An aeronautical component induction heating power control method and system based on BP neural network

Abstract

This invention provides a method and system for induction heating power control based on a BP neural network. The method includes: determining a corresponding power setpoint based on the heat treatment process curve of an aerospace component; calculating the power error and error change rate based on the power setpoint and actual output power, and constructing a five-dimensional input vector for the BP neural network; mapping PI parameters based on the neural network output parameters, and obtaining the proportional and integral coefficients of the current sampling period by combining the basic PI parameters and the mapping gain, to calculate the phase shift angle control quantity; calculating a dynamic index based on the absolute value of the power error and the absolute value of the error change rate, and obtaining a dynamic factor after amplitude limiting; calculating the dynamic learning rate of the proportional and integral channels based on the dynamic factor; locking the current PI parameters and pausing the neural network weight update when the output power continuously meets the steady-state criterion. This invention can significantly improve power control efficiency.

Description

Technical Field

[0001] This invention relates to the field of power control technology, and in particular to a method and system for controlling the induction heating power of aerospace components based on a BP neural network. Background Technology

[0002] PI control is a classic control strategy, consisting of proportional and integral elements. By adjusting Kp and Ki, it matches the dynamic and static performance of the system. It is widely used in industrial control due to its advantages such as simple structure, no need for precise models, ease of implementation, and strong robustness. However, its performance is highly dependent on parameter tuning. Traditional tuning relies on experience and trial and error, which has the drawbacks of strong blindness, long debugging cycle, and inability to be corrected online. Fixed parameters are difficult to adapt to nonlinear and strongly time-varying objects.

[0003] Induction heating is a core technology for precision heat treatment of aerospace components. It relies on electromagnetic induction to generate heat inside the workpiece and features rapid heating, low energy consumption, and easy automation. It is widely used in the heat treatment processes of key components such as aero-engine blades and landing gear. However, during the heating process, the resistivity, permeability, and coil coupling state of the workpiece continuously change with temperature, and the equivalent load impedance fluctuates dynamically. The power regulation object exhibits strong nonlinearity and fast time-varying characteristics. Traditional fixed-parameter PI control is difficult to balance fast response and steady-state accuracy, and is prone to problems such as large overshoot, slow adjustment, and large steady-state error during startup, power step jumps, and sudden load changes.

[0004] To address the aforementioned issues, the BP-PI control strategy has been gradually applied due to its nonlinear mapping and self-learning capabilities. However, traditional BP neural networks use a fixed global learning rate, lack adaptability to operating conditions, and suffer from slow convergence and susceptibility to getting trapped in local minima; furthermore, K... p With K i Due to differences in parameter magnitude and mechanism of action, sharing a global learning rate can easily lead to gradient mismatch and parameter coupling problems. Summary of the Invention

[0005] Based on this, the purpose of this invention is to provide a method and system for controlling the induction heating power of aerospace components based on a BP neural network, so as to solve the problems of existing technologies that use a fixed global learning rate, lack the ability to adapt to working conditions, have slow convergence, are prone to getting trapped in local minima, and cause gradient mismatch and parameter coupling.

[0006] The first aspect of the present invention proposes: A method for controlling induction heating power based on a BP neural network, wherein the method includes: Collect voltage and current operating data on the output side of the inverter bridge, calculate the actual output power in the current sampling period, and determine the corresponding power setpoint based on the heat treatment process curve of aerospace components; The power error and error change rate are calculated based on the power setpoint and the actual output power. After normalization, a five-dimensional input vector of the BP neural network is constructed for forward propagation calculation of the BP neural network. PI parameter mapping is performed based on the output parameters of the neural network. The proportional coefficient and integral coefficient of the current sampling period are obtained by combining the basic PI parameters and the mapping gain, so as to calculate the phase shift angle control quantity. The dynamic index is calculated based on the absolute value of the power error and the absolute value of the error change rate, and the dynamic factor is obtained after amplitude limiting. The dynamic learning rates of the proportional and integral channels are calculated based on the dynamic factor, and a steady-state criterion is set. When the output power continuously meets the steady-state criterion, the current PI parameter is locked and the neural network weight update is paused. When the output power does not meet the steady-state criterion and the phase shift control quantity has not entered the saturation state, the quadratic function of the power tracking error is used as the performance index, and the gradient descent method combined with the dynamic learning rate is used to complete the backpropagation weight update of the BP neural network.

[0007] The beneficial effects of this invention are as follows: This technical solution can achieve precise adaptive control of the output power of the heat treatment power supply for aerospace components. By collecting the voltage and current on the output side of the inverter bridge to calculate the actual output power, matching the heat treatment process curve to determine the power setpoint, and combining the power error and error change rate to construct a five-dimensional input vector to drive the BP neural network, the online dynamic mapping and tuning of the PI control parameters is realized. By introducing a dynamic factor and a channel-specific dynamic learning rate mechanism, the weight update rate can be adaptively adjusted according to the system operating state. With the steady-state criterion, the PI parameters are locked and the neural network weight update is paused when the power is stable. This not only ensures rapid response and power tracking accuracy under dynamic conditions, but also avoids parameter oscillation and redundant calculation under steady state, effectively improving the stability and control accuracy of the power supply output power, and ensuring the consistency and processing quality of the heat treatment process for aerospace components.

[0008] Furthermore, the step of collecting voltage and current operating data on the inverter bridge output side and calculating the actual output power includes: In the kth sampling period, the voltage and current on the output side of the inverter bridge are collected by the voltage sampling circuit and the current sampling circuit, respectively. Perform a moving average processing on the power calculation results of one or more sampling periods; A nonlinear equivalent model of output power with respect to phase shift angle is established, wherein the output power changes nonlinearly with phase shift angle, and the temperature change of aerospace components causes the equivalent load parameters to be time-varying; the actual feedback quantity of the power closed loop is the actual output power obtained by sampling calculation.

[0009] Furthermore, the step of constructing the five-dimensional input vector of the BP neural network includes: Calculate the current power error and the rate of change of error based on the power setpoint and the actual output power; Min-Max normalization was performed on the power error, error rate of change and phase shift angle respectively to eliminate the influence of different input variable dimensions and numerical ranges; The normalized current error, the previous sampling error, the rate of change of the current error, the rate of change of the previous sampling error, and the current phase shift angle are combined to form a five-dimensional input vector.

