Permanent magnet motor starting control method and system based on harmonic injection

Abstract

The invention discloses a permanent magnet motor starting control method and system based on harmonic injection, and aims to improve the harmonic injection precision and the dynamic adaptive capacity in the motor starting process. According to the method, frequency domain features are extracted by adopting CNN of a multi-scale convolution kernel, dynamic changes of a time sequence are captured in combination with Bi-LSTM, and an attention mechanism is introduced to dynamically weight key transient signals, so that comprehensive modeling and optimization of harmonic signals on frequency domain, time domain and transient features are realized; and an efficient closed-loop dynamic optimization system is constructed by further combining the global search capability and the real-time error feedback mechanism of the genetic algorithm. The method can adaptively capture the key change when the working condition suddenly changes, correct the harmonic injection parameters in real time, significantly reduce the power loss caused by torque fluctuation and high-frequency harmonic in millisecond response, and improve the stability of harmonic injection and the energy efficiency optimization effect.

Description

Technical Field

[0001] The present invention relates to the technical field of motor control and signal optimization, and in particular to a permanent magnet motor starting control method and system based on harmonic injection. Background Art

[0002] Harmonic signal optimization technology plays a key role in the startup and operation of permanent magnet motors. Its main purpose is to reduce torque fluctuations, power losses and high-frequency interference by controlling the injection of harmonic signals, thereby improving the overall performance of the motor. During the motor startup process, due to factors such as load changes, instability of starting torque and complex dynamic conditions, the precise injection and real-time optimization of harmonic signals have become technical problems that need to be solved in the industry.

[0003] In the prior art, harmonic signal optimization usually adopts a control method based on a fixed model, such as classic PI control, multi-rotation PI control, or harmonic analysis technology based on fast Fourier transform (FFT). These methods adjust the harmonic injection signal through pre-set fixed parameters to suppress the motor torque fluctuation. However, these traditional methods have obvious limitations when dealing with complex dynamic conditions, which are mainly manifested in the following points:

[0004] 1. Lack of real-time performance: Traditional control methods rely on fixed parameters or offline optimization models, which make it difficult to timely capture the dynamic characteristics of signals caused by sudden changes in operating conditions or load changes. The response time of optimization and adjustment is long, making it difficult to meet the real-time requirements in actual operating conditions.

[0005] 2. Poor accuracy: Although methods such as FFT can extract the frequency domain characteristics of harmonic signals, they are insufficient in capturing transient signals or dynamic changes in timing, resulting in a lack of accurate modeling of dynamic signals in the optimization results, which is prone to increased power loss caused by high-frequency interference.

[0006] 3. Poor adaptability to working conditions: Traditional methods are mostly based on static optimization models, which show insufficient adaptability under complex working conditions (such as startup transients or scenarios with drastic load changes) and cannot dynamically adjust the injection amplitude and frequency of harmonic signals, resulting in unstable motor performance.

[0007] In order to solve the above problems, the industry usually introduces optimization technologies such as deep learning and genetic algorithms to enhance optimization capabilities. However, the single application of these technologies still has new challenges. For example, although deep learning has strong nonlinear modeling capabilities, it still faces problems such as lagging prediction results and insufficient dynamic adjustment when dealing with scenarios with high requirements for real-time and working condition adaptability; genetic algorithms perform well in global optimization, but their convergence speed is relatively slow, and it is difficult to meet the millisecond-level response requirements during motor startup when used alone; and although simple feedback mechanisms can improve the flexibility of dynamic adjustment to a certain extent, conventional feedback mechanisms are mostly linear adjustments, and their ability to process nonlinear signal characteristics in complex dynamic conditions is limited.

[0008] Therefore, how to achieve a deep combination of real-time performance, accuracy and dynamic adaptability in harmonic signal optimization has become a technical problem to be solved by the present invention. Summary of the invention

[0009] The technical problem solved by the present invention is to provide a permanent magnet motor starting control method and system based on harmonic injection in response to the defects existing in the above-mentioned prior art, so as to solve the problem raised in the above-mentioned background technology that it is impossible to take into account real-time performance, accuracy and adaptability to dynamic working conditions at the same time when solving the harmonic signal optimization problem.

