A load disturbance control method for permanent magnet synchronous motor

Abstract

The present application provides a method for controlling a permanent magnet synchronous motor against load disturbances, which relates to the technical field of load disturbance resistance, and includes: configuring an autoregressive integrated moving average model and a memory network based on the obtained time domain characteristics, frequency domain characteristics, rate of change, and normalized operating physical signals to obtain short-term prediction results and long-term prediction results; integrating the short-term prediction results and the long-term prediction results to obtain a final disturbance prediction result; and calculating an error based on the comparison of the final disturbance prediction result with the operating physical signal. By combining the autoregressive integrated moving average model and the time memory network for disturbance prediction, and by fusing the short-term and long-term prediction results, the operating state of the motor under load disturbance can be accurately predicted, the accuracy of disturbance prediction can be significantly improved, and a more reliable basis can be provided for the calculation of feedforward compensation current.

Description

Technical Field

[0001] The present invention relates to the technical field of load disturbance resistance, and in particular to a load disturbance resistance control method for a permanent magnet synchronous motor. Background Art

[0002] Permanent magnet synchronous motors are widely used in industrial automation, home appliances, electric vehicles and other fields due to their high efficiency, low loss and high power density. As the motor load changes, especially under complex load disturbances, the operating stability of the motor will be seriously affected. In order to maintain the efficient operation of the motor system, the control system needs to be able to perceive and effectively compensate for load disturbances in real time.

[0003] Existing load disturbance resistance control methods mainly rely on traditional PID control or simple prediction models, which have two major shortcomings: First, when dealing with complex nonlinear disturbances, traditional methods cannot accurately predict the type and degree of disturbances, resulting in poor compensation effect; second, most control methods lack a dynamic adjustment mechanism for compensation errors and cannot effectively deal with disturbances with large frequency changes, resulting in slow system response and insufficient compensation accuracy in actual applications. Summary of the Invention

[0004] The purpose of the present invention is to address the shortcomings of the existing technology and provide a load disturbance resistance control method for a permanent magnet synchronous motor. The method aims to combine time domain and frequency domain feature analysis, an autoregressive integrated moving average model and a hybrid prediction method of a memory network, so as to accurately capture and predict the dynamic changes of load disturbances. By dynamically adjusting the sliding mode control surface and the fusion weight, the operating errors of the motor under various load disturbances can be effectively compensated, thereby improving the stability and compensation accuracy, and overcoming the shortcomings of insufficient prediction accuracy and weak dynamic adjustment capability in the existing technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A method for controlling a permanent magnet synchronous motor against load disturbances includes the following steps:

[0007] Step S100: Acquire the operating physical signal of the permanent magnet synchronous motor, perform preprocessing and normalization processing, analyze the local characteristics of the operating physical signal after normalization through a sliding time window, and obtain the disturbance type and disturbance degree.

[0008] Step S200: Based on the obtained time domain features, frequency domain features, change rate and normalized operating physical signal, an autoregressive integrated moving average model and a memory network are configured to obtain short-term prediction results and long-term prediction results. The short-term prediction results and the long-term prediction results are integrated to obtain a final disturbance prediction result. The error is calculated based on the comparison between the final disturbance prediction result and the operating physical signal.

[0009] Step S300: Calculate the feedforward current compensation based on the final disturbance prediction result to obtain a feedforward compensation current signal; extract the low-frequency component and the high-frequency component of the feedforward compensation current signal through a low-pass filter and a band-pass filter to obtain a low-frequency compensation signal and a high-frequency compensation signal; obtain the reference current through the motor control system of the permanent magnet synchronous motor; superimpose the obtained low-frequency compensation signal, the high-frequency compensation signal and the reference current to obtain a total compensation current; perform dynamic compensation adjustment based on the obtained total compensation current, and compare before and after the dynamic compensation adjustment.

[0010] Step S400: Adjust the parameters of the autoregressive integrated moving average model based on the obtained compensated error data and compare them; optimize the memory network based on the obtained compensated error data; optimize the fusion weights based on the optimized autoregressive integrated moving average model and the memory network; fuse the short-term prediction results and the long-term prediction results based on the fusion weight optimization results to obtain a fused disturbance prediction result; and verify the results based on the obtained fused disturbance prediction result.

[0011] Step S500: Calculate the residual error based on the obtained fusion disturbance prediction result and the running physical signal error data to obtain the residual error data, configure the sliding mode control surface based on the obtained residual error data, adopt the non-singular sliding mode control law to make the sliding mode control surface converge quickly, obtain the sliding mode control output signal, combine the sliding mode control output signal with the total compensation current, and generate the final compensation signal.

[0012] Step S600: Compare the obtained final compensation signal with the running physical signal and the fusion disturbance prediction result to obtain the final residual error, perform frequency domain decomposition to obtain high-frequency error components and low-frequency error components, dynamically adjust the sliding mode control surface weight parameters and fusion weights based on the obtained high-frequency error components and low-frequency error components, optimize the final compensation signal generation logic, generate the final optimized compensation signal based on the dynamically adjusted sliding mode control surface weight parameters and fusion weights, and inject the final optimized compensation signal into the motor control loop.

[0013] As a preferred solution of the present invention, the operating physical signal of the permanent magnet synchronous motor is obtained, and preprocessed and normalized. Based on the normalized operating physical signal, the local characteristics are analyzed through a sliding time window to obtain the disturbance type and disturbance degree, which specifically includes:

[0014] Step S100.1: Acquire the operating physical signal of the permanent magnet synchronous motor.

[0015] Operating physical signals include: current, voltage, speed and torque.

[0016] The current in the running physical signal is obtained through the current sensor, the voltage in the running physical signal is obtained through the voltage sensor, the speed in the running physical signal is obtained through the speed sensor, and the torque in the running physical signal is obtained through the load torque sensor.

[0017] Step S100.2: Pre-process the obtained operating physical signal through a filter.

[0018] Step S100.2.1: Perform normalization processing based on the pre-processed running physical signal.

[0019] It should be noted here that, since preprocessing and normalization processing are already well known technologies to those skilled in the art, they will not be described in detail here.

[0020] Step S100.3: Based on the normalized running physical signal, analyze the local characteristics through a sliding time window.

[0021] The time window is set to be greater than 0mS, the sliding window is set to be greater than 0mS, the normalized running physical signal is divided by greater than 0mS through the time window, and more than 0 time data points are obtained. The normalized running physical signal is divided by greater than 0mS through the sliding window, and more than 0 sliding data points are obtained.

[0022] Feature extraction is performed based on the obtained time data points and sliding data points.

[0023] Feature extraction calculates the mean and variance of the current, voltage, speed and torque of the time data points and sliding data points through mean calculation and variance calculation to obtain the time domain features, performs frequency domain analysis on the time data points and sliding data points through fast Fourier transform, extracts the frequency domain features, identifies the high-frequency disturbance and low-frequency fluctuation components of the frequency domain features through the spectrum diagram, calculates the instantaneous change rate of the time data points and sliding data points, and obtains the change rate data.

[0024] Based on the time domain features, frequency domain features and change rate data obtained by feature extraction, a machine learning algorithm is used to classify the disturbance.

[0025] The categories include: load mutation, inertial disturbance and periodic disturbance.