[0010] Furthermore, the step of performing forward propagation calculations in the BP neural network includes: Before control begins, the weight matrix W1 from the input layer to the hidden layer, the weight matrix W2 from the hidden layer to the output layer, and the bias vectors b1 and b2 are initialized; the hidden layer input is calculated: Where W1(k) is a 5×5 weight matrix, b1(k) is a 5-dimensional bias vector, X(k) is a 5-dimensional input vector, and B in (k) represents the hidden layer input; The hidden layer uses the Tanh function as the activation function, and the output of the hidden layer is:

[0011] Where g represents the hidden layer activation function, B out (k) represents the hidden layer output. The Tanh function has an output range of (-1, 1) and can simultaneously characterize both positive and negative power error states. The output layer input is calculated as follows: Where W2(k) is a 2×5 weight matrix, b2(k) is a 2-dimensional bias vector, and C in (k) represents the input of the output layer; The output layer uses the Sigmoid function as the activation function. The output range of the Sigmoid function is (0,1), which matches the output value range mapped by the PI parameter.

[0012] Furthermore, the step of mapping PI parameters based on neural network output parameters and calculating the phase shift angle control quantity includes: The scaling factor K for the current sampling period is calculated based on the two parameters output by the neural network, combined with the basic PI parameters and the mapping gain. p (k) and the integral coefficient K i (k):

[0013] Among them, K pbase and K ibase These are the basic proportional coefficient and basic integral coefficient, determined by the conventional PI tuning results under rated operating conditions of the induction heating power supply. They are used to ensure that the closed-loop system still has basic stability when the neural network has not completed online adjustment; a and b are K... pChannel and K i Channel mapping gain; C out1 (k) C out2 (k) represents the output values ​​of the two neurons in the output layer; Calculate the phase shift angle control increment based on the PI parameters; The current phase shift control value is obtained based on the phase shift control value of the previous sampling period.

[0014] Furthermore, the steps of calculating the dynamic factor and performing dynamic learning rate and steady-state locking include: The dynamic index is calculated based on the absolute value of the power error and the absolute value of the error change rate; the dynamic index is subjected to a limiting process to obtain the dynamic factor; among them, the dynamic factor approaches 1 when the power setpoint changes abruptly during the start-up phase or when the output power approaches the setpoint and the change slows down; the dynamic factor approaches 0 when the output power approaches the setpoint and the change slows down. Calculate K based on the dynamic factor. p Channel and K i Channel learning rate:

[0015] in, η p0 , η i0 K p Channel and K i The base learning rate of the channel, and α and β are the dynamic gain coefficients of the corresponding channels, respectively. F df ( k ) represents a dynamic factor; Set a steady-state criterion, and after the steady-state criterion is met, maintain the current K. p (k) and K i (k) Pauses the update of neural network weights and biases, while power sampling, PI operation and phase shifting drive continue to be executed; when the power error exceeds the threshold, the parameter lock is released and the online update of the neural network is resumed.

[0016] Furthermore, the steps of performing backpropagation and weight update in the BP neural network include: When the output power does not meet the steady-state criterion and the phase shift control quantity has not entered the saturation state, a quadratic performance index function of power tracking error is constructed; the gradient descent method is used to update the backpropagation weights, and the local error gradient of the two neurons in the output layer is calculated based on the chain rule; Construct a dual-channel learning rate vector, calculate the partial derivatives of the objective function with respect to the output layer weight matrix W2(k) and the bias term b2(k) according to the chain rule, and multiply the dual-channel learning rate with the corresponding gradient element by element through the Hadamard product to obtain the weight and bias update rules from the hidden layer to the output layer. Calculate the local gradient vector of the hidden layer, update the hidden layer parameters using a dynamic learning rate with proportional channels, calculate the partial derivatives of the objective function with respect to the input layer weight matrix W1(k) and the bias term b1(k), and obtain the weight and bias update rules from the input layer to the hidden layer; complete the iterative update of the weight matrix and bias vector of each layer according to the update rules.

[0017] The second aspect of the present invention proposes: An induction heating power control system based on a BP neural network, wherein the system comprises: The acquisition module is used to acquire voltage and current operating data on the output side of the inverter bridge, calculate the actual output power in the current sampling period, and determine the corresponding power setpoint based on the heat treatment process curve of aerospace components. The calculation module is used to calculate the power error and error change rate based on the power setpoint and the actual output power. After normalization, it constructs a five-dimensional input vector for the BP neural network to perform forward propagation calculation. The mapping module is used to perform PI parameter mapping based on the neural network output parameters. It combines the basic PI parameters and the mapping gain to obtain the proportional coefficient and integral coefficient of the current sampling period, so as to calculate the phase shift angle control quantity. The processing module is used to calculate dynamic indicators based on the absolute value of power error and the absolute value of error change rate, and obtain dynamic factors after amplitude limiting processing; The update module is used to calculate the dynamic learning rate of the proportional channel and the integral channel according to the dynamic factor, and set the steady-state criterion. When the output power continuously meets the steady-state criterion, the current PI parameter is locked and the neural network weight update is paused. The iterative module, when the output power does not meet the steady-state criterion and the phase shift angle control quantity has not entered the saturation state, uses the quadratic function of the power tracking error as the performance index, and adopts the gradient descent method combined with the dynamic learning rate to complete the backpropagation weight iterative update of the BP neural network.

[0018] Furthermore, the step of collecting voltage and current operating data on the inverter bridge output side and calculating the actual output power includes: In the kth sampling period, the voltage and current on the output side of the inverter bridge are collected by the voltage sampling circuit and the current sampling circuit, respectively. Perform a moving average processing on the power calculation results of one or more sampling periods; A nonlinear equivalent model of output power with respect to phase shift angle is established, wherein the output power changes nonlinearly with phase shift angle, and the temperature change of aerospace components causes the equivalent load parameters to be time-varying; the actual feedback quantity of the power closed loop is the actual output power obtained by sampling calculation.