[0010] In order to solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0011] A permanent magnet motor starting control method based on harmonic injection comprises the following steps:

[0012] Step 1, harmonic parameter extraction: During the starting process of the permanent magnet motor, the real-time starting current signal is collected, and the current signal is decomposed in real time using an adaptive filter based on the least mean square algorithm to extract the harmonic frequency, amplitude and phase information in the current signal to generate a harmonic parameter set;

[0013] The harmonic parameter set includes frequency characteristics, amplitude and phase information collected in real time, wherein the frequency characteristics will be used as one of the input features of the deep learning model and used in subsequent optimization steps;

[0014] Step 2, multi-frequency harmonic dynamic optimization injection: According to the harmonic parameter set, the amplitude and phase of the harmonic injection signal are optimized and adjusted using a genetic algorithm, wherein:

[0015] The optimization goal is to minimize the torque fluctuation in the low-speed area during the motor startup phase and reduce the energy loss caused by the harmonic injection signal;

[0016] The optimization objective function is defined as:

[0017]

[0018] Among them, M actual,i is the actual torque after the i-th harmonic injection, M target,i is the set target torque, P loss,i is the power loss caused by harmonic injection, w i To optimize the weight coefficient;

[0019] The weight coefficient w i The calculation formula is:

[0020]

[0021] If ΔM actual，i =0 or ΔM actual，i =0, then the weight coefficient w i Set to the default value of 0.1 to avoid calculation anomalies;

[0022] The result of the dynamic optimization generates an adjusted harmonic injection signal, which includes an optimized amplitude, frequency and phase, and is used as an input for subsequent steps;

[0023] Step 3, optimization of spatial distribution of harmonic injection: according to the harmonic injection signal generated in step 2, the spatial distribution of the signal is optimized, wherein:

[0024] Divide the motor stator winding into multiple spatial injection areas:

[0025] The area close to the load end is defined as the high-frequency priority injection area, into which high-frequency harmonic signals are preferentially injected, with a frequency range of 1 kHz to 10 kHz to enhance the transient electromagnetic response capability;

[0026] The area far away from the load end is defined as the low-frequency priority injection area, into which low-frequency harmonic signals are preferentially injected, with a frequency range of 100 Hz to 1 kHz to reduce high-frequency electromagnetic interference;

[0027] The goals of spatial distribution optimization include: the mean square error of magnetic field intensity distribution is less than 10A / m 2 ; The high-frequency harmonic injection amount does not exceed 30% of the total winding injection amount;

[0028] When optimizing the spatial distribution of harmonic signals, the root mean square value RMS of the winding current is calculated, and the RMS value is calculated according to the following formula:

[0029]

[0030] Where i(t) is the function of the winding current changing with time, T is the period of the current signal; the RMS value is used to evaluate the stability of the energy distribution injected by the harmonic signal;

[0031] Step 4, harmonic injection based on time dynamic phase delay: by real-time monitoring of the torque sensor output or based on the dynamic analysis of the current signal, the influence of the harmonic injection signal on the transient torque change is obtained, and the phase delay angle of the harmonic injection signal is dynamically adjusted using the transient torque tracking control algorithm;

[0032] The phase delay angle is calculated based on the following formula:

[0033]

[0034] Among them, θ optimal is the initial phase angle; ΔT is the transient torque change, Δt is the time interval, K is the dynamic adjustment coefficient, and the specific adjustment rules are as follows:

[0035] When ΔT / Δt≤50rad / s 2 ,K=0.1;

[0036] When 50rad / s2<ΔT / Δt≤200rad / s 2 ,K=0.5;

[0037] When ΔT / Δt>200rad / s 2 ,K=1.0;