[0026] According to the obtained time domain characteristics, frequency domain characteristics and change rate data, the amplitude and frequency of the disturbance are judged, and the disturbance amplitude and disturbance frequency are obtained.

[0027] As a preferred solution of the present invention, an autoregressive integrated moving average model and a memory network are configured based on the obtained time domain features, frequency domain features, change rate and normalized operating physical signals to obtain short-term prediction results and long-term prediction results. The short-term prediction results and long-term prediction results are integrated to obtain a final disturbance prediction result. The error is calculated based on the comparison between the final disturbance prediction result and the operating physical signal, specifically:

[0028] Step S200.1: Based on the obtained time domain features, frequency domain features, change rate and normalized operating physical signals, an autoregressive integrated moving average model is configured to obtain a short-term prediction result.

[0029] Step S200.2: Based on the obtained short-term prediction results, time domain features, frequency domain features and change rates, a memory network is configured to obtain a long-term prediction result.

[0030] Step S200.3: Based on the obtained short-term prediction results and long-term prediction results, weighted fusion is used to integrate them to obtain the final disturbance prediction result.

[0031] Step S200.4: Calculate the error based on the comparison between the final disturbance prediction result and the operating physical signal.

[0032] As a preferred solution of the present invention, the feedforward current compensation is calculated based on the final disturbance prediction result to obtain a feedforward compensation current signal, and the low-frequency component and high-frequency component of the feedforward compensation current signal are extracted by a low-pass filter and a band-pass filter to obtain a low-frequency compensation signal and a high-frequency compensation signal. The reference current is obtained by the motor control system of the permanent magnet synchronous motor, and the obtained low-frequency compensation signal, high-frequency compensation signal and reference current are superimposed to obtain a total compensation current. Dynamic compensation adjustment is performed based on the obtained total compensation current, and a comparison before and after the dynamic compensation adjustment is performed, specifically as follows:

[0033] Step S300.1: Calculate feedforward current compensation based on the final disturbance prediction result to obtain a feedforward compensation current signal.

[0034] Step S300.1.1: Extract the low-frequency component in the feedforward compensation current signal through a low-pass filter to obtain a low-frequency compensation signal.

[0035] Step S300.1.2: Extract the high-frequency component in the feedforward compensation current signal through a bandpass filter to obtain a high-frequency compensation signal.

[0036] Step S300.1.3: Obtain a reference current of a current control loop from a motor control system of the permanent magnet synchronous motor.

[0037] Step S300.2: derive a total compensation current based on the superposition of the obtained low-frequency compensation signal, the high-frequency compensation signal, and the reference current.

[0038] Based on the obtained total compensation current, it is injected into the permanent magnet synchronous motor control loop through the current controller to dynamically compensate and adjust the input current of the permanent magnet synchronous motor to offset the influence of the predicted disturbance.

[0039] Step S300.3: perform comparison based on the results of dynamic compensation adjustment.

[0040] The compensation error data is obtained by calculating the error of the disturbance amplitude and disturbance frequency before and after dynamic compensation adjustment.

[0041] As a preferred solution of the present invention, the parameters of the autoregressive integrated moving average model are adjusted based on the obtained compensation error data, and a comparison is performed. The memory network is optimized based on the obtained compensation error data. The fusion weight is optimized based on the optimized autoregressive integrated moving average model and the memory network. The short-term prediction results and the long-term prediction results are fused based on the fusion weight optimization results to obtain a fused disturbance prediction result. The fused disturbance prediction result is verified based on the obtained fused disturbance prediction result, specifically:

[0042] Step S400.1: Optimize the parameters p, d, and q of the autoregressive integrated moving average model based on the obtained compensated error data.

[0043] The grid search method is used to find the optimal parameter combination p, autoregressive term order d, difference order q, and moving average term order again. After obtaining the optimal parameter combination, residual analysis is performed.

[0044] Step S400.2: Compare the parameters p, d, and q of the optimized autoregressive integrated moving average model.

[0045] Step S400.3: Optimize the memory network based on the obtained compensation error data.

[0046] Step S400.4: Optimize the fusion weight based on the optimized autoregressive integrated moving average model and memory network.

[0047] Step S400.4.1: Based on the fusion weight optimization result, the short-term prediction result and the long-term prediction result are fused to obtain the fusion disturbance prediction result.

[0048] Step S400.5: Verify based on the obtained fusion disturbance prediction result.

[0049] The fusion disturbance prediction results and the running physical signal are compared by the mean square error and mean square error to obtain the running physical signal error data. If the error meets the threshold requirement: MSE threshold ≤ 0.05Nm 2, when the MAE threshold is ≤ 0.1Nm, the optimized autoregressive integrated moving average model and memory network are applied. If the error still does not meet the requirements, the iterative optimization is continued until it reaches the threshold range.

[0050] As a preferred solution of the present invention, the residual error is calculated based on the obtained fusion disturbance prediction result and the running physical signal error data to obtain the residual error data, the sliding mode control surface is configured based on the obtained residual error data, the non-singular sliding mode control law is used to make the sliding mode control surface converge quickly, the sliding mode control output signal is obtained, and the sliding mode control output signal is combined with the total compensation current to generate the final compensation signal, which specifically includes:

[0051] Step S500.1: Calculate the residual error based on the obtained fusion disturbance prediction result and the running physical signal error data to obtain the residual error data, specifically:

[0052] E residual (t+τ)=ΔT(t+τ)-(ΔT)^(t+τ)

[0053] Where: E residual (t+τ) is the residual error, which represents the difference between the prediction and the actual value after compensation. ΔT(t+τ) is the actual disturbance amplitude of the running physical signal. (ΔT)^(t+τ) is the disturbance prediction result after fusion.

[0054] To E residual (t+τ) is decomposed in the frequency domain to obtain high-frequency error and low-frequency error components.

[0055] Step S500.2: Configure the sliding mode control surface based on the obtained residual error data, specifically:

[0056]

[0057] Where: S(t) is the sliding mode control surface, which is used to quantify the system error state, e(t) is the real-time residual error, and E residual (t) is equal, e(t)=E residual (t), is the rate of change of the residual error, which is used to reflect the dynamic change of the error over time. c1(t) is the dynamic weight parameter for error amplification, and c2(t) is the dynamic weight parameter for error change rate amplification.

[0058] The sliding mode gain is dynamically adjusted based on the obtained high-frequency error and low-frequency error components, specifically:

[0059]

[0060] Where: k1, k2 are reference gains, α, β are error sensitivity adjustment coefficients, is the absolute value of the low-frequency error component, is the absolute value of the high-frequency error component.

[0061] The non-singular sliding mode control law is used to make the sliding mode control surface converge quickly, and the sliding mode control output signal is obtained, which is specifically:

[0062] u(t)=-K·sgn(S(t))

[0063] Where u(t) is the sliding mode control output signal, which is used to dynamically adjust the compensation current; K is the sliding mode gain coefficient, which controls the compensation response amplitude; and sgn(S(t)) is the sliding surface sign function, which is used to indicate the control direction.

[0064] The sliding mode control output signal is combined with the total compensation current to generate the final compensation signal, which is:

[0065]

[0066] Where: I final (t) is the total control current finally injected into the motor, I total (t) is the synthesis result of feedforward and frequency domain compensation signals, is the sliding mode feedback compensation signal.