[0019] Furthermore, the step of constructing the five-dimensional input vector of the BP neural network includes: Calculate the current power error and the rate of change of error based on the power setpoint and the actual output power; Min-Max normalization was performed on the power error, error rate of change and phase shift angle respectively to eliminate the influence of different input variable dimensions and numerical ranges; The normalized current error, the previous sampling error, the rate of change of the current error, the rate of change of the previous sampling error, and the current phase shift angle are combined to form a five-dimensional input vector.

[0020] Furthermore, the step of performing forward propagation calculations in the BP neural network includes: Before control begins, the weight matrix W1 from the input layer to the hidden layer, the weight matrix W2 from the hidden layer to the output layer, and the bias vectors b1 and b2 are initialized; the hidden layer input is calculated: Where W1(k) is a 5×5 weight matrix, b1(k) is a 5-dimensional bias vector, X(k) is a 5-dimensional input vector, and B in (k) represents the hidden layer input; The hidden layer uses the Tanh function as the activation function, and the output of the hidden layer is:

[0021] Where g represents the hidden layer activation function, B out (k) represents the hidden layer output. The Tanh function has an output range of (-1, 1) and can simultaneously characterize both positive and negative power error states. The output layer input is calculated as follows: Where W2(k) is a 2×5 weight matrix, b2(k) is a 2-dimensional bias vector, and C in (k) represents the input of the output layer; The output layer uses the Sigmoid function as the activation function. The output range of the Sigmoid function is (0,1), which matches the output value range mapped by the PI parameter.

[0022] Furthermore, the step of mapping PI parameters based on neural network output parameters and calculating the phase shift angle control quantity includes: The scaling factor K for the current sampling period is calculated based on the two parameters output by the neural network, combined with the basic PI parameters and the mapping gain. p (k) and the integral coefficient K i (k):

[0023] Among them, K pbase and K ibase These are the basic proportional coefficient and basic integral coefficient, determined by the conventional PI tuning results under rated operating conditions of the induction heating power supply. They are used to ensure that the closed-loop system still has basic stability when the neural network has not completed online adjustment; a and b are K...p Channel and K i Channel mapping gain; C out1 (k) C out2 (k) represents the output values ​​of the two neurons in the output layer; Calculate the phase shift angle control increment based on the PI parameters; The current phase shift control value is obtained based on the phase shift control value of the previous sampling period.

[0024] Furthermore, the steps of calculating the dynamic factor and performing dynamic learning rate and steady-state locking include: The dynamic index is calculated based on the absolute value of the power error and the absolute value of the error change rate; the dynamic index is subjected to a limiting process to obtain the dynamic factor; among them, the dynamic factor approaches 1 when the power setpoint changes abruptly during the start-up phase or when the output power approaches the setpoint and the change slows down; the dynamic factor approaches 0 when the output power approaches the setpoint and the change slows down. Calculate K based on the dynamic factor. p Channel and K i Channel learning rate:

[0025] in, η p0 , η i0 K p Channel and K i The base learning rate of the channel, and α and β are the dynamic gain coefficients of the corresponding channels, respectively. F df ( k ) represents a dynamic factor; Set a steady-state criterion, and after the steady-state criterion is met, maintain the current K. p (k) and K i (k) Pauses the update of neural network weights and biases, while power sampling, PI operation and phase shifting drive continue to be executed; when the power error exceeds the threshold, the parameter lock is released and the online update of the neural network is resumed.

[0026] Furthermore, the steps of performing backpropagation and weight update in the BP neural network include: When the output power does not meet the steady-state criterion and the phase shift control quantity has not entered the saturation state, a quadratic performance index function of power tracking error is constructed; the gradient descent method is used to update the backpropagation weights, and the local error gradient of the two neurons in the output layer is calculated based on the chain rule; Construct a dual-channel learning rate vector, calculate the partial derivatives of the objective function with respect to the output layer weight matrix W2(k) and the bias term b2(k) according to the chain rule, and multiply the dual-channel learning rate with the corresponding gradient element by element through the Hadamard product to obtain the weight and bias update rules from the hidden layer to the output layer. Calculate the local gradient vector of the hidden layer, update the hidden layer parameters using a dynamic learning rate with proportional channels, calculate the partial derivatives of the objective function with respect to the input layer weight matrix W1(k) and the bias term b1(k), and obtain the weight and bias update rules from the input layer to the hidden layer; complete the iterative update of the weight matrix and bias vector of each layer according to the update rules.

[0027] The third aspect of the present invention proposes: A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the induction heating power control method based on a BP neural network as described above.

[0028] The fourth aspect of the present invention proposes: A readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the induction heating power control method based on a BP neural network as described above.

[0029] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0030] Figure 1 A block diagram illustrating the BP-PI power control principle in the BP neural network-based induction heating power control method provided in the second embodiment of the present invention. Figure 2 The diagram shows the BP neural network structure in the induction heating power control method based on BP neural network provided in the second embodiment of the present invention. Figure 3 The flowchart of the BP neural network algorithm in the induction heating power control method based on BP neural network provided in the second embodiment of the present invention is shown.

[0031] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0032] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0033] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0034] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0035] The induction heating power control method based on BP neural network provided in the first embodiment of the present invention can ensure both rapid response and power tracking accuracy under dynamic operating conditions, and avoid parameter oscillation and redundant calculation under steady state, effectively improving the stability and control accuracy of power output, and ensuring the consistency and processing quality of heat treatment process for aerospace components.

[0036] Specifically, this embodiment provides: A method for controlling induction heating power based on a BP neural network, wherein the method includes: Collect voltage and current operating data on the output side of the inverter bridge, calculate the actual output power in the current sampling period, and determine the corresponding power setpoint based on the heat treatment process curve of aerospace components; The power error and error change rate are calculated based on the power setpoint and the actual output power. After normalization, a five-dimensional input vector of the BP neural network is constructed for forward propagation calculation of the BP neural network. PI parameter mapping is performed based on the output parameters of the neural network. The proportional coefficient and integral coefficient of the current sampling period are obtained by combining the basic PI parameters and the mapping gain, so as to calculate the phase shift angle control quantity. The dynamic index is calculated based on the absolute value of the power error and the absolute value of the error change rate, and the dynamic factor is obtained after amplitude limiting. The dynamic learning rates of the proportional and integral channels are calculated based on the dynamic factor, and a steady-state criterion is set. When the output power continuously meets the steady-state criterion, the current PI parameter is locked and the neural network weight update is paused. When the output power does not meet the steady-state criterion and the phase shift control quantity has not entered the saturation state, the quadratic function of the power tracking error is used as the performance index, and the gradient descent method combined with the dynamic learning rate is used to complete the backpropagation weight update of the BP neural network.