[0038] The total adjustment angle of dynamic phase delay shall not exceed 180 rad / s 2 To avoid winding magnetic field mismatch or reverse harmonic interference;

[0039] Step 5: Intelligent predictive control of harmonics: Combined with the deep learning model, intelligent predictive adjustment of harmonic injection signals is realized, where:

[0040] The deep learning model includes:

[0041] Convolutional neural network (CNN) extracts the frequency characteristics of real-time current signals, including high-frequency and low-frequency harmonic components;

[0042] The recurrent neural network (RNN) combines real-time features and historical harmonic data through the long short-term memory (LSTM) network to predict the amplitude, frequency, and phase of the harmonic injection signal at the next moment.

[0043] The input features of the deep learning model further include ambient temperature signals and torque change rates collected in real time to improve the model's prediction ability for complex working conditions;

[0044] The inference time of the deep learning model is controlled within 1 millisecond, meeting the real-time requirement of the harmonic signal update period of 5 milliseconds;

[0045] Step 6, safety protection mechanism: During the harmonic injection process, safety protection is achieved through the following steps:

[0046] The thermistor sensor samples the stator winding temperature in real time with a period of 10 milliseconds, and the temperature threshold is no more than 125 degrees Celsius;

[0047] The Hall effect sensor array collects the magnetic field intensity distribution in the winding in real time at a frequency of 100 times per second, and the optimization goal is to control the magnetic field intensity deviation within 5A / m;

[0048] When the winding temperature or magnetic field strength exceeds the threshold for 10 consecutive sampling cycles, the emergency protection mode is triggered, the harmonic signal injection is stopped and the motor speed is gradually reduced to a safe range.

[0049] As a further solution of the present invention, the fitness function of the genetic algorithm is further defined as follows:

[0050] The square error of the motor starting torque fluctuation:

[0051]

[0052] Actual power loss caused by harmonic injection signal:

[0053]

[0054] Mean square error of magnetic field intensity distribution:

[0055]

[0056] Among them, B j is the magnetic field strength at the jth point in the winding, is the average value of the magnetic field strength.

[0057] As a further solution of the present invention, the optimizer of the deep learning model adopts Adam optimizer with a learning rate of 0.001. During the model training process, the prediction error target of frequency features and time series data is controlled within 5%.

[0058] As a further solution of the present invention, the dynamic optimization model meets the following constraints when performing optimization calculations: the stator winding temperature shall not exceed 125 degrees Celsius; the magnetic field strength deviation is controlled within 5%; and the optimization calculation time shall not exceed 10% of the dynamic adjustment period.

[0059] As a further solution of the present invention, the spatial distribution optimization of harmonic injection in step 3 also includes the following objectives: in the high-frequency priority injection area, the injection power of the harmonic signal accounts for no more than 40% of the total injection power to avoid eddy current losses caused by excessive high-frequency signals; in the low-frequency priority injection area, the root mean square value RMS of the winding current after the harmonic signal is injected is maintained between 90% and 110% of the rated value to ensure the stability of the winding energy distribution.

[0060] As a further scheme of the present invention, the dynamic adjustment rules of the fitness function of the genetic algorithm include: in the initial stage of starting, the weight ratio of the torque fluctuation error is preferentially set to 60%-80% to reduce the starting oscillation in the low-speed zone; in the middle and late stages of starting, the weight ratio of power loss is dynamically adjusted to 40%-60% to optimize the motor starting efficiency; the weight ratio of the magnetic field uniformity deviation is always kept within 20% to avoid the impact of excessive weight on other optimization objectives.

[0061] As a further solution of the present invention, the input features of the deep learning model further include: a torque change rate signal with a sampling frequency of 500 Hz; a real-time waveform of the winding current caused by harmonic injection with a sampling frequency of 1 kHz; and an ambient temperature signal during the startup phase with a sampling period of 10 milliseconds.