[0067] As a preferred solution of the present invention, the final compensation signal is compared with the running physical signal and the fusion disturbance prediction result to obtain the final residual error, and the frequency domain decomposition is performed to obtain the high-frequency error component and the low-frequency error component. The sliding mode control surface weight parameters and the fusion weight are dynamically adjusted based on the obtained high-frequency error component and the low-frequency error component, the final compensation signal generation logic is optimized, and the final optimized compensation signal is generated based on the dynamically adjusted sliding mode control surface weight parameters and the fusion weight. The final optimized compensation signal is injected into the motor control loop, which specifically includes:

[0068] Step S600.1: Compare the obtained final compensation signal with the running physical signal and the fusion disturbance prediction result to obtain the final residual error, which is specifically:

[0069]

[0070] Where: E residual (t) is the residual error, which represents the difference between the predicted value after compensation and the actual disturbance value, ΔT(t) is the actual disturbance amplitude in the running physical signal, is the final disturbance prediction result.

[0071] By performing frequency domain decomposition on the final residual error, high-frequency error components and low-frequency error components are obtained.

[0072] Step S600.1.1: Dynamically adjust the sliding mode control surface weight parameters and fusion weights based on the obtained high-frequency error components and low-frequency error components to optimize the final compensation signal generation logic. Specifically:

[0073]

[0074] Where: S(t) is the sliding mode control surface, which is used to quantify the residual error state, e(t) is the real-time residual error, and E residual (t) are equal, is the rate of change of the residual error, c1(t) and c2(t) are dynamically adjusted weight parameters:

[0075]

[0076] k1 and k2 are reference gains, α and β are error sensitivity coefficients, and the compensation signal ratio is adjusted according to the high-frequency error component and the low-frequency error component. Specifically:

[0077]

[0078] Where: α is the dynamic proportional coefficient of the low-frequency compensation signal, 1-α is the dynamic proportional coefficient of the high-frequency compensation signal, is the absolute value of the low-frequency error in the residual error, is the absolute value of the high-frequency error in the residual error.

[0079] Step S600.2: Generate a final optimized compensation signal based on the dynamically adjusted sliding mode control surface weight parameters and fusion weights, specifically:

[0080]

[0081] Where: I final (t) is the total control current finally injected into the motor control system. I total (t) is the fused compensation signal, including feedforward compensation and frequency domain compensation. is the sliding mode control feedback signal.

[0082] The final optimized compensation signal is injected into the motor control loop.

[0083] Obtain the running physical signal and compare it with the fusion disturbance prediction result, specifically:

[0084] E final (t) = ΔT(t) - ΔT adjusted (t)

[0085] Where: E final(t) is the final residual error, which represents the difference between the actual disturbance state and the target disturbance value after the final optimized compensation signal is injected. ΔT(t) is the target disturbance value in the running physical signal, which represents the actual disturbance amplitude of the motor operation under ideal conditions. adjusted (t) is the final optimized compensation signal I injected final The adjusted disturbance amplitude after (t) reflects the actual operating state of the motor after closed-loop control.

[0086] If E final (t) Meet the error threshold condition MSE≤0.05Nm 2 , MAE≤0.1Nm, the final optimized compensation signal is considered valid. If the error exceeds the set threshold, the sliding mode control surface weight parameters and fusion weight are adjusted cyclically, and the process stops when the error is satisfied.

[0087] Compared with the prior art, the present invention has the following beneficial effects:

[0088] 1. By combining the autoregressive integrated moving average model and the time memory network for disturbance prediction, the autoregressive integrated moving average model is good at processing linear features, while the memory physics can capture long-term dependencies and nonlinear characteristics. By fusing the short-term and long-term prediction results, the operating state of the motor under load disturbance can be accurately predicted, which significantly improves the accuracy of disturbance prediction and provides a more reliable basis for the calculation of feedforward compensation current.

[0089] 2. By introducing low-pass filters and band-pass filters, the present invention can decompose the feedforward compensation current signal into low-frequency and high-frequency components, adopt different compensation strategies for the low-frequency and high-frequency disturbance components of the motor, and then calculate the total compensation current. This method can accurately control the current compensation, ensuring that the impact of load disturbances is adjusted in real time and accurately during the operation of the motor, effectively improving the dynamic response capability and control accuracy of the motor.

[0090] 3. The system adopts an autoregressive integrated moving average model and a memory network optimization mechanism based on the compensation error data. By adjusting the model parameters and fusion weights, the disturbance prediction results are continuously optimized, and the sliding mode control surface converges quickly, thereby reducing the disturbance residual error. Combined with frequency domain analysis, the system can dynamically adjust the sliding mode control surface weight to ensure that the motor can be efficiently and accurately compensated under complex disturbance conditions, and can adjust the system control strategy in real time to reduce error accumulation.

[0091] 4. By performing frequency domain decomposition on the final compensation signal and dynamically adjusting the sliding mode control surface and fusion weights, adaptive optimization of the motor operation error is achieved. This optimization mechanism can dynamically adjust the control logic according to the frequency domain characteristics of the disturbance, so that the motor can maintain an efficient and stable operation state under different load conditions. This method is particularly suitable for complex load disturbance environments, effectively improving the adaptability and stability of the motor system in practical applications and meeting the needs of high-precision control. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 A flow chart of a method for controlling load disturbances of a permanent magnet synchronous motor provided in an embodiment of the present application. DETAILED DESCRIPTION

[0093] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0094] See also Figure 1 , Figure 1 A flowchart of a method for controlling load disturbances of a permanent magnet synchronous motor is provided in accordance with an embodiment of the present application.

[0095] In this embodiment, a method for controlling a permanent magnet synchronous motor against load disturbance may include steps S100, S200, S300, S400, S500 and S600;

[0096] Step S100: Acquire the operating physical signal of the permanent magnet synchronous motor, perform preprocessing and normalization processing, analyze the local characteristics of the operating physical signal after normalization through a sliding time window, and obtain the disturbance type and disturbance degree.

[0097] Step S200: Based on the obtained time domain features, frequency domain features, change rate and normalized operating physical signal, an autoregressive integrated moving average model and a memory network are configured to obtain short-term prediction results and long-term prediction results. The short-term prediction results and the long-term prediction results are integrated to obtain a final disturbance prediction result. The error is calculated based on the comparison between the final disturbance prediction result and the operating physical signal.

[0098] Step S300: Calculate the feedforward current compensation based on the final disturbance prediction result to obtain a feedforward compensation current signal; extract the low-frequency component and the high-frequency component of the feedforward compensation current signal through a low-pass filter and a band-pass filter to obtain a low-frequency compensation signal and a high-frequency compensation signal; obtain the reference current through the motor control system of the permanent magnet synchronous motor; superimpose the obtained low-frequency compensation signal, the high-frequency compensation signal and the reference current to obtain a total compensation current; perform dynamic compensation adjustment based on the obtained total compensation current, and compare before and after the dynamic compensation adjustment.

[0099] Step S400: Adjust the parameters of the autoregressive integrated moving average model based on the obtained compensated error data and compare them; optimize the memory network based on the obtained compensated error data; optimize the fusion weights based on the optimized autoregressive integrated moving average model and the memory network; fuse the short-term prediction results and the long-term prediction results based on the fusion weight optimization results to obtain a fused disturbance prediction result; and verify the results based on the obtained fused disturbance prediction result.