[0037] The second embodiment of the present invention provides: Furthermore, the step of collecting voltage and current operating data on the inverter bridge output side and calculating the actual output power includes: In the kth sampling period, the voltage and current on the output side of the inverter bridge are collected by the voltage sampling circuit and the current sampling circuit, respectively. Perform a moving average processing on the power calculation results of one or more sampling periods; A nonlinear equivalent model of output power with respect to phase shift angle is established, wherein the output power changes nonlinearly with phase shift angle, and the temperature change of aerospace components causes the equivalent load parameters to be time-varying; the actual feedback quantity of the power closed loop is the actual output power obtained by sampling calculation.

[0038] Furthermore, the step of constructing the five-dimensional input vector of the BP neural network includes: Calculate the current power error and the rate of change of error based on the power setpoint and the actual output power; Min-Max normalization was performed on the power error, error rate of change and phase shift angle respectively to eliminate the influence of different input variable dimensions and numerical ranges; The normalized current error, the previous sampling error, the rate of change of the current error, the rate of change of the previous sampling error, and the current phase shift angle are combined to form a five-dimensional input vector.

[0039] Furthermore, the step of performing forward propagation calculations in the BP neural network includes: Before control begins, the weight matrix W1 from the input layer to the hidden layer, the weight matrix W2 from the hidden layer to the output layer, and the bias vectors b1 and b2 are initialized; the hidden layer input is calculated: Where W1(k) is a 5×5 weight matrix, b1(k) is a 5-dimensional bias vector, X(k) is a 5-dimensional input vector, and B in (k) represents the hidden layer input; The hidden layer uses the Tanh function as the activation function, and the output of the hidden layer is:

[0040] Where g represents the hidden layer activation function, B out (k) represents the hidden layer output. The Tanh function has an output range of (-1, 1) and can simultaneously characterize both positive and negative power error states. The output layer input is calculated as follows: Where W2(k) is a 2×5 weight matrix, b2(k) is a 2-dimensional bias vector, and C in (k) represents the input of the output layer; The output layer uses the Sigmoid function as the activation function. The output range of the Sigmoid function is (0,1), which matches the output value range mapped by the PI parameter.

[0041] Furthermore, the step of mapping PI parameters based on neural network output parameters and calculating the phase shift angle control quantity includes: The scaling factor K for the current sampling period is calculated based on the two parameters output by the neural network, combined with the basic PI parameters and the mapping gain. p (k) and the integral coefficient K i (k):

[0042] Among them, K pbase and K ibase These are the basic proportional coefficient and basic integral coefficient, determined by the conventional PI tuning results under rated operating conditions of the induction heating power supply. They are used to ensure that the closed-loop system still has basic stability when the neural network has not completed online adjustment; a and b are K... p Channel and K i Channel mapping gain; C out1 (k) C out2 (k) represents the output values ​​of the two neurons in the output layer; Calculate the phase shift angle control increment based on the PI parameters; The current phase shift control value is obtained based on the phase shift control value of the previous sampling period.

[0043] Furthermore, the steps of calculating the dynamic factor and performing dynamic learning rate and steady-state locking include: The dynamic index is calculated based on the absolute value of the power error and the absolute value of the error change rate; the dynamic index is subjected to a limiting process to obtain the dynamic factor; among them, the dynamic factor approaches 1 when the power setpoint changes abruptly during the start-up phase or when the output power approaches the setpoint and the change slows down; the dynamic factor approaches 0 when the output power approaches the setpoint and the change slows down. Calculate K based on the dynamic factor. p Channel and K i Channel learning rate:

[0044] in, η p0 , η i0 K p Channel and K i The base learning rate of the channel, and α and β are the dynamic gain coefficients of the corresponding channels, respectively. F df ( k ) represents a dynamic factor; Set a steady-state criterion, and after the steady-state criterion is met, maintain the current K. p (k) and K i(k) Pauses the update of neural network weights and biases, while power sampling, PI operation and phase shifting drive continue to be executed; when the power error exceeds the threshold, the parameter lock is released and the online update of the neural network is resumed.

[0045] Furthermore, the steps of performing backpropagation and weight update in the BP neural network include: When the output power does not meet the steady-state criterion and the phase shift control quantity has not entered the saturation state, a quadratic performance index function of power tracking error is constructed; the gradient descent method is used to update the backpropagation weights, and the local error gradient of the two neurons in the output layer is calculated based on the chain rule; Construct a dual-channel learning rate vector, calculate the partial derivatives of the objective function with respect to the output layer weight matrix W2(k) and the bias term b2(k) according to the chain rule, and multiply the dual-channel learning rate with the corresponding gradient element by element through the Hadamard product to obtain the weight and bias update rules from the hidden layer to the output layer. Calculate the local gradient vector of the hidden layer, update the hidden layer parameters using a dynamic learning rate with proportional channels, calculate the partial derivatives of the objective function with respect to the input layer weight matrix W1(k) and the bias term b1(k), and obtain the weight and bias update rules from the input layer to the hidden layer; complete the iterative update of the weight matrix and bias vector of each layer according to the update rules.

[0046] In addition, in this embodiment, it should be noted that, as Figures 1 to 3 As shown, this embodiment presents a power control method for induction heating of aerospace components based on dynamic optimization of PI parameters using a BP neural network. During operation, the induction heating power supply for aerospace components experiences dynamic fluctuations in the equivalent impedance of the series resonant load due to changes in the workpiece resistivity, permeability, and coupling state between the induction coil and the workpiece with temperature. This results in nonlinear and time-varying characteristics between the output power and the phase shift angle. The precision heat treatment processes of aerospace components such as aero-engine blades, landing gear, and structural parts place high demands on heating power response speed, steady-state accuracy, and process consistency. Traditional fixed-parameter PI control is insufficient to meet the power control requirements under complex operating conditions. A power control system for induction heating of aerospace components is proposed as follows: Figure 1 As shown, it includes a phase-shifted full-bridge inverter circuit, a series resonant load, a voltage sampling circuit, a current sampling circuit, a controller, and a phase-shifted drive circuit. The BP neural network structure is as follows: Figure 2 As shown, the algorithm control flow is as follows: Figure 3 As shown.

