desired set current of the q axis in the two-phase rotating dq coordinate system as the control input and the angular velocity of the motor rotor as the output. Next, the first-order forward Euler method is used to discretize the hyperlocal model, and a discrete prediction model of the motor speed is constructed. Based on the deadbeat principle, the deadbeat predictive control law of the deadbeat predictive controller of the speed loop is obtained. Finally, a real-time BP neural network identifier is established, with the motor speed at the current sampling time as input and the lumped disturbance estimate of the speed loop at the next time as output. This is used to identify the lumped disturbance of the speed loop online, and to use the obtained lumped disturbance estimate to perform feedforward compensation on the deadbeat predictive control law, generating the final motor speed control law, which is then applied to the motor control system to achieve fast, accurate, and robust control of the motor speed.

[0020] To achieve the above objectives, the present invention provides the following specific embodiments: Example 1: A method for predictive control of a motor without deadbeat based on a real-time neural network identifier, comprising the following steps: S01. Taking the speed loop of the permanent magnet synchronous motor control system as the controlled object, based on the lumped disturbance of the controlled object, the motion equation of the permanent magnet synchronous motor in the two-phase rotating dq coordinate system is expressed as: , In the formula, This represents the moment of inertia of the motor rotor. Indicates the angular velocity of the motor rotor; Indicates the number of pole pairs of the motor; Indicates permanent magnet flux linkage; and These represent the rotating dq-axis currents of the two phases, respectively. and These represent the two-phase rotating dq-axis inductances, respectively. Indicates load torque; Indicates torque pulsation; This represents the coefficient of viscous friction.

[0021] S02, based on the angular velocity of the motor rotor As output, the desired set current for the q-axis. As the control input, the hyperlocal model of the permanent magnet synchronous motor is obtained as follows: , In the formula, The lumped disturbance representing the velocity loop, and Represented as: .

[0022] S03. The hyperlocal model is discretized using the first-order forward Euler method, resulting in the discrete prediction model for the motor speed: , In the formula, The sampling period of the velocity loop. and They represent the first k The and the first k +1 motor speed value at sampling time, This is the desired setpoint for the motor speed. The desired set current for the q-axis. Indicates permanent magnet flux linkage. Indicates the number of pole pairs of the motor. This represents the moment of inertia of the motor rotor. This represents the lumped disturbance of the velocity loop.

[0023] S04. Based on the principle of no beat time, the discrete prediction model in... use Substituting the desired setpoint for the motor speed, the control law of the deadbeat predictive controller for the speed loop is obtained as follows: .

[0024] S05. Establish a real-time BP neural network identifier that takes the motor speed at the current sampling time as input and outputs the lumped disturbance estimate of the speed loop at the next time step, for online identification of lumped disturbances in the speed loop. The established real-time BP neural network identifier is as follows: , In the formula, for k The estimated value of the motor speed at time +1. Indicates the first k Motor speed at each sampling time, The desired set current for the q-axis. Indicates permanent magnet flux linkage. Indicates the number of pole pairs of the motor. This represents the moment of inertia of the motor rotor. The lumped disturbance estimate for the rotational speed loop is obtained from a real-time BP neural network with a single input and single output and a single hidden layer. A real-time backpropagation (BP) neural network specifically includes an input layer, hidden layers, and an output layer. The input layer uses... k Motor speed at any given time The input layer to the hidden layer uses weight coefficients of... The layer has a 1×N weight matrix and a bipolar sigmoid activation function, with multiple nonlinear nodes N in the hidden layer; the weight coefficients from the hidden layer to the output layer are... The N×1 weight vector and linear activation function, the output layer node L, and the output estimate. k Lumped disturbance of the velocity loop at any moment ; in, , , These represent the number of nodes in the hidden layer, input layer, and output layer, respectively. At the same time, the number of neurons in both the input layer and the output layer is 1; The number of neurons in the hidden layer is determined during training. The activation function between the hidden layer and the output layer is a linear function, and the activation function between the input layer and the hidden layer is a bipolar sigmoid function. , And the cost function is: , The lumped disturbance of the speed loop is identified online, where, and They are respectively k The estimated and actual values ​​of the motor speed at time +1 express k The rotational speed identification error at time +1, in practical applications The desired setpoint for the motor speed is used instead, and the cost function is gradually reduced during the identification process to ensure the estimation... k Lumped disturbance of the velocity loop at any moment Always approaching k Moment-time velocity lumped disturbance.