[0100] Step S500: Calculate the residual error based on the obtained fusion disturbance prediction result and the running physical signal error data to obtain the residual error data, configure the sliding mode control surface based on the obtained residual error data, adopt the non-singular sliding mode control law to make the sliding mode control surface converge quickly, obtain the sliding mode control output signal, combine the sliding mode control output signal with the total compensation current, and generate the final compensation signal.

[0101] Step S600: Compare the obtained final compensation signal with the running physical signal and the fusion disturbance prediction result to obtain the final residual error, perform frequency domain decomposition to obtain high-frequency error components and low-frequency error components, dynamically adjust the sliding mode control surface weight parameters and fusion weights based on the obtained high-frequency error components and low-frequency error components, optimize the final compensation signal generation logic, generate the final optimized compensation signal based on the dynamically adjusted sliding mode control surface weight parameters and fusion weights, and inject the final optimized compensation signal into the motor control loop.

[0102] In a specific embodiment, the operating physical signal of the permanent magnet synchronous motor is obtained, and preprocessed and normalized. Based on the normalized operating physical signal, the local characteristics are analyzed through a sliding time window to obtain the disturbance type and disturbance degree, which specifically includes:

[0103] Step S100.1: Acquire the operating physical signal of the permanent magnet synchronous motor.

[0104] Operating physical signals include: current, voltage, speed and torque.

[0105] The current in the running physical signal is obtained through the current sensor, the voltage in the running physical signal is obtained through the voltage sensor, the speed in the running physical signal is obtained through the speed sensor, and the torque in the running physical signal is obtained through the load torque sensor.

[0106] The frequency of running physical signal acquisition is at least 1kHz, that is, collecting data one thousand times per second.

[0107] Step S100.2: Pre-process the obtained operating physical signal through a filter.

[0108] The outliers and high-frequency noise in the running physical signal are removed by filtering.

[0109] Step S100.2.1: Perform normalization processing based on the pre-processed running physical signal.

[0110] The pre-processed running physical signal is converted into a standardized unit through normalization processing to eliminate the influence of the magnitude difference of different sensors.

[0111] It should be noted here that, since preprocessing and normalization processing are already well known technologies to those skilled in the art, they will not be described in detail here.

[0112] Step S100.3: Based on the normalized running physical signal, analyze the local characteristics through a sliding time window.

[0113] The time window is set to 50 mS, and the sliding window is set to 20 mS. The normalized running physical signal is divided into 50 mS intervals through the time window to obtain 50 time data points. The normalized running physical signal is divided into 20 mS intervals through the sliding window to obtain 20 sliding data points.

[0114] Feature extraction is performed based on the obtained time data points and sliding data points.

[0115] Feature extraction calculates the mean and variance of the current, voltage, speed and torque of the time data points and sliding data points through mean calculation and variance calculation to obtain the time domain features, performs frequency domain analysis on the time data points and sliding data points through fast Fourier transform, extracts the frequency domain features, identifies the high-frequency disturbance and low-frequency fluctuation components of the frequency domain features through the spectrum diagram, calculates the instantaneous change rate of the time data points and sliding data points, and obtains the change rate data.

[0116] Based on the time domain features, frequency domain features and change rate data obtained by feature extraction, a machine learning algorithm is used to classify the disturbance.

[0117] The categories include: load mutation, inertial disturbance and periodic disturbance.

[0118] According to the obtained time domain characteristics, frequency domain characteristics and change rate data, the amplitude and frequency of the disturbance are judged, and the disturbance amplitude and disturbance frequency are obtained.

[0119] In a specific embodiment, an autoregressive integrated moving average model and a memory network are configured based on the obtained time domain features, frequency domain features, change rate, and normalized operating physical signals to obtain short-term prediction results and long-term prediction results. The short-term prediction results and long-term prediction results are integrated to obtain a final disturbance prediction result. The error is calculated based on the comparison between the final disturbance prediction result and the operating physical signal, specifically:

[0120] Step S200.1: Based on the obtained time domain features, frequency domain features, change rate and normalized operating physical signals, an autoregressive integrated moving average model is configured to obtain a short-term prediction result.

[0121] The normalized running physical signal is subjected to differential operation to remove the non-stationary part, and the obtained time domain features, frequency domain features and change rate are decomposed into the stationary part and the non-stationary part.

[0122] The order of the autoregressive integrated moving average model is configured through the autocorrelation function and partial autocorrelation function. The disturbance amplitude and disturbance frequency are linearly configured through the autoregressive integrated moving average model.

[0123] The basic trend and periodic characteristics of the operating physical signals are captured by the autoregressive integrated moving average model to obtain short-term forecast results.

[0124] Step S200.2: Based on the obtained short-term prediction results, time domain features, frequency domain features and change rates, a memory network is configured to obtain a long-term prediction result.

[0125] First, the obtained short-term prediction results, time domain features and frequency domain features are used as the input layer. Secondly, a hidden layer is composed of multiple memory network units. Each unit contains an input gate, a forget gate and an output gate to extract the long-term dependencies and nonlinear characteristics of the running physical signals. Finally, the real-time disturbance amplitude, disturbance frequency and disturbance trend are output.

[0126] The memory network uses mean square error as the loss function and Adam optimizer as the optimization function. The weight parameters of the memory network are dynamically adjusted according to the error of the short-term prediction results.

[0127] Step S200.3: Based on the obtained short-term prediction results and long-term prediction results, weighted fusion is used to integrate them to obtain the final disturbance prediction result, which is specifically:

[0128] ΔT(t+τ)=α·ΔT ARIMA (t+τ)+(1-α)·ΔT LSTM (t+τ)

[0129] Where: ΔT(t+τ) is the predicted value of the disturbance amplitude at the real-time t+τ after fusion, ΔTA RIMA (t+τ) is the short-term forecast result output by the autoregressive integrated moving average model, ΔTL STM (t+τ) is the long-term prediction result output by the memory network, and α is the weight coefficient of the prediction result of the autoregressive integrated moving average model. The value range is 0≤α≤1 and is dynamically adjusted by actual verification and error evaluation. When the data is mainly linear, α is increased, and when the data has obvious nonlinear characteristics, α is decreased.

[0130] Step S200.4: Calculate the error based on the comparison between the final disturbance prediction result and the operating physical signal.

[0131] The error is calculated by comparing the mean square error and the mean square error, specifically:

[0132]

[0133] Where n is the number of data samples, is the predicted disturbance amplitude of the i-th sample point, ΔT i is the actual disturbance amplitude of the i-th sample point, (ΔT i -ΔT i ) 2 It is the square of the error between the predicted value and the actual value of each sample point.

[0134]

[0135] Where: n is the number of data samples, is the predicted disturbance amplitude of the i-th sample point, ΔTi is the actual disturbance amplitude of the i-th sample point, |ΔT i -ΔT i | is the absolute value of the error between the predicted value and the actual value of each sample point.

[0136] The threshold is set to: MSE threshold 0.1Nm 2 , MAE threshold 0.2Nm.