[0047] This invention, from the perspective of closed-loop power control of induction heating power supplies, combines a BP neural network with traditional PI control in a phase-shifted full-bridge induction heating power supply. This enables closed-loop control of the output power of the induction heating power supply. The specific implementation steps are as follows: S101, Acquire operating data and calculate actual output power: In the k-th sampling cycle, the voltage sampling circuit and the current sampling circuit respectively acquire the output voltage and current of the inverter bridge. The controller calculates the actual output power y(k) based on the voltage and current sampling values ​​within the same power calculation window. Assuming a power calculation window contains M sets of synchronously sampled data, the actual output power can be expressed as:

[0048] Among them, u o (k,m) and i o (k, m) represent the m-th sampled output voltage and current values ​​within the k-th sampling period, respectively. To reduce the impact of switching noise and sampling glitches, a moving average can be applied to the power calculation results of one or more switching periods. The power setpoint r(k) is set by the controller based on the heat treatment process curve of the aerospace component; when a constant power heating process is executed, r(k) remains constant; when a segmented heating process is executed, r(k) changes according to the process curve.

[0049] Inverter bridge output voltage fundamental RMS value U out With DC bus voltage U dc The mathematical relationship between the phase shift angle θ and the phase shift angle θ is:

[0050] The total complex impedance of the series resonant load is:

[0051] The magnitude of the total complex impedance is:

[0052] According to circuit theory, the active power P of the system is mainly determined by the fundamental component, and its expression is:

[0053] Substituting the expressions for the fundamental effective value of the inverter bridge output voltage and the total complex impedance of the series resonant load into the active power expression, we can obtain the nonlinear function of the output power with respect to the phase shift angle:

[0054] Where R, L, and C represent the equivalent resistance, equivalent inductance, and equivalent capacitance of the series resonant load, respectively, and ω is the inverter angular frequency. From the above relationships, it can be seen that the output power changes nonlinearly with the phase shift angle. Simultaneously, temperature changes in aerospace components cause changes in the workpiece resistivity, permeability, and coupling state between the coil and the workpiece, thus altering the equivalent load parameters. Therefore, the induction heating power regulation object exhibits nonlinear and time-varying characteristics. The above equivalent model is used to illustrate the characteristics of the controlled object and determine the sensitivity of the output power to the direction of phase shift angle changes. The actual feedback quantity in the power closed loop is still y(k) calculated based on voltage and current sampling data; the controller is not required to rely on an accurate identification model for power control.

[0055] S102, Calculate the error state and construct the neural network input vector: Calculate the current power error based on the power setpoint r(k) and the actual output power y(k): e(k) = r(k) - y(k), where the error change rate is defined as Δe(k) = e(k) - e(k-1). To eliminate the influence of different input variable dimensions and numerical ranges, Min-Max normalization is performed on the power error, error change rate, and phase shift angle. The normalized current error, previous sampling error, current error change rate, previous sampling error change rate, and current phase shift angle constitute a five-dimensional input vector.

[0056] Where u(k) represents the phase shift angle state currently stored and applied to the phase-shifted full-bridge inverter circuit by the controller in the k-th sampling period. After completing the PI increment calculation for the current sampling period, the controller refreshes this phase shift angle state.

[0057] S103, Perform forward propagation of the BP neural network: Before control begins, initialize the weight matrix W1 from the input layer to the hidden layer, the weight matrix W2 from the hidden layer to the output layer, and the bias vectors b1 and b2.

[0058] The hidden layer input is:

[0059] Where W1(k) is a 5×5 weight matrix and b1(k) is a 5-dimensional bias vector. The output range of the Tanh function is (-1, 1), which can characterize the positive and negative power error states. Therefore, the Tanh function is used for nonlinear mapping in the hidden layer activation function, and its expression and derivative form are as follows:

[0060] The hidden layer output is:

[0061] The input to the output layer is:

[0062] Where W2(k) is a 2×5 weight matrix and b2(k) is a 2-dimensional bias vector. The output range of the Sigmoid function is (0,1), which satisfies the output range of the output parameters. Therefore, the Sigmoid function is used as the activation function of the output layer, and its expression and derivative form are as follows:

[0063] The output layer outputs:

[0064] S104, Perform PI parameter mapping and phase shift control quantity calculation: The neural network output is not directly used as the PI parameter, but the proportional coefficient and integral coefficient of the current sampling period are obtained through the basic PI parameter and the mapping gain:

[0065] Among them, K pbase and K ibase These are the basic proportional coefficient and the basic integral coefficient, respectively, and a and b are K... p Channel and K i The channel's mapping gain. The basic PI parameters are determined based on the conventional PI tuning results of the induction heating power supply under rated operating conditions, and are used to ensure that the closed-loop system still has basic stability when the neural network has not completed online adjustment.

[0066] Calculate the phase shift angle control increment based on the output parameters:

[0067] The current phase shift control value is obtained based on the phase shift control value of the previous sampling period:

[0068] S105, Calculate the dynamic factor F df (k): Calculate dynamic indicators based on the absolute value of power error and the absolute value of error change rate.

[0069] By limiting the dynamic index, we obtain the dynamic factor:

[0070] Dynamic factor F df The range of (k) is [0,1]. During the startup phase or when the power setpoint undergoes a step change, the power error is relatively large, F df (k) approaches 1; when the workpiece temperature change causes a sudden change in the equivalent load impedance, the error change rate increases rapidly, F df(k) increases accordingly; when the actual output power is close to the set power and the trend of change slows down, F df (k) gradually approaches 0.

[0071] S106, Implement dynamic learning rate and steady-state locking: Calculate K based on the dynamic factor. p Channel and K i Channel learning rate:

[0072] Where, η p0 η i0 K p Channel and K i The base learning rate for each channel, and α and β are the dynamic gain coefficients for the corresponding channels, respectively. When F df When (k) approaches 1, the two-channel learning rate increases relative to the corresponding base learning rate; when F df When (k) approaches 0, the learning rates of the two channels fall back to the corresponding base learning rates.