[0025] S06. An adaptive gradient descent algorithm is used to identify lumped disturbances in the rotational speed loop in real time, specifically including: , , In the formula, and They represent k The changes in weight coefficients from the input layer to the hidden layer and from the hidden layer to the output layer at any given time. For hidden layers in k Output at any moment; This is the input to the hidden layer at time k; For learning rate, ; The inertia coefficient, ; This is the rate adjustment coefficient. , , , These represent the number of nodes in the hidden layer, input layer, and output layer of the real-time BP neural network identifier, respectively. Used when the cost function value is too large. The weight learning rate will be proportional to the deviation of the option value from the previous week, and the learning rate will automatically increase to shorten the dynamic response time. When the cost function approaches zero, The learning rate tends to This slows down the convergence speed compared to the original gradient descent algorithm, meaning the learning rate is automatically reduced to suppress jitter and ensure asymptotic stability.

[0026] S07. Obtain the lumped disturbance estimate of the speed loop using a real-time BP neural network identifier. By performing feedforward correction on the deadbeat predictive speed controller, the control law for the motor speed is obtained as follows: , In the formula, The desired set current for the q-axis. This represents the moment of inertia of the motor rotor. Indicates the number of pole pairs of the motor. Indicates permanent magnet flux linkage. The sampling period of the velocity loop. This is the desired setpoint for the motor speed. This indicates the motor speed value. This is the estimated value of the lumped disturbance in the rotational speed loop.

[0027] Example 2: The present invention also provides a motor deadbeat predictive control system based on a real-time neural network identifier as in Example 1, including a deadbeat predictive controller that executes a motor deadbeat predictive control method based on a real-time neural network identifier, a permanent magnet synchronous motor, a three-phase inverter, a current loop controller, Clark and Park coordinate transformation, and a space vector pulse width modulation module. The three-phase current output by the permanent magnet synchronous motor is transformed into the current in the α-β stationary coordinate system in the Clark coordinate transformation module, and then transformed into the feedback current value in the dq rotating coordinate system by the Park coordinate transformation module. After setting the desired motor operating speed in the deadbeat predictive controller, the real-time BP neural network in the deadbeat predictive controller collects the actual motor speed and current at the current moment, and estimates the estimated motor speed at the next moment based on the actual motor speed and current. Then, the deadbeat predictive controller uses the estimated motor speed and the desired motor operating speed to calculate the cost function, and minimizes the cost function through the adaptive gradient descent algorithm to obtain the lumped disturbance estimate of the current loop. Then, the lumped disturbance estimate is used to feedforward compensate the control law of the deadbeat predictive controller to generate the motor speed compensation control law. The motor speed control law and the lumped disturbance estimate are input to the current loop controller. The resulting control voltage is transformed to the voltage in the α-β stationary coordinate system by the inverse Park coordinate transformation module, and then sent to the space vector pulse width modulation module to generate a PWM control switch signal. The PWM control switch signal drives the inverter to output three-phase voltage and apply it to the permanent magnet synchronous motor again, realizing that the motor speed compensation control law acts on the motor, thereby completing the power stage closed loop.

[0028] It also includes position and velocity detection sensors and an external load; The position and speed detection sensors extract the rotor position and speed of the permanent magnet synchronous motor from the shaft end. The rotor position is used for the real-time coordinate rotation of the Park coordinate transformation module and the inverse Park coordinate transformation module, while the speed is used to provide current for the speed outer loop. The torque of the external load acts directly on the shaft of the permanent magnet synchronous motor. The disturbance is quickly suppressed by the current loop through the motor current, so that the entire control system forms a high-performance drive link that includes power, signal and parameter identification triple closed loops.

[0029] The system is the control system in the artillery servo system, which is used to maintain sub-millisecond response, zero steady-state error and steady-state speed fluctuation within ±1 rpm under conditions of sudden load changes, parameter drift and torque pulsation.