[0137] If the error exceeds the set threshold, the parameters p, d, and q of the autoregressive integrated moving average model are dynamically adjusted to improve the fitting accuracy of the linear modeling, and the weight and learning rate of the memory network are adjusted by the Adam optimizer to reduce the prediction error of the model.

[0138] In a specific embodiment, the feedforward current compensation is calculated based on the final disturbance prediction result to obtain a feedforward compensation current signal, and the low-frequency component and high-frequency component of the feedforward compensation current signal are extracted by a low-pass filter and a band-pass filter to obtain a low-frequency compensation signal and a high-frequency compensation signal. The reference current is obtained by the motor control system of the permanent magnet synchronous motor, and the obtained low-frequency compensation signal, high-frequency compensation signal and reference current are superimposed to obtain a total compensation current. Dynamic compensation adjustment is performed based on the obtained total compensation current, and a comparison before and after the dynamic compensation adjustment is performed, specifically as follows:

[0139] Step S300.1: Calculate the feedforward current compensation based on the final disturbance prediction result to obtain a feedforward compensation current signal, specifically:

[0140]

[0141] Where: ΔT(t+τ) is the predicted disturbance torque, k t is the torque constant of the permanent magnet synchronous motor, I comp (t+τ) is the predicted real-time compensation current signal.

[0142] Step S300.1.1: Extract the low-frequency component in the feedforward compensation current signal through a low-pass filter to obtain a low-frequency compensation signal.

[0143] Step S300.1.2: Extract the high-frequency component in the feedforward compensation current signal through a bandpass filter to obtain a high-frequency compensation signal.

[0144] Step S300.1.3: Obtain a reference current of a current control loop from a motor control system of the permanent magnet synchronous motor.

[0145] Step S300.2: The obtained low-frequency compensation signal, high-frequency compensation signal and reference current are superimposed to obtain a total compensation current, specifically:

[0146] I total (t+τ)=I ref (t)+I comp,HF (t+τ)+I comp,LF (t+τ)

[0147] Where: I total (t+τ) is the final total compensation current, I ref (t) is the reference current of the current control loop, I comp,HF (t+τ) is from I comp The high-frequency compensation current component extracted from (t+τ), I comp,LF (t+τ) is from I comp The low-frequency compensation current component extracted from (t+τ).

[0148] Based on the obtained total compensation current, it is injected into the permanent magnet synchronous motor control loop through the current controller to dynamically compensate and adjust the input current of the permanent magnet synchronous motor to offset the influence of the predicted disturbance.

[0149] Step S300.3: perform comparison based on the results of dynamic compensation adjustment.

[0150] The compensation error data is obtained by comparing the disturbance amplitude and disturbance frequency before and after dynamic compensation adjustment, specifically:

[0151]

[0152] Where: MSE 补偿 is the mean square error after compensation, which is used to measure the residual error of disturbance before and after compensation. n is the number of data samples, ΔT i is the actual disturbance amplitude of the i-th sample point, ΔT comp,i is the compensated disturbance amplitude of the i-th sample point,

[0153]

[0154] Where: MAE 补偿 is the mean absolute error after compensation, which is used to measure the average magnitude of the perturbation residual error, n is the number of data samples, ΔT i is the actual disturbance amplitude of the i-th sample point, ΔT comp,i is the compensated disturbance amplitude of the i-th sample point.

[0155] The error judgment threshold is set to: MSE threshold = 0.05Nm 2 , MAE threshold = 0.1Nm.

[0156] If the error exceeds the threshold, the parameters of the autoregressive integrated moving average model are optimized and the memory network is optimized using the Adam optimizer.

[0157] In a specific implementation, the parameters of the autoregressive integrated moving average model are adjusted based on the obtained compensation error data and compared, the memory network is optimized based on the obtained compensation error data, the fusion weight is optimized based on the optimized autoregressive integrated moving average model and the memory network, the short-term prediction results and the long-term prediction results are fused based on the fusion weight optimization result, and the fusion disturbance prediction result is obtained. The verification is performed based on the obtained fusion disturbance prediction result, specifically:

[0158] Step S400.1: Optimize the parameters p, d, and q of the autoregressive integrated moving average model based on the obtained compensated error data.

[0159] The grid search method is used to find the optimal parameter combination p, autoregressive term order d, difference order q, and moving average term order again. After obtaining the optimal parameter combination, residual analysis is performed.

[0160] The residual analysis is as follows: detect whether the mean of the residuals is close to zero. When the mean of the residuals deviates from zero, optimization needs to be continued. Detect whether the variance of the residuals is small and stable. When the variance of the residuals fluctuates significantly over time, optimization needs to be continued. Use the autocorrelation function to detect whether the residuals have significant autocorrelation. If the residuals show obvious autocorrelation at multiple time lag points, optimization needs to be continued. Analyze whether the distribution of the residuals conforms to the normal distribution characteristics. If the residuals show obvious skewness or peak characteristics, optimization needs to be continued. Perform fast Fourier transform analysis on the residuals to detect whether there are significant periodic components. If the residuals contain periodic components, optimization needs to be continued.

[0161] When the mean of the residuals is -10 -3 ~10 -3 , the residual variance is kept at σ 2 <0.01, and there is no trend of fluctuation over time. In the autocorrelation analysis of the residuals, the values ​​of the autocorrelation function at all lag points fall within the 95% confidence interval, the residual distribution conforms to normality, the p value is greater than 0.05, and there is no obvious periodic peak in the spectrum analysis of the residuals, then the optimization is stopped.

[0162] Step S400.2: Compare the parameters p, d, and q of the optimized autoregressive integrated moving average model.

[0163] The optimized short-term forecast results are obtained by operating the optimized autoregressive integrated moving average model, and the unoptimized short-term forecast results are obtained by operating the unoptimized autoregressive integrated moving average model.

[0164] Based on the comparison between the optimized short-term prediction results and the unoptimized short-term prediction results, the error reduction rate is calculated, specifically:

[0165]

[0166] Where: MSE 旧 is the mean square error between the short-term forecast results generated by the unoptimized autoregressive integrated moving average model and the actual value, and is calculated as:

[0167]

[0168] ΔT i is the actual disturbance amplitude of the i-th sample point, ΔT 旧，i is the disturbance amplitude of the i-th sample point predicted by the unoptimized autoregressive integrated moving average model, n is the total number of samples, and MSE 新It is the mean square error between the short-term forecast results generated by the optimized autoregressive integrated moving average model and the actual value, and the calculation method is the same as MSE 旧 Same as, but using the predicted value ΔT from the optimized model 新，i The error reduction rate (MSE) is the improvement degree of the mean square error of the optimized autoregressive integrated moving average model compared with the unoptimized model, expressed as a percentage. If the error reduction rate is a positive value, it means that the optimization is effective and the mean square error of the short-term forecast is reduced.

[0169]

[0170] Where: MAE 旧 is the mean absolute error between the short-term forecast results generated by the unoptimized autoregressive integrated moving average model and the actual value, and is calculated as:

[0171]

[0172] ΔT i is the actual disturbance amplitude of the i-th sample point, ΔT 旧，i is the disturbance amplitude of the i-th sample point predicted by the unoptimized autoregressive integrated moving average model, n is the total number of samples, MAE 新 It is the mean absolute error between the short-term forecast results generated by the optimized autoregressive integrated moving average model and the actual value, calculated in the same way as MAE 旧 Same as, but using the predicted value ΔT from the optimized model 新，i The error reduction rate (MAE) is the degree of improvement in the mean absolute error of the optimized autoregressive integrated moving average model compared to the unoptimized model, expressed as a percentage. If the error reduction rate is positive, it means that the optimization is effective and the mean absolute error of short-term predictions is reduced.