[0073] The output power is considered to have entered a steady state when the following conditions are met consecutively for N sampling periods:

[0074] Where, ε e This represents the steady-state error threshold. After satisfying the steady-state criterion, maintain the current K value. p (k) and K i (k) Pauses the update of neural network weights and biases, but power sampling, PI calculation, and phase shifting drive continue to execute. When the power error exceeds the threshold, the parameter lock is released and the online update of the neural network resumes.

[0075] S107, perform backpropagation and weight update: when the output power does not meet the steady-state criterion and the phase shift angle has not entered the saturation state, the quadratic function of the power tracking error is used as the performance index:

[0076] The BP neural network uses gradient descent to update weights through backpropagation. The specific rules are as follows: The local error gradients of the two neurons in the output layer are δ1 and δ2, respectively. (2) (k) and δ2 (2) (k), expanded based on the chain rule, is as follows:

[0077] Let H be the learning rate vector of the two independent channels during weight update. η (k)=[η p (k),η iAccording to the chain rule, the partial derivatives of the objective function with respect to the output layer weight coefficient matrix W2(k) and the bias term b2(k) can be expanded as follows:

[0078] Therefore, the update rules for the weight coefficient matrix W2(k) and the bias term b2(k) from the hidden layer to the output layer are as follows:

[0079] In the formula, This is the Hadamard product, which enables element-wise multiplication of the dual-channel learning rate with the corresponding gradient.

[0080] Let the local gradient vector of the hidden layer be δ. (1) (k), according to the backpropagation chain rule, the gradient calculation formula for the hidden layer is:

[0081] The hidden layer uses a dynamic learning rate η with proportional channels. p (k) is updated. According to the chain rule, the partial derivatives of the objective function with respect to the hidden layer input weight coefficient matrix and bias terms can be expanded as follows:

[0082] Therefore, the update rules for the weight matrix W1(k) and bias term b1(k) from the input layer to the hidden layer are as follows:

[0083] S108 executes phase-shift control and closed-loop iteration: The phase-shift angle control value obtained from PI control is applied to the phase-shifted full-bridge inverter circuit, changing the phase difference between the bridge arm drive signals, thereby changing the effective energy obtained by the resonant load. To improve engineering safety, upper and lower limits and single-cycle change rate limits can be set for the phase-shift angle to avoid excessive phase shift abrupt changes causing inverter bridge switching stress or resonant current surges. After the phase-shift execution is completed, the next sampling cycle begins and S101 to S108 are executed again, ensuring that the output power of the induction heating power supply continuously tracks the target power.

[0084] The aforementioned method combines a BP neural network intelligent algorithm with traditional PI control for power control of induction heating power supplies in aerospace components. The resulting controller can adjust the PI control parameters Kp and Ki online based on power error, error rate of change, and phase angle state. Furthermore, it adjusts the dynamic learning rate through a dynamic factor mechanism, enabling the controller to adapt to the nonlinear and time-varying characteristics of the induction heating load caused by temperature changes. In the precision heat treatment of aerospace components, this method can improve power response speed and reduce overshoot and steady-state power fluctuations under conditions such as startup, power step changes, and load disturbances. It addresses the difficulty of traditional fixed-parameter PI control in balancing dynamic response and steady-state accuracy, thereby improving the accuracy, stability, and engineering adaptability of induction heating power control.

[0085] The third embodiment of the present invention provides: An induction heating power control system based on a BP neural network, wherein the system comprises: The acquisition module is used to acquire voltage and current operating data on the output side of the inverter bridge, calculate the actual output power in the current sampling period, and determine the corresponding power setpoint based on the heat treatment process curve of aerospace components. The calculation module is used to calculate the power error and error change rate based on the power setpoint and the actual output power. After normalization, it constructs a five-dimensional input vector for the BP neural network to perform forward propagation calculation. The mapping module is used to perform PI parameter mapping based on the neural network output parameters. It combines the basic PI parameters and the mapping gain to obtain the proportional coefficient and integral coefficient of the current sampling period, so as to calculate the phase shift angle control quantity. The processing module is used to calculate dynamic indicators based on the absolute value of power error and the absolute value of error change rate, and obtain dynamic factors after amplitude limiting processing; The update module is used to calculate the dynamic learning rate of the proportional channel and the integral channel according to the dynamic factor, and set the steady-state criterion. When the output power continuously meets the steady-state criterion, the current PI parameter is locked and the neural network weight update is paused. The iterative module, when the output power does not meet the steady-state criterion and the phase shift angle control quantity has not entered the saturation state, uses the quadratic function of the power tracking error as the performance index, and adopts the gradient descent method combined with the dynamic learning rate to complete the backpropagation weight iterative update of the BP neural network.

[0086] Furthermore, the step of collecting voltage and current operating data on the inverter bridge output side and calculating the actual output power includes: In the kth sampling period, the voltage and current on the output side of the inverter bridge are collected by the voltage sampling circuit and the current sampling circuit, respectively. Perform a moving average processing on the power calculation results of one or more sampling periods; A nonlinear equivalent model of output power with respect to phase shift angle is established, wherein the output power changes nonlinearly with phase shift angle, and the temperature change of aerospace components causes the equivalent load parameters to be time-varying; the actual feedback quantity of the power closed loop is the actual output power obtained by sampling calculation.

[0087] Furthermore, the step of constructing the five-dimensional input vector of the BP neural network includes: Calculate the current power error and the rate of change of error based on the power setpoint and the actual output power; Min-Max normalization was performed on the power error, error rate of change and phase shift angle respectively to eliminate the influence of different input variable dimensions and numerical ranges; The normalized current error, the previous sampling error, the rate of change of the current error, the rate of change of the previous sampling error, and the current phase shift angle are combined to form a five-dimensional input vector.