[0030] like Figure 1-7 As shown, to further illustrate the technical solution and technical effects of the present invention, the following specific examples are provided: Specific example 1: Step S1: Using the speed loop of the permanent magnet synchronous motor as the controlled object, establish a hyperlocal model of the permanent magnet synchronous motor; Step S2: Design a deadbeat predictive controller for the speed loop of the permanent magnet synchronous motor control system within the model; Step S3: Design a real-time neural network identifier for the speed loop of the permanent magnet synchronous motor control system within the model to compensate for the deadbeat predictive controller in step S2.

[0031] In step S1, in the two-phase rotating coordinate system (dq coordinate system), the equation of motion of the permanent magnet synchronous motor can be expressed as: (1) In the formula, Represents the angular velocity and moment of inertia of the rotor; Indicates the angular velocity of the rotor; Indicates the number of pole pairs of the motor; Indicates permanent magnet flux linkage; and These represent the d-axis and q-axis currents, respectively. and These represent the d-axis and q-axis inductances, respectively. Indicates load torque; Indicates torque pulsation; This represents the coefficient of viscous friction.

[0032] With rotor angular velocity As output, the desired set current for the q-axis. As a control input, it can be obtained (2) In the formula, the definition is... The lumped disturbance representing the velocity loop.

[0033] In step S2, the hyperlocal model (2) of the permanent magnet synchronous motor established in S1 is discretized using the first-order forward Euler method to obtain the prediction model of the motor speed. (3) To achieve deadbeat predictive control, let ,in, This is the desired setpoint for the motor speed. The value in equation (3) is... use By substitution, we can obtain: (4) The lumped disturbance exists in the dq-axis current and speed prediction equations. When the motor is subjected to large changes in external load or when the inductance, permanent magnet flux linkage, and moment of inertia parameters become mismatched with changes in the working environment, the speed prediction will have a large deviation, which will affect the selection of the optimal voltage vector of the system and thus cause large current and speed tracking errors. Therefore, this invention designs a real-time neural network disturbance observer to identify the lumped disturbances in the current loop and speed loop online, which improves the parameter robustness of the controller while retaining the fast response advantage of deadbeat predictive control.

[0034] In step S3, a neural network identifier model is used as... (5) In the formula, This is the estimated rotational speed at time k+1. The estimate of the lumped perturbation of the speed loop can be achieved by an improved BP neural network with hidden layers, where the neural network identifier of the speed loop is as follows: Figure 1 As shown.

[0035] Figure 1 In this context, N represents a non-linear node, and L represents a linear node. and These are the weight coefficients from the input layer to the hidden layer and from the hidden layer to the output layer, respectively. , , These represent the number of nodes in the hidden layer, input layer, and output layer, respectively. Since the input of this network is... The output should be close to Therefore, the number of neurons in both the input and output layers is set to 1. The number of neurons in the hidden layers is determined during training. A linear function is chosen between the hidden and output layers, and a bipolar sigmoid function is chosen between the input and hidden layers, as shown in the following equation. (6) Define the cost function for neural network learning as follows: (7) During the identification process, the cost function gradually decreases, causing the estimated value of the lumped perturbation to approach the lumped perturbation. Traditional gradient descent algorithms approach the optimal solution at a constant rate. When the learning rate is too small, the convergence speed and adjustment process are too slow. If the learning rate is too large, the convergence speed and adjustment process are fast. However, when approaching the optimal solution, excessive speed can lead to excessive jitter in the steady state and poor motion quality. Therefore, this invention addresses the shortcomings of traditional gradient descent algorithms by designing an adaptive gradient descent algorithm: (8) (9) In the formula, This represents the output of the hidden layer at time k. This is the input to the hidden layer at time k; For learning rate, ; The inertia coefficient, ; This is the rate adjustment coefficient. .