[0062] As a further solution of the present invention, the training process of the deep learning model includes the following enhancement mechanism:

[0063] Data enhancement: By randomly perturbing the historical harmonic signals, an extended data set of no less than 1.5 times the original data volume is generated to improve the robustness of the model;

[0064] Dynamic adjustment of model training accuracy: When the training error convergence rate is lower than 1% / epoch, the learning rate is automatically reduced to 50% of the original learning rate;

[0065] Overfitting protection: During the training process, the model performance is monitored in real time through cross-validation. When the error of the validation set is greater than twice the error of the training set, the model early stopping mechanism is triggered.

[0066] As a further solution of the present invention, the security protection mechanism further includes:

[0067] When the winding temperature is close to 125 degrees Celsius, such as 120 ± 2 degrees Celsius, it is preferred to reduce the injection amplitude of the high-frequency harmonic signal to within 50% of the original amplitude to reduce the thermal stress of the winding;

[0068] When the magnetic field intensity deviation is close to 10%, such as 8%-10%, the injection ratio of the low-frequency harmonic signal is increased in real time to balance the magnetic field distribution in the winding;

[0069] If the temperature or magnetic field strength exceeds the threshold and cannot be restored, the emergency protection mode is triggered, the harmonic signal injection is stopped and the motor speed is gradually reduced to a safe range.

[0070] As a further solution of the present invention, the dynamic phase adjustment in step 4 further includes the following rules:

[0071] When the amplitude of the harmonic injection signal exceeds 20% of the rated amplitude, the adjustment rate of the dynamic phase delay angle is limited to within 1° / ms to avoid motor oscillation caused by phase mutation;

[0072] When the transient torque change rate is greater than 200 rad / s 2 When , the dynamic adjustment coefficient K can be dynamically increased to 1.5 to improve the sensitivity of phase adjustment;

[0073] During the dynamic phase delay adjustment process, the phase angle change after each adjustment shall not exceed 30° to ensure the stability of the control system.

[0074] The steps 2 and 5 are combined with a real-time feedback mechanism to further optimize the harmonic injection signal through the following steps.

[0075] Compared with the prior art, the present invention has the following beneficial effects:

[0076] 1. The frequency domain features are extracted by CNN with multi-scale convolution kernels, and the dynamic changes of time series are captured by Bi-LSTM. The attention mechanism is introduced to dynamically weight the key transient signals, thus realizing the comprehensive modeling and optimization of harmonic signals in frequency domain, time domain and transient features. Different from the traditional separation feature processing method, the present invention can accurately capture the dynamic change law and key points of harmonic signals under complex working conditions, which not only improves the prediction accuracy, but also greatly improves the system's response flexibility and adaptability to complex dynamic signals.

[0077] 2. Through the synergistic mechanism of deep learning prediction and genetic algorithm optimization, combined with the real-time error feedback system, a real-time and efficient closed-loop dynamic optimization system is formed. In actual working conditions, the injection amplitude and frequency of harmonic signals are optimized through the millisecond-level dynamic adjustment mechanism. Compared with the optimization strategy of traditional fixed models, it can quickly adapt to the changes in complex dynamic conditions during the startup process. The impact of torque fluctuations and high-frequency interference on system performance during motor startup is significantly reduced, and the stability of harmonic signals and the simultaneous optimization of energy efficiency are achieved.

[0078] 3. Through the multi-layer synergy of the algorithm architecture, when processing complex dynamic signals, the dynamic adjustment strategy can adaptively capture changes in key operating conditions, correct harmonic signal injection parameters in real time, and effectively reduce power loss and torque fluctuations caused by high-frequency harmonics. At the same time, the real-time closed-loop design based on the feedback optimization mechanism enables the system to respond quickly and maintain optimized high precision and stability when the operating conditions change suddenly, significantly improving the dynamic adaptability and operating efficiency of the motor system in complex environments. Compared with existing technologies, it demonstrates higher real-time, accuracy and stability.

[0079] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0081] Figure 1 This is a system framework diagram of the invention.

[0082] Figure 2 It is a dynamic optimization flow chart of the present invention.