[0173] Step S400.3: Optimize the memory network based on the obtained compensation error data.

[0174] The mean square error is used as the loss function to measure the difference between the predicted value and the actual value, specifically:

[0175]

[0176] Where: Loss is the loss function value, which is used to measure the deviation between the predicted value and the actual value. The goal is to minimize the loss function, ΔT i is the actual disturbance amplitude of the i-th sample point, is the perturbation amplitude of the i-th sample point predicted by the memory network, and n is the total number of samples, that is, the number of data points included in the short-term prediction.

[0177] The Adam optimizer is used to adaptively update the weights according to the direction and size of the gradient, specifically:

[0178]

[0179] Where: W (t+1) is the weight parameter for the next step, W (t) is the current weight parameter, η is the learning rate, which controls the step size of the weight update, It is the gradient of the loss function with respect to the weight, indicating the direction and magnitude of the weight adjustment.

[0180] According to the size of the compensation error data, the number of neurons in each layer is adjusted. If the prediction error is large, memory network units are added to enhance the feature extraction capability, and the number of memory network layers is increased to adapt to the long-term dependence characteristics of complex disturbances. If the error has stabilized, overfitting is avoided, the network complexity is reduced, and the number of layers is reduced to improve computational efficiency and adapt to simple disturbance environments.

[0181] Step S400.4: Optimize the fusion weight based on the optimized autoregressive integrated moving average model and memory network, specifically:

[0182]

[0183] Where: α is the dynamically adjusted fusion weight coefficient, ranging from 0≤α≤1, reflecting the weight of the autoregressive integrated moving average model in the fusion prediction, MAE ARIMA is the mean absolute error of the autoregressive integrated moving average model in the current forecast period, MAE LSTM is the mean absolute error of the memory network during the current prediction period.

[0184] If the error of the autoregressive integrated moving average model is small, α is large, and it relies more on the short-term prediction results of ARIMA. If the error of the memory network is small, α is small, and it relies more on the long-term prediction results of the memory network.

[0185] Step S400.4.1: Based on the fusion weight optimization result, the short-term prediction result and the long-term prediction result are fused to obtain the fusion disturbance prediction result, which is specifically:

[0186] ΔT(t+τ)=α·ΔT ARIMA (t+τ)+(1-α)·ΔT LSTM (t+τ)

[0187] Where: ΔT(t+τ) is the prediction result of fusion perturbation, ΔT ARIMA (t+τ) is the disturbance amplitude predicted by the autoregressive integrated moving average model, ΔT LSTM(t+τ) is the perturbation amplitude predicted by the memory network, α is the dynamically adjusted fusion weight coefficient, representing the contribution of the autoregressive integrated moving average model. 1-α is the fusion contribution weight of the memory network.

[0188] Step S400.5: Verify based on the obtained fusion disturbance prediction result.

[0189] The fusion disturbance prediction results and the running physical signal are compared by the mean square error and mean square error to obtain the running physical signal error data. If the error meets the threshold requirement: MSE threshold ≤ 0.05Nm 2 , when the MAE threshold is ≤ 0.1Nm, the optimized autoregressive integrated moving average model and memory network are applied. If the error still does not meet the requirements, the iterative optimization is continued until it reaches the threshold range.

[0190] In a specific embodiment, a residual error is calculated based on the obtained fusion disturbance prediction result and the running physical signal error data to obtain residual error data, a sliding mode control surface is configured based on the obtained residual error data, a non-singular sliding mode control law is used to make the sliding mode control surface converge quickly, a sliding mode control output signal is obtained, and the sliding mode control output signal is combined with the total compensation current to generate a final compensation signal, which specifically includes:

[0191] Step S500.1: Calculate the residual error based on the obtained fusion disturbance prediction result and the running physical signal error data to obtain the residual error data, specifically:

[0192] E residual (t+τ)=ΔT(t+τ)-(ΔT)^(t+τ)

[0193] Where: E residual (t+τ) is the residual error, which represents the difference between the prediction and the actual value after compensation. ΔT(t+τ) is the actual disturbance amplitude of the running physical signal. (ΔT)^(t+τ) is the disturbance prediction result after fusion.

[0194] To E residual (t+τ) is decomposed in the frequency domain to obtain high-frequency error and low-frequency error components.

[0195] Step S500.2: Configure the sliding mode control surface based on the obtained residual error data, specifically:

[0196]

[0197] Where: S(t) is the sliding mode control surface, which is used to quantify the system error state, e(t) is the real-time residual error, and E residual (t) is equal, e(t)=E residual (t), is the rate of change of the residual error, which is used to reflect the dynamic change of the error over time. c1(t) is the dynamic weight parameter for error amplification, and c2(t) is the dynamic weight parameter for error change rate amplification.

[0198] The sliding mode gain is dynamically adjusted based on the obtained high-frequency error and low-frequency error components, specifically:

[0199]

[0200] Where: k1, k2 are reference gains, α, β are error sensitivity adjustment coefficients, is the absolute value of the low-frequency error component, is the absolute value of the high-frequency error component.

[0201] The non-singular sliding mode control law is used to make the sliding mode control surface converge quickly, and the sliding mode control output signal is obtained, which is specifically:

[0202] u(t)=-K·sgn(S(t))

[0203] Where u(t) is the sliding mode control output signal, which is used to dynamically adjust the compensation current; K is the sliding mode gain coefficient, which controls the compensation response amplitude; and sgn(S(t)) is the sliding surface sign function, which is used to indicate the control direction.

[0204] The sliding mode control output signal is combined with the total compensation current to generate the final compensation signal, which is:

[0205]

[0206] Where: I final (t) is the total control current finally injected into the motor, I total (t) is the synthesis result of feedforward and frequency domain compensation signals, is the sliding mode feedback compensation signal.

[0207] In a specific embodiment, the final compensation signal is compared with the running physical signal and the fusion disturbance prediction result to obtain the final residual error, and the frequency domain decomposition is performed to obtain the high-frequency error component and the low-frequency error component. The sliding mode control surface weight parameters and the fusion weight are dynamically adjusted based on the obtained high-frequency error component and the low-frequency error component, the final compensation signal generation logic is optimized, and the final optimized compensation signal is generated based on the dynamically adjusted sliding mode control surface weight parameters and the fusion weight. The final optimized compensation signal is injected into the motor control loop, which specifically includes:

[0208] Step S600.1: Compare the obtained final compensation signal with the running physical signal and the fusion disturbance prediction result to obtain the final residual error, which is specifically:

[0209]

[0210] Where: E residual (t) is the residual error, which represents the difference between the predicted value after compensation and the actual disturbance value, ΔT(t) is the actual disturbance amplitude in the running physical signal, is the final disturbance prediction result.

[0211] By performing frequency domain decomposition on the final residual error, high-frequency error components and low-frequency error components are obtained.