[0036] The improved gradient descent algorithm has the following characteristics: When the cost function value is too large The weight learning rate will be proportional to the deviation of the option value from the previous week, so it has a shorter dynamic response time than the traditional gradient descent algorithm. When the cost function approaches 0 The learning rate tends to This makes its convergence speed slower than the original gradient descent algorithm, which helps reduce system jitter and improve the quality of system control. Meanwhile, for Perform Taylor expansion (local linearization): (10) As can be seen from the above formula, as long as ,So It is a non-positive value, therefore The weights are not increased in each update step. Therefore, this weight update formula can guarantee the asymptotic stability of the system.

[0037] The improved neural network identifier is used to obtain an estimate of the lumped perturbation. Feedforward correction is applied to the deadbeat predictive speed controller to obtain the control law as follows: (11) Figure 2 This is a block diagram of a deadbeat predictive control system for a permanent magnet synchronous motor based on an improved real-time neural network identifier, according to a specific embodiment 1 of the present invention. It consists of a permanent magnet synchronous motor, a three-phase inverter, a position sensor, a deadbeat predictive controller based on an improved real-time neural network identifier, a current loop controller, a coordinate transformation module, and a space vector pulse width modulation module.

[0038] A beat-free predictive controller based on an improved real-time neural network identifier, such as... Figure 3 As shown, when the desired operating speed is set... Then, the improved neural network identifier collects the actual rotation speed at the current moment and... Current And based on these collected values, estimate the motor speed at the next moment. Simultaneously, the estimated value and the expected operating speed are used to calculate the cost function. By minimizing the cost function using the adaptive gradient descent algorithm, an estimate of the lumped perturbation can be obtained. Finally, the estimated value is used to compensate for the control law acting on the motor. The flowchart in this specific example is as follows: Figure 4 As shown.

[0039] To verify the effectiveness of the proposed algorithm, the speed response of the invented deadbeat predictive controller based on an improved real-time neural network identifier was compared with that of a traditional PI controller under different operating conditions (including startup, sudden loading, and constant speed). For ease of comparison, the same controller was used for the current loop of both methods. The main parameters of the motor are: , , , , , The DC bus voltage is 353V, the current loop sampling frequency is 10kHz, and the speed loop sampling frequency is 1kHz. The parameters of the PI speed controller are as follows: , The number of hidden layer neurons in the neural network identifier is set to 5, and the control parameters are set to... , , The comparison results are as follows: Figure 5 — Figure 8 As shown. Figure 5 The traditional PI controller and the algorithm of this invention track the same step signal. The step response of ). Figure 6 The algorithm of this invention is compared with that of a traditional PI controller. Suddenly applied load torque ( ) speed response; Figure 7 Applying torque ripple to a traditional PI controller and the algorithm of this invention under no-load conditions is as follows: steady-state speed fluctuations; Figure 8 The results are obtained from the Fast Fourier Transform analysis of the no-load steady-state speed of the traditional PI controller and the algorithm of this invention.

[0040] Specific example 2: Please see Figure 9 The diagram below illustrates the device structure of a specific example 2 of this application. The device 11 includes a processor 12 and a memory 13 coupled to the processor 12. The memory 13 stores a program file 14 for implementing all the methods described above. The processor 12 executes the program file 14 stored in the memory 13 to implement a deadbeat predictive control method for motors based on a real-time neural network identifier.

[0041] The processor 12, also known as a CPU (Central Processing Unit), can be an integrated circuit chip, a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components, possessing signal processing capabilities. The memory 13 can be any medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, or a device such as a computer, server, mobile phone, or tablet. The program file 14 can be stored in the aforementioned storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of the present invention.

[0042] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predictive control of a motor without deadbeat based on a real-time neural network identifier, characterized in that, Includes the following steps: Step 1: Taking the speed loop of the permanent magnet synchronous motor control system as the controlled object, based on the lumped disturbance of the controlled object, establish a hyperlocal model of the permanent magnet synchronous motor with the desired set current of the q axis in the two-phase rotating dq coordinate system as the control input and the angular velocity of the motor rotor as the output. Step 2: Discretize the hyperlocal model using the first-order forward Euler method to construct a discrete prediction model of the motor speed, and obtain the deadbeat predictive control law of the deadbeat predictive controller of the speed loop based on the deadbeat principle. Step 3: Establish a real-time BP neural network identifier with the motor speed at the current sampling time as input and the lumped disturbance estimate of the speed loop at the next time as output. This is used to identify the lumped disturbance of the speed loop online and to use the obtained lumped disturbance estimate to perform feedforward compensation on the deadbeat predictive control law, generating the final motor speed control law, which is then applied to the motor control system to achieve fast, accurate, and robust control of the motor speed.