[0083] Figure 3 It is a schematic diagram of optimizing the spatial distribution of harmonic signals of the present invention.

[0084] Figure 4 It is the experimental result and comparative data of the present invention. DETAILED DESCRIPTION

[0085] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0086] See also Figure 1 — Figure 4 In an embodiment of the present invention, a permanent magnet motor starting control method based on harmonic injection includes the following steps:

[0087] Step 1, harmonic parameter extraction: During the starting process of the permanent magnet motor, the real-time starting current signal is collected, and the current signal is decomposed in real time using an adaptive filter based on the least mean square algorithm to extract the harmonic frequency, amplitude and phase information in the current signal to generate a harmonic parameter set;

[0088] The harmonic parameter set includes frequency characteristics, amplitude and phase information collected in real time, wherein the frequency characteristics will be used as one of the input features of the deep learning model and used in subsequent optimization steps;

[0089] Step 2, multi-frequency harmonic dynamic optimization injection: According to the harmonic parameter set, the amplitude and phase of the harmonic injection signal are optimized and adjusted using a genetic algorithm, wherein:

[0090] The optimization goal is to minimize the torque fluctuation in the low-speed area during the motor startup phase and reduce the energy loss caused by the harmonic injection signal;

[0091] The optimization objective function is defined as:

[0092]

[0093] Among them, M actual,i is the actual torque after the i-th harmonic injection, M target,i is the set target torque, P loss,i is the power loss caused by harmonic injection, w i To optimize the weight coefficient;

[0094] The weight coefficient w i The calculation formula is:

[0095]

[0096] If ΔM actual，i =0 or ΔM actual，i =0, then the weight coefficient w i Set to the default value of 0.1 to avoid calculation anomalies;

[0097] The result of the dynamic optimization generates an adjusted harmonic injection signal, which includes an optimized amplitude, frequency and phase, and is used as an input for subsequent steps;

[0098] Step 3, optimization of spatial distribution of harmonic injection: according to the harmonic injection signal generated in step 2, the spatial distribution of the signal is optimized, wherein:

[0099] Divide the motor stator winding into multiple spatial injection areas:

[0100] The area close to the load end is defined as the high-frequency priority injection area, into which high-frequency harmonic signals are preferentially injected, with a frequency range of 1 kHz to 10 kHz to enhance the transient electromagnetic response capability;

[0101] The area far away from the load end is defined as the low-frequency priority injection area, into which low-frequency harmonic signals are preferentially injected, with a frequency range of 100 Hz to 1 kHz to reduce high-frequency electromagnetic interference;

[0102] The goals of spatial distribution optimization include: the mean square error of magnetic field intensity distribution is less than 10A / m 2 ; The high-frequency harmonic injection amount does not exceed 30% of the total winding injection amount;

[0103] When optimizing the spatial distribution of harmonic signals, the root mean square value RMS of the winding current is calculated, and the RMS value is calculated according to the following formula:

[0104]

[0105] Where i(t) is the function of the winding current changing with time, T is the period of the current signal; the RMS value is used to evaluate the stability of the energy distribution injected by the harmonic signal;

[0106] Step 4, harmonic injection based on time dynamic phase delay: by real-time monitoring of the torque sensor output or based on the dynamic analysis of the current signal, the influence of the harmonic injection signal on the transient torque change is obtained, and the phase delay angle of the harmonic injection signal is dynamically adjusted using the transient torque tracking control algorithm;

[0107] The phase delay angle is calculated based on the following formula:

[0108]

[0109] Among them, θ optimal is the initial phase angle; ΔT is the transient torque change, Δt is the time interval, K is the dynamic adjustment coefficient, and the specific adjustment rules are as follows:

[0110] When ΔT / Δt≤50rad / s 2 ,K=0.1;

[0111] When 50rad / s2<ΔT / Δt≤200rad / s 2 ,K=0.5;

[0112] When ΔT / Δt>200rad / s 2 ,K=1.0;