[0212] Step S600.1.1: Dynamically adjust the sliding mode control surface weight parameters and fusion weights based on the obtained high-frequency error components and low-frequency error components to optimize the final compensation signal generation logic. Specifically:

[0213]

[0214] Where: S(t) is the sliding mode control surface, which is used to quantify the residual error state, e(t) is the real-time residual error, and E residual (t) are equal, is the rate of change of the residual error, c1(t) and c2(t) are dynamically adjusted weight parameters:

[0215]

[0216] k1 and k2 are reference gains, α and β are error sensitivity coefficients, and the compensation signal ratio is adjusted according to the high-frequency error component and the low-frequency error component. Specifically:

[0217]

[0218] Where: α is the dynamic proportional coefficient of the low-frequency compensation signal, 1-α is the dynamic proportional coefficient of the high-frequency compensation signal, is the absolute value of the low-frequency error in the residual error, is the absolute value of the high-frequency error in the residual error.

[0219] Step S600.2: Generate a final optimized compensation signal based on the dynamically adjusted sliding mode control surface weight parameters and fusion weights, specifically:

[0220]

[0221] Where: I final (t) is the total control current finally injected into the motor control system. I total (t) is the fused compensation signal, including feedforward compensation and frequency domain compensation. is the sliding mode control feedback signal.

[0222] The final optimized compensation signal is injected into the motor control loop.

[0223] Obtain the running physical signal and compare it with the fusion disturbance prediction result, specifically:

[0224] E final (t) = ΔT(t) - ΔT adjusted (t)

[0225] Where: E final (t) is the final residual error, which represents the difference between the actual disturbance state and the target disturbance value after the final optimized compensation signal is injected. ΔT(t) is the target disturbance value in the running physical signal, which represents the actual disturbance amplitude of the motor operation under ideal conditions. adjusted (t) is the final optimized compensation signal I injected final The adjusted disturbance amplitude after (t) reflects the actual operating state of the motor after closed-loop control.

[0226] If E final (t) Meet the error threshold condition MSE≤0.05Nm 2 , MAE≤0.1Nm, the final optimized compensation signal is considered valid. If the error exceeds the set threshold, the sliding mode control surface weight parameters and fusion weight are adjusted cyclically, and the process stops when the error is satisfied.

[0227] In the above content, in actual application, first, the operating physical signals of the permanent magnet synchronous motor are obtained through current sensors, voltage sensors, speed sensors, and load torque sensors. The signal acquisition frequency is set to at least 1 kHz. Then, the collected signals are preprocessed using filters to remove outliers and high-frequency noise. Then, normalization is used to convert the preprocessed signals into standardized units, thereby eliminating the differences in the magnitudes of different sensors and ensuring signal consistency.

[0228] Based on the normalized signal, the local characteristics of the signal are analyzed through a sliding time window, and the time domain features, frequency domain features and change rate data are extracted. The autoregressive integrated moving average model is used to make short-term disturbance predictions, and the memory network is used to make long-term disturbance predictions. The short-term prediction results are weighted and fused with the long-term prediction results to obtain the final disturbance prediction value.

[0229] Based on the final disturbance prediction result, the feedforward compensation current signal is calculated, and the low-frequency component and high-frequency component of the feedforward compensation current signal are extracted through a low-pass filter and a band-pass filter to obtain a low-frequency compensation signal and a high-frequency compensation signal. The reference current is obtained through the motor control system of the permanent magnet synchronous motor. The obtained low-frequency compensation signal, high-frequency compensation signal and reference current are superimposed to calculate the total compensation current, which is then injected into the motor control loop for dynamic compensation adjustment.

[0230] The autoregressive integrated moving average model and memory network are optimized by compensating the error data. With the support of the compensated error data, the parameters of the autoregressive integrated moving average model are adjusted, and the memory network is optimized to perform fusion weight optimization. Based on the optimized model, the short-term and long-term disturbance prediction results are fused to calculate the final disturbance prediction result. The prediction result is compared with the operating physical signal, the residual error is calculated, and the sliding mode control surface is adjusted based on the residual error to achieve rapid convergence.

[0231] Based on the final residual error data, frequency domain decomposition is performed to extract high-frequency and low-frequency error components. The sliding mode control surface weight parameters and fusion weights are dynamically adjusted according to the obtained high-frequency and low-frequency error components. The final compensation signal generation logic is optimized. The sliding mode control is combined with the dynamically adjusted compensation signal to ensure that the motor can respond quickly and remain stable under complex load disturbances.

[0232] The final optimized compensation signal is injected into the motor control loop. The final residual error is obtained by comparing it with the actual operating physical signal and the fusion disturbance prediction results. Frequency domain analysis is performed and the compensation effect is verified based on whether the error meets the set threshold condition. If the error is lower than the set threshold, the motor system maintains the optimal operating state; if the error exceeds the threshold, the sliding mode control surface and fusion weight are continuously optimized until the error standard is met, ensuring stable operation of the motor in a complex disturbance environment.

[0233] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for controlling a permanent magnet synchronous motor against load disturbance, characterized in that: The steps include: S100, obtaining an operating physical signal of the permanent magnet synchronous motor, performing preprocessing and normalization processing, and analyzing local characteristics of the normalized operating physical signal through a sliding time window to obtain a disturbance type and disturbance degree; S200, configuring an autoregressive integrated moving average model and a memory network based on the obtained time domain features, frequency domain features, change rate, and normalized operating physical signal to obtain short-term prediction results and long-term prediction results, integrating the short-term prediction results and the long-term prediction results to obtain a final disturbance prediction result, and calculating an error based on comparing the final disturbance prediction result with the operating physical signal; S300, calculating feedforward current compensation based on the final disturbance prediction result to obtain a feedforward compensation current signal, extracting low-frequency components and high-frequency components from the feedforward compensation current signal through a low-pass filter and a band-pass filter to obtain a low-frequency compensation signal and a high-frequency compensation signal, obtaining a reference current through a motor control system of the permanent magnet synchronous motor, superimposing the obtained low-frequency compensation signal, the high-frequency compensation signal, and the reference current to obtain a total compensation current, performing dynamic compensation adjustment based on the obtained total compensation current, and performing a comparison before and after the dynamic compensation adjustment; S400, adjusting the parameters of the autoregressive integrated moving average model based on the obtained compensation error data and performing comparisons, optimizing the memory network based on the obtained compensation error data, optimizing the fusion weights based on the optimized autoregressive integrated moving average model and the memory network, fusing the short-term prediction results and the long-term prediction results based on the fusion weight optimization results, obtaining a fused disturbance prediction result, and performing verification based on the obtained fused disturbance prediction result; S500, calculating a residual error based on the obtained fused disturbance prediction result and the operating physical signal error data to obtain residual error data, configuring a sliding mode control surface based on the obtained residual error data, using a non-singular sliding mode control law to enable the sliding mode control surface to converge rapidly, obtaining a sliding mode control output signal, and combining the sliding mode control output signal with the total compensation current to generate a final compensation signal; S600. Compare the obtained final compensation signal with the running physical signal and the fusion disturbance prediction result to obtain the final residual error, and perform frequency domain decomposition to obtain the high-frequency error component and the low-frequency error component. Dynamically adjust the sliding mode control surface weight parameters and the fusion weight based on the obtained high-frequency error component and the low-frequency error component, optimize the final compensation signal generation logic, generate the final optimized compensation signal based on the dynamically adjusted sliding mode control surface weight parameters and the fusion weight, and inject the final optimized compensation signal into the motor control loop.