2. The motor deadbeat predictive control method based on a real-time neural network identifier as described in claim 1, characterized in that, Step one, as described above, specifically includes: In the two-phase rotating dq coordinate system, the motion equation of the permanent magnet synchronous motor is expressed as: ， In the formula, This represents the moment of inertia of the motor rotor. Indicates the angular velocity of the motor rotor; Indicates the number of pole pairs of the motor; Indicates permanent magnet flux linkage; and These represent the rotating dq-axis currents of the two phases, respectively. and These represent the two-phase rotating dq-axis inductances, respectively. Indicates load torque; Indicates torque pulsation; Indicates the coefficient of viscous friction; Then, with the angular velocity of the motor rotor As output, the desired set current for the q-axis. As the control input, the hyperlocal model of the permanent magnet synchronous motor is obtained as follows: ， In the formula, The lumped disturbance representing the velocity loop, and Represented as: 。 3. The motor deadbeat predictive control method based on a real-time neural network identifier as described in claim 1, characterized in that, Step two specifically includes: The hyperlocal model is discretized using the first-order forward Euler method, resulting in the discrete prediction model for motor speed: ， In the formula, The sampling period of the velocity loop. and They represent the first k The and the first k +1 motor speed value at sampling time, This is the desired setpoint for the motor speed. The desired set current for the q-axis. Indicates permanent magnet flux linkage. Indicates the number of pole pairs of the motor. This represents the moment of inertia of the motor rotor. The lumped disturbance representing the velocity loop; Based on the principle of no beat time, the discrete prediction model is... use Substituting the desired setpoint for the motor speed, the control law of the deadbeat predictive controller for the speed loop is obtained as follows: 。 4. The motor deadbeat predictive control method based on a real-time neural network identifier as described in claim 1, characterized in that, In step three, the established real-time BP neural network identifier is as follows: ， In the formula, for k The estimated value of the motor speed at time +1. Indicates the first k Motor speed at each sampling time, The desired set current for the q-axis. Indicates permanent magnet flux linkage. Indicates the number of pole pairs of the motor. This represents the moment of inertia of the motor rotor. The lumped disturbance estimate for the rotational speed loop is obtained by a real-time BP neural network with a single input and single output and a single hidden layer.

5. The motor deadbeat predictive control method based on a real-time neural network identifier as described in claim 4, characterized in that, The real-time BP neural network includes an input layer, hidden layers, and an output layer. The input layer uses... k Motor speed at any given time The input layer to the hidden layer uses weight coefficients of... The layer has a 1×N weight matrix and a bipolar sigmoid activation function, with multiple nonlinear nodes N in the hidden layer; the weight coefficients from the hidden layer to the output layer are... The N×1 weight vector and linear activation function, the output layer node L, and the output estimate. k Lumped disturbance of the velocity loop at any moment ; in, , , These represent the number of nodes in the hidden layer, input layer, and output layer, respectively. At the same time, the number of neurons in both the input layer and the output layer is 1; The number of neurons in the hidden layer is determined during training. The activation function between the hidden layer and the output layer is a linear function, and the activation function between the input layer and the hidden layer is a bipolar sigmoid function. ， And the cost function is: ， The lumped disturbance of the speed loop is identified online, where, and They are respectively k The estimated and actual values ​​of the motor speed at time +1 express k The rotational speed identification error at time +1, in practical applications The desired setpoint for the motor speed is used instead, and the cost function is gradually reduced during the identification process to ensure the estimation... k Lumped disturbance of the velocity loop at any moment Always approaching k Moment-time velocity lumped disturbance.