[0088] Furthermore, the step of performing forward propagation calculations in the BP neural network includes: Before control begins, the weight matrix W1 from the input layer to the hidden layer, the weight matrix W2 from the hidden layer to the output layer, and the bias vectors b1 and b2 are initialized; the hidden layer input is calculated: Where W1(k) is a 5×5 weight matrix, b1(k) is a 5-dimensional bias vector, X(k) is a 5-dimensional input vector, and B in (k) represents the hidden layer input; The hidden layer uses the Tanh function as the activation function, and the output of the hidden layer is:

[0089] Where g represents the hidden layer activation function, B out (k) represents the hidden layer output. The Tanh function has an output range of (-1, 1) and can simultaneously characterize both positive and negative power error states. The output layer input is calculated as follows: Where W2(k) is a 2×5 weight matrix, b2(k) is a 2-dimensional bias vector, and C in (k) represents the input of the output layer; The output layer uses the Sigmoid function as the activation function. The output range of the Sigmoid function is (0,1), which matches the output value range mapped by the PI parameter.

[0090] Furthermore, the step of mapping PI parameters based on neural network output parameters and calculating the phase shift angle control quantity includes: The scaling factor K for the current sampling period is calculated based on the two parameters output by the neural network, combined with the basic PI parameters and the mapping gain. p (k) and the integral coefficient K i (k):

[0091] Among them, K pbase and K ibase These are the basic proportional coefficient and basic integral coefficient, determined by the conventional PI tuning results under rated operating conditions of the induction heating power supply. They are used to ensure that the closed-loop system still has basic stability when the neural network has not completed online adjustment; a and b are K... p Channel and K i Channel mapping gain; C out1 (k) C out2 (k) represents the output values ​​of the two neurons in the output layer; Calculate the phase shift angle control increment based on the PI parameters; The current phase shift control value is obtained based on the phase shift control value of the previous sampling period.

[0092] Furthermore, the steps of calculating the dynamic factor and performing dynamic learning rate and steady-state locking include: The dynamic index is calculated based on the absolute value of the power error and the absolute value of the error change rate; the dynamic index is subjected to a limiting process to obtain the dynamic factor; among them, the dynamic factor approaches 1 when the power setpoint changes abruptly during the start-up phase or when the output power approaches the setpoint and the change slows down; the dynamic factor approaches 0 when the output power approaches the setpoint and the change slows down. Calculate K based on the dynamic factor. p Channel and K i Channel learning rate:

[0093] in, η p0 , η i0 K p Channel and K i The base learning rate of the channel, and α and β are the dynamic gain coefficients of the corresponding channels, respectively. F df ( k ) represents a dynamic factor; Set a steady-state criterion, and after the steady-state criterion is met, maintain the current K. p (k) and K i (k) Pauses the update of neural network weights and biases, while power sampling, PI operation and phase shifting drive continue to be executed; when the power error exceeds the threshold, the parameter lock is released and the online update of the neural network is resumed.

[0094] Furthermore, the steps of performing backpropagation and weight update in the BP neural network include: When the output power does not meet the steady-state criterion and the phase shift control quantity has not entered the saturation state, a quadratic performance index function of power tracking error is constructed; the gradient descent method is used to update the backpropagation weights, and the local error gradient of the two neurons in the output layer is calculated based on the chain rule; Construct a dual-channel learning rate vector, calculate the partial derivatives of the objective function with respect to the output layer weight matrix W2(k) and the bias term b2(k) according to the chain rule, and multiply the dual-channel learning rate with the corresponding gradient element by element through the Hadamard product to obtain the weight and bias update rules from the hidden layer to the output layer. Calculate the local gradient vector of the hidden layer, update the hidden layer parameters using a dynamic learning rate with proportional channels, calculate the partial derivatives of the objective function with respect to the input layer weight matrix W1(k) and the bias term b1(k), and obtain the weight and bias update rules from the input layer to the hidden layer; complete the iterative update of the weight matrix and bias vector of each layer according to the update rules.

[0095] The fourth embodiment of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the induction heating power control method based on the BP neural network as described above.

[0096] The fifth embodiment of the present invention provides a readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, it implements the induction heating power control method based on a BP neural network as described above.

[0097] In summary, the above embodiments of the present invention can ensure both rapid response and power tracking accuracy under dynamic operating conditions, while avoiding parameter oscillations and redundant calculations under steady-state conditions, effectively improving the stability and control accuracy of power supply output power, and ensuring the consistency and processing quality of heat treatment processes for aerospace components.

[0098] It should be noted that the above modules can be functional modules or program modules, and can be implemented through software or hardware. For modules implemented through hardware, the above modules can reside in the same processor; or the above modules can be located in different processors in any combination.

[0099] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0100] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0101] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0102] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0103] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for controlling induction heating power based on a BP neural network, characterized in that, The method includes: Collect voltage and current operating data on the output side of the inverter bridge, calculate the actual output power in the current sampling period, and determine the corresponding power setpoint based on the heat treatment process curve of aerospace components; The power error and error change rate are calculated based on the power setpoint and the actual output power. After normalization, a five-dimensional input vector of the BP neural network is constructed for forward propagation calculation of the BP neural network. PI parameter mapping is performed based on the output parameters of the neural network. The proportional coefficient and integral coefficient of the current sampling period are obtained by combining the basic PI parameters and the mapping gain, so as to calculate the phase shift angle control quantity. The dynamic index is calculated based on the absolute value of the power error and the absolute value of the error change rate, and the dynamic factor is obtained after amplitude limiting. The dynamic learning rates of the proportional and integral channels are calculated based on the dynamic factor, and a steady-state criterion is set. When the output power continuously meets the steady-state criterion, the current PI parameter is locked and the neural network weight update is paused. When the output power does not meet the steady-state criterion and the phase shift control quantity has not entered the saturation state, the quadratic function of the power tracking error is used as the performance index, and the gradient descent method combined with the dynamic learning rate is used to complete the backpropagation weight update of the BP neural network.

2. The induction heating power control method based on a BP neural network according to claim 1, characterized in that, The steps of collecting voltage and current operating data on the output side of the inverter bridge and calculating the actual output power include: In the kth sampling period, the voltage and current on the output side of the inverter bridge are collected by the voltage sampling circuit and the current sampling circuit, respectively. Perform a moving average processing on the power calculation results of one or more sampling periods; A nonlinear equivalent model of output power with respect to phase shift angle is established, wherein the output power changes nonlinearly with phase shift angle, and the temperature change of aerospace components causes the equivalent load parameters to be time-varying; the actual feedback quantity of the power closed loop is the actual output power obtained by sampling calculation.