2. The method for controlling a permanent magnet synchronous motor against load disturbance according to claim 1, wherein: The step 100 specifically includes: S100.

1. Obtaining physical operating signals of the permanent magnet synchronous motor; Operational physical signals include: current, voltage, speed and torque; The current in the running physical signal is obtained through the current sensor, the voltage in the running physical signal is obtained through the voltage sensor, the speed in the running physical signal is obtained through the speed sensor, and the torque in the running physical signal is obtained through the load torque sensor; S100.

2. Preprocess the obtained operating physical signal through a filter; S100.2.

1. Perform normalization processing based on the preprocessed operating physical signal; S100.

3. Analyze local characteristics using a sliding time window based on the normalized operational physical signal. The time window is set to be greater than 0ms, the sliding window is set to be greater than 0ms, and the normalized running physical signal is divided by greater than 0ms through the time window to obtain more than 0 time data points. The normalized running physical signal is divided by greater than 0ms through the sliding window to obtain more than 0 sliding data points; Perform feature extraction based on the obtained time data points and sliding data points; Feature extraction calculates the mean and variance of the current, voltage, speed, and torque of the time data points and the sliding data points through mean calculation and variance calculation to obtain time domain features. The time data points and the sliding data points are analyzed in the frequency domain through fast Fourier transform to extract frequency domain features. The frequency domain features are identified through the spectrum diagram to identify the high-frequency disturbance and low-frequency fluctuation components. The instantaneous rate of change of the time data points and the sliding data points is calculated to obtain the rate of change data. Based on the time domain features, frequency domain features and change rate data obtained by feature extraction, a machine learning algorithm is used to classify the disturbance; The categories include: load mutation, inertial disturbance and periodic disturbance; According to the obtained time domain characteristics, frequency domain characteristics and change rate data, the amplitude and frequency of the disturbance are judged, and the disturbance amplitude and disturbance frequency are obtained.

3. The method for controlling a permanent magnet synchronous motor against load disturbance according to claim 1, wherein: The step 200 is specifically as follows: S200.

1. Based on the obtained time-domain characteristics, frequency-domain characteristics, rate of change, and normalized operational physical signals, configure an autoregressive integrated moving average model to obtain short-term prediction results. S200.

2. Based on the obtained short-term prediction results, time domain characteristics, frequency domain characteristics, and change rates, a memory network is configured to obtain a long-term prediction result. S200.

3. Based on the obtained short-term prediction results and long-term prediction results, weighted fusion is used to integrate them to obtain the final disturbance prediction result; S200.

4. Calculate the error based on the comparison between the final disturbance prediction result and the operating physical signal.

4. The method for controlling a permanent magnet synchronous motor against load disturbance according to claim 1, wherein: The step 300 is specifically as follows: S300.

1. Calculate feedforward current compensation based on the final disturbance prediction result to obtain a feedforward compensation current signal; S300.1.

1. Extract the low-frequency component from the feedforward compensation current signal through a low-pass filter to obtain a low-frequency compensation signal; S300.1.

2. Extract the high-frequency component from the feedforward compensation current signal through a bandpass filter to obtain a high-frequency compensation signal; S300.1.

3. Obtain a reference current for a current control loop from the motor control system of the permanent magnet synchronous motor; S300.

2. Superimpose the obtained low-frequency compensation signal, high-frequency compensation signal, and reference current to obtain a total compensation current; Based on the obtained total compensation current, the current controller is injected into the permanent magnet synchronous motor control loop to dynamically compensate and adjust the input current of the permanent magnet synchronous motor; S300.

3. Compare the results based on dynamic compensation adjustment; Compensation error data is obtained by calculating the errors of disturbance amplitude and disturbance frequency before and after dynamic compensation adjustment.

5. The method for controlling a permanent magnet synchronous motor against load disturbance according to claim 1, wherein: The step 400 is specifically as follows: S400.

1. Optimize the parameters p, d, and q of the autoregressive integrated moving average model based on the obtained compensation error data; The grid search method is used to find the optimal parameter combination p, the order of the autoregressive term d, the order of the difference term q, and the order of the moving average term. After obtaining the optimal parameter combination, the residual analysis is performed. S400.

2. Comparison of the parameters p, d, and q based on the optimized autoregressive integrated moving average model; S400.

3. Optimize the memory network based on the obtained compensation error data; S400.

4. Optimize fusion weights based on the optimized autoregressive integrated moving average model and memory network; S400.4.

1. Based on the fusion weight optimization results, the short-term prediction results and the long-term prediction results are fused to obtain the fusion disturbance prediction results; S400.

5. Verify based on the obtained fusion perturbation prediction result; The fusion disturbance prediction results and the running physical signal are compared by the mean square error and mean square error to obtain the running physical signal error data. If the error meets the threshold requirement: MSE threshold ≤ 0.05Nm 2 , when the MAE threshold is ≤ 0.1Nm, the optimized autoregressive integrated moving average model and memory network are applied. If the error still does not meet the requirements, the iterative optimization is continued until it reaches the threshold range.

6. The method for controlling a permanent magnet synchronous motor against load disturbance according to claim 1, wherein: The step 500 specifically includes: S500.

1. Calculate the residual error based on the obtained fused disturbance prediction result and the running physical signal error data to obtain the residual error data, specifically: E residual (t+τ)=ΔT(t+τ)-(ΔT)^(t+τ) Where: E residual (t+τ) is the residual error, which represents the difference between the prediction and the actual value after compensation. ΔT(t+τ) is the actual disturbance amplitude of the running physical signal. (ΔT)^(t+τ) is the disturbance prediction result after fusion. To E residual (t+τ) performs frequency domain decomposition to obtain high-frequency error and low-frequency error components; S500.

2. Configure a sliding mode control surface based on the obtained residual error data; Dynamically adjust the sliding mode gain based on the obtained high-frequency error and low-frequency error components; The non-singular sliding mode control law is used to make the sliding mode control surface converge quickly and obtain the sliding mode control output signal; The sliding mode control output signal is combined with the total compensation current to generate the final compensation signal, which is: Where: I final (t) is the total control current finally injected into the motor, I total (t) is the synthesis result of feedforward and frequency domain compensation signals, is the sliding mode feedback compensation signal.

7. The method for controlling a permanent magnet synchronous motor against load disturbance according to claim 1, wherein: The step 600 specifically includes: S600.

1. Compare the final compensation signal with the running physical signal and the fusion disturbance prediction result to obtain the final residual error, which is: Where: E residual (t) is the residual error, which represents the difference between the predicted value after compensation and the actual disturbance value, ΔT(t) is the actual disturbance amplitude in the running physical signal, is the final disturbance prediction result; By performing frequency domain decomposition on the final residual error, high-frequency error components and low-frequency error components are obtained; S600.1.

1. Dynamically adjust the sliding mode control surface weight parameters and fusion weights based on the obtained high-frequency error components and low-frequency error components to optimize the final compensation signal generation logic; S600.

2. Generate a final optimized compensation signal based on the dynamically adjusted sliding mode control surface weight parameters and fusion weights; The final optimized compensation signal is injected into the motor control loop.

Images