6. The motor deadbeat predictive control method based on a real-time neural network identifier as described in claim 4, characterized in that, It also includes the use of an adaptive gradient descent algorithm to identify lumped disturbances in the speed loop in real time, specifically including: ， ， In the formula, and They represent k The changes in weight coefficients from the input layer to the hidden layer and from the hidden layer to the output layer at any given time. For hidden layers in k Output at any moment; This is the input to the hidden layer at time k; For learning rate, ; The inertia coefficient, ; This is the rate adjustment coefficient. , , , These represent the number of nodes in the hidden layer, input layer, and output layer of the real-time BP neural network identifier, respectively. Used when the cost function value is too large. The weight learning rate will be proportional to the deviation of the option value from the previous week, and the learning rate will automatically increase to shorten the dynamic response time. When the cost function approaches zero, The learning rate tends to This slows down the convergence speed compared to the original gradient descent algorithm, meaning the learning rate is automatically reduced to suppress jitter and ensure asymptotic stability.

7. The motor deadbeat predictive control method based on a real-time neural network identifier as described in any one of claims 1-6, characterized in that, The lumped disturbance estimate of the speed loop is obtained using the aforementioned real-time BP neural network identifier. By performing feedforward correction on the deadbeat predictive speed controller, the control law for the motor speed is obtained as follows: ， In the formula, The desired set current for the q-axis. This represents the moment of inertia of the motor rotor. Indicates the number of pole pairs of the motor. Indicates permanent magnet flux linkage. The sampling period of the velocity loop. This is the desired setpoint for the motor speed. This indicates the motor speed value. This is the estimated value of the lumped disturbance in the rotational speed loop.

8. A motor beat-free predictive control system based on a real-time neural network identifier as described in claims 1-7, characterized in that, The system includes a deadbeat predictive controller for executing the deadbeat predictive control method for motors based on a real-time neural network identifier, a permanent magnet synchronous motor, a three-phase inverter, a current loop controller, Clark and Park coordinate transformation, and a space vector pulse width modulation module. The three-phase current output by the permanent magnet synchronous motor is transformed into the current in the α-β stationary coordinate system in the Clark coordinate transformation module, and then transformed into the feedback current value in the dq rotating coordinate system by the Park coordinate transformation module. After setting the desired motor operating speed in the deadbeat predictive controller, the real-time BP neural network in the deadbeat predictive controller collects the actual motor speed and current at the current moment, and estimates the estimated motor speed at the next moment based on the actual motor speed and current. Then, the deadbeat predictive controller uses the estimated motor speed and the desired motor operating speed to calculate the cost function, and minimizes the cost function through the adaptive gradient descent algorithm to obtain the lumped disturbance estimate of the current loop. Then, the lumped disturbance estimate is used to feedforward compensate the control law of the deadbeat predictive controller to generate the motor speed compensation control law. The motor speed control law and the lumped disturbance estimate are input to the current loop controller. The resulting control voltage is transformed to the voltage in the α-β stationary coordinate system by the inverse Park coordinate transformation module, and then sent to the space vector pulse width modulation module to generate a PWM control switch signal. The PWM control switch signal drives the inverter to output three-phase voltage and apply it to the permanent magnet synchronous motor again, realizing that the motor speed compensation control law acts on the motor, thereby completing the power stage closed loop.

9. The motor beat-free predictive control system based on a real-time neural network identifier as described in claim 8 further includes position and speed detection sensors and an external load; The position and speed detection sensors extract the rotor position and speed of the permanent magnet synchronous motor from the shaft end. The rotor position is used for the real-time coordinate rotation of the Park coordinate transformation module and the inverse Park coordinate transformation module, while the speed is used to provide current for the speed outer loop. The torque of the external load acts directly on the shaft of the permanent magnet synchronous motor. The disturbance is quickly suppressed by the current loop through the motor current, so that the entire control system forms a high-performance drive link that includes power, signal and parameter identification triple closed loops.

10. The motor deadbeat predictive control system based on a real-time neural network identifier as described in claim 8, wherein the system is a control system in an artillery servo system, used to maintain sub-millisecond response, zero steady-state error and steady-state speed fluctuation within ±1 rpm under conditions of sudden load changes, parameter drift and torque pulsation.

Images