[0113] The total adjustment angle of dynamic phase delay shall not exceed 180 rad / s 2 To avoid winding magnetic field mismatch or reverse harmonic interference;

[0114] Step 5: Intelligent predictive control of harmonics: Combined with the deep learning model, intelligent predictive adjustment of harmonic injection signals is realized, where:

[0115] The deep learning model includes:

[0116] Convolutional neural network (CNN) extracts the frequency characteristics of real-time current signals, including high-frequency and low-frequency harmonic components;

[0117] The recurrent neural network (RNN) combines real-time features and historical harmonic data through the long short-term memory (LSTM) network to predict the amplitude, frequency, and phase of the harmonic injection signal at the next moment.

[0118] The input features of the deep learning model further include ambient temperature signals and torque change rates collected in real time to improve the model's prediction ability for complex working conditions;

[0119] The inference time of the deep learning model is controlled within 1 millisecond, meeting the real-time requirement of the harmonic signal update period of 5 milliseconds;

[0120] Step 6, safety protection mechanism: During the harmonic injection process, safety protection is achieved through the following steps:

[0121] The thermistor sensor samples the stator winding temperature in real time with a period of 10 milliseconds, and the temperature threshold is no more than 125 degrees Celsius;

[0122] The Hall effect sensor array collects the magnetic field intensity distribution in the winding in real time at a frequency of 100 times per second, and the optimization goal is to control the magnetic field intensity deviation within 5A / m;

[0123] When the winding temperature or magnetic field strength exceeds the threshold for 10 consecutive sampling cycles, the emergency protection mode is triggered, the harmonic signal injection is stopped and the motor speed is gradually reduced to a safe range.

[0124] As a further solution of the present invention, the fitness function of the genetic algorithm is further defined as follows:

[0125] The square error of the motor starting torque fluctuation:

[0126]

[0127] Actual power loss caused by harmonic injection signal:

[0128]

[0129] Mean square error of magnetic field intensity distribution:

[0130]

[0131] Among them, B j is the magnetic field strength at the jth point in the winding, is the average value of the magnetic field strength.

[0132] As a further solution of the present invention, the optimizer of the deep learning model adopts Adam optimizer with a learning rate of 0.001. During the model training process, the prediction error target of frequency features and time series data is controlled within 5%.

[0133] As a further solution of the present invention, the dynamic optimization model meets the following constraints when performing optimization calculations: the stator winding temperature shall not exceed 125 degrees Celsius; the magnetic field strength deviation is controlled within 5%; and the optimization calculation time shall not exceed 10% of the dynamic adjustment period.

[0134] As a further solution of the present invention, the spatial distribution optimization of harmonic injection in step 3 also includes the following objectives: in the high-frequency priority injection area, the injection power of the harmonic signal accounts for no more than 40% of the total injection power to avoid eddy current losses caused by excessive high-frequency signals; in the low-frequency priority injection area, the root mean square value RMS of the winding current after the harmonic signal is injected is maintained between 90% and 110% of the rated value to ensure the stability of the winding energy distribution.

[0135] As a further scheme of the present invention, the dynamic adjustment rules of the fitness function of the genetic algorithm include: in the initial stage of starting, the weight ratio of the torque fluctuation error is preferentially set to 60%-80% to reduce the starting oscillation in the low-speed zone; in the middle and late stages of starting, the weight ratio of power loss is dynamically adjusted to 40%-60% to optimize the motor starting efficiency; the weight ratio of the magnetic field uniformity deviation is always kept within 20% to avoid the impact of excessive weight on other optimization objectives.

[0136] As a further solution of the present invention, the input features of the deep learning model further include: a torque change rate signal with a sampling frequency of 500 Hz; a real-time waveform of the winding current caused by harmonic injection with a sampling frequency of 1 kHz; and an ambient temperature signal during the startup phase with a sampling period of 10 milliseconds.