3. The induction heating power control method based on a BP neural network according to claim 1, characterized in that, The steps for constructing the five-dimensional input vector of the BP neural network include: Calculate the current power error and the rate of change of error based on the power setpoint and the actual output power; Min-Max normalization was performed on the power error, error rate of change and phase shift angle respectively to eliminate the influence of different input variable dimensions and numerical ranges; The normalized current error, the previous sampling error, the rate of change of the current error, the rate of change of the previous sampling error, and the current phase shift angle are combined to form a five-dimensional input vector.

4. The induction heating power control method based on a BP neural network according to claim 1, characterized in that, The steps for performing forward propagation calculations in a BP neural network include: Before control begins, the weight matrix W1 from the input layer to the hidden layer, the weight matrix W2 from the hidden layer to the output layer, and the bias vectors b1 and b2 are initialized; the hidden layer input is calculated: Where W1(k) is a 5×5 weight matrix, b1(k) is a 5-dimensional bias vector, X(k) is a 5-dimensional input vector, and B... in (k) represents the hidden layer input; The hidden layer uses the Tanh function as the activation function, and the output of the hidden layer is: Where g represents the hidden layer activation function, B out (k) represents the hidden layer output. The Tanh function has an output range of (-1, 1) and can simultaneously characterize both positive and negative power error states. The output layer input is calculated as follows: Where W2(k) is a 2×5 weight matrix, b2(k) is a 2-dimensional bias vector, and C in (k) represents the input of the output layer; The output layer uses the Sigmoid function as the activation function. The output range of the Sigmoid function is (0,1), which matches the output value range mapped by the PI parameter.

5. The induction heating power control method based on a BP neural network according to claim 4, characterized in that, The steps of mapping PI parameters based on neural network output parameters and calculating the phase shift angle control quantity include: The scaling factor K for the current sampling period is calculated based on the two parameters output by the neural network, combined with the basic PI parameters and the mapping gain. p (k) and integral coefficient K i (k): Among them, K pbase and K ibase These are the basic proportional coefficient and basic integral coefficient, determined by the conventional PI tuning results under rated operating conditions of the induction heating power supply. They are used to ensure that the closed-loop system still has basic stability when the neural network has not completed online adjustment; a and b are K... p Channel and K i Channel mapping gain; C out1 (k) C out2 (k) represents the output values ​​of the two neurons in the output layer; Calculate the phase shift angle control increment based on the PI parameters; The current phase shift control value is obtained based on the phase shift control value of the previous sampling period.

6. The induction heating power control method based on a BP neural network according to claim 5, characterized in that, The steps of calculating the dynamic factor and performing dynamic learning rate and steady-state locking include: The dynamic index is calculated based on the absolute value of the power error and the absolute value of the error change rate; the dynamic index is subjected to a limiting process to obtain the dynamic factor; among them, the dynamic factor approaches 1 when the power setpoint changes abruptly during the start-up phase or when the output power approaches the setpoint and the change slows down; the dynamic factor approaches 0 when the output power approaches the setpoint and the change slows down. Calculate K based on the dynamic factor. p Channel and K i Channel learning rate: in, η p0 , η i0 K p Channel and K i The base learning rate for each channel, and α and β are the dynamic gain coefficients for the corresponding channels, respectively. F df ( k ) represents a dynamic factor; Set a steady-state criterion, and after the steady-state criterion is met, maintain the current K. p (k) and K i (k) Pauses the update of neural network weights and biases, while power sampling, PI operation and phase shifting drive continue to be executed; when the power error exceeds the threshold, the parameter lock is released and the online update of the neural network is resumed.

7. The induction heating power control method based on a BP neural network according to claim 6, characterized in that, The steps for backpropagation and weight update of the BP neural network include: When the output power does not meet the steady-state criterion and the phase shift control quantity has not entered the saturation state, a quadratic performance index function of power tracking error is constructed; the gradient descent method is used to update the backpropagation weights, and the local error gradient of the two neurons in the output layer is calculated based on the chain rule; Construct a dual-channel learning rate vector, calculate the partial derivatives of the objective function with respect to the output layer weight matrix W2(k) and the bias term b2(k) according to the chain rule, and multiply the dual-channel learning rate with the corresponding gradient element by element through the Hadamard product to obtain the weight and bias update rules from the hidden layer to the output layer. Calculate the local gradient vector of the hidden layer, update the hidden layer parameters using a dynamic learning rate with proportional channels, calculate the partial derivatives of the objective function with respect to the input layer weight matrix W1(k) and the bias term b1(k), and obtain the weight and bias update rules from the input layer to the hidden layer; complete the iterative update of the weight matrix and bias vector of each layer according to the update rules.

8. An induction heating power control system based on a BP neural network, characterized in that, The system includes: The acquisition module is used to acquire voltage and current operating data on the output side of the inverter bridge, calculate the actual output power in the current sampling period, and determine the corresponding power setpoint based on the heat treatment process curve of aerospace components. The calculation module is used to calculate the power error and error change rate based on the power setpoint and the actual output power. After normalization, it constructs a five-dimensional input vector for the BP neural network to perform forward propagation calculation. The mapping module is used to perform PI parameter mapping based on the neural network output parameters. It combines the basic PI parameters and the mapping gain to obtain the proportional coefficient and integral coefficient of the current sampling period, so as to calculate the phase shift angle control quantity. The processing module is used to calculate dynamic indicators based on the absolute value of power error and the absolute value of error change rate, and obtain dynamic factors after amplitude limiting processing; The update module is used to calculate the dynamic learning rate of the proportional channel and the integral channel according to the dynamic factor, and set the steady-state criterion. When the output power continuously meets the steady-state criterion, the current PI parameter is locked and the neural network weight update is paused. The iterative module, when the output power does not meet the steady-state criterion and the phase shift angle control quantity has not entered the saturation state, uses the quadratic function of the power tracking error as the performance index, and adopts the gradient descent method combined with the dynamic learning rate to complete the backpropagation weight iterative update of the BP neural network.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the induction heating power control method based on a BP neural network as described in any one of claims 1 to 7.

10. A readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the induction heating power control method based on a BP neural network as described in any one of claims 1 to 7.

Images