[0137] As a further solution of the present invention, the training process of the deep learning model includes the following enhancement mechanism:

[0138] Data enhancement: By randomly perturbing the historical harmonic signals, an extended data set of no less than 1.5 times the original data volume is generated to improve the robustness of the model;

[0139] Dynamic adjustment of model training accuracy: When the training error convergence rate is lower than 1% / epoch, the learning rate is automatically reduced to 50% of the original learning rate;

[0140] Overfitting protection: During the training process, the model performance is monitored in real time through cross-validation. When the error of the validation set is greater than twice the error of the training set, the model early stopping mechanism is triggered.

[0141] As a further solution of the present invention, the security protection mechanism further includes:

[0142] When the winding temperature is close to 125 degrees Celsius, such as 120 ± 2 degrees Celsius, it is preferred to reduce the injection amplitude of the high-frequency harmonic signal to within 50% of the original amplitude to reduce the thermal stress of the winding;

[0143] When the magnetic field intensity deviation is close to 10%, such as 8%-10%, the injection ratio of the low-frequency harmonic signal is increased in real time to balance the magnetic field distribution in the winding;

[0144] If the temperature or magnetic field strength exceeds the threshold and cannot be restored, the emergency protection mode is triggered, the harmonic signal injection is stopped and the motor speed is gradually reduced to a safe range.

[0145] As a further solution of the present invention, the dynamic phase adjustment in step 4 further includes the following rules:

[0146] When the amplitude of the harmonic injection signal exceeds 20% of the rated amplitude, the adjustment rate of the dynamic phase delay angle is limited to within 1° / ms to avoid motor oscillation caused by phase mutation;

[0147] When the transient torque change rate is greater than 200 rad / s 2 When , the dynamic adjustment coefficient K can be dynamically increased to 1.5 to improve the sensitivity of phase adjustment;

[0148] During the dynamic phase delay adjustment process, the phase angle change after each adjustment shall not exceed 30° to ensure the stability of the control system.

[0149] The steps 2 and 5 are combined with a real-time feedback mechanism to further optimize the harmonic injection signal through the following steps.

[0150] Embodiment 1:

[0151] Combined with the actual starting process of the permanent magnet motor, the specific implementation steps are explained:

[0152] During the startup phase of the permanent magnet motor, the three-phase current signal of the stator winding is collected by a real-time current sensor, and the sampling frequency is set to 10 kHz. The current signal is decomposed using an adaptive filter based on the least mean square algorithm to extract the harmonic frequency, amplitude and phase information, and generate a harmonic parameter set. The extracted harmonic frequency range is 100 Hz to 10 kHz, the amplitude range is 0 to 50% of the rated current, and the initial phase range is 0 to 360 degrees. The frequency features in the harmonic parameter set are also used as input data and provided to the subsequent deep learning model to predict the real-time change trend of the harmonic signal at different startup stages.

[0153] The amplitude and phase of the harmonic injection signal are dynamically optimized using a genetic algorithm. The initial population size is set to 50, and each generation of the population generates the next generation through selection, crossover and mutation operations, with an iteration number of 100. The fitness function adopts the optimization objective function in claim 1, and the balance of the optimization effect is improved by adding the mean square error index of the magnetic field intensity distribution to the constraint condition. During the optimization process, if the torque fluctuation error is found to be less than 0.5 and the power loss is less than 2, the algorithm stops iterating and outputs the current optimal solution.

[0154] When the optimized harmonic signal is injected into the stator winding of the motor, the spatial distribution of the harmonic signal is optimized according to the regional division principle of high frequency priority and low frequency priority. In the winding area close to the load end, a high-frequency harmonic signal with a frequency range of 5 kHz to 10 kHz is injected to enhance the transient response capability of the motor; in the winding area far away from the load end, a low-frequency harmonic signal with a frequency range of 100 Hz to 1 kHz is injected to reduce the electromagnetic interference caused by the high-frequency signal. The root mean square value RMS of the injected signal is dynamically monitored, and its calculation formula is:

[0155]

[0156] Where N is the total number of sampling points, which depends on the sampling frequency and signal duration. For example, if the sampling frequency is 10kHz and the signal duration is 1 second, then N = 10000. k is the current value of the kth sampling point, in ampere (A). RMS : RMS current, the unit of the calculation result is ampere (A); and the sampling frequency of the current signal should be higher than twice the harmonic frequency to meet the Nyquist sampling theorem. For example, when the highest harmonic frequency is 5kHz, the sampling frequency should be at least 10kHz.

[0157] Based on the harmonic signal after spatial distribution optimization, the phase delay angle of the harmonic signal is dynamically adjusted using the real-time monitored torque change rate. When the transient torque change rate is low, the dynamic adjustment coefficient k is set to 0.1 to ensure the stability of the phase adjustment; when the transient torque change rate is high, the dynamic adjustment coefficient k is dynamically increased to 1.0 to enhance the sensitivity of the adjustment. In order to avoid phase jumps during the adjustment process, an additional maximum phase adjustment rate limit is set, and the adjustment rate does not exceed 1 degree per millisecond.

[0158] A deep learning model is used to predict and adjust the harmonic injection signal. The model structure is as follows: the convolutional neural network (CNN) is used to extract the harmonic frequency characteristics collected in real time, and the time series characteristics of high-frequency and low-frequency harmonics are captured respectively through multi-scale convolution kernels; the long short-term memory network (LSTM) in the recurrent neural network (RNN) is used to predict the amplitude, frequency and phase of the harmonic signal.

[0159] Model input: real-time extracted harmonic frequency characteristics; ambient temperature signal during motor operation, with a sampling period of 10 milliseconds; torque change rate, with a sampling frequency of 500 Hz.

[0160] Model training: About 1,000 sets of historical harmonic signal samples are used to train the model, and the prediction error in the test phase is controlled within 5. Model output: Predict the parameters of the harmonic injection signal at the next moment, which is used to correct the dynamic optimization results and improve the real-time performance of harmonic injection.

[0161] During the entire harmonic injection process, the stator winding temperature and magnetic field strength distribution are monitored in real time. The thermistor sensor samples the temperature signal with a sampling period of 10 milliseconds and a temperature threshold set to 125 degrees Celsius; the Hall effect sensor collects the magnetic field strength distribution signal with a target magnetic field strength deviation of no more than 5. If the temperature or magnetic field strength exceeds the set range for 10 consecutive sampling periods, the safety protection mode is triggered. In the safety protection mode, the harmonic signal injection is stopped and the motor speed is gradually reduced to a safe range. After the protection mode is triggered, the harmonic injection signal parameters generated by the last optimization are recorded for subsequent analysis and recovery.

[0162] Embodiment 2:

[0163] In the startup phase of the permanent magnet motor, the present invention collects the three-phase current signal of the stator winding through a real-time current sensor with a sampling frequency of 10kHz. The current signal is decomposed by an adaptive filter based on the least mean square algorithm to extract the harmonic frequency, amplitude and phase information and generate a harmonic parameter set.

[0164] The extracted harmonic frequency range is 100Hz to 10kHz, the amplitude range is 0 to 50% of the rated current, and the initial phase range is 0 to 360°. The frequency and amplitude features in the parameter set are directly used as inputs to the deep learning model for subsequent harmonic prediction.

[0165] In terms of genetic algorithm optimization, the amplitude and phase of the harmonic injection signal are dynamically adjusted in combination with the genetic algorithm.

[0166] Objective function: It includes three parts: square error of torque fluctuation, harmonic injection power loss and mean square error of magnetic field intensity distribution. Optimization conditions: The initial population size is 50, the number of iterations of the genetic algorithm is 100 generations, and the roulette selection mechanism, single-point crossover (crossover probability 0.8) and uniform mutation (mutation probability 0.1) are used. The output standard of the optimization result is: when the square error of torque fluctuation is less than 0.5N·m 2 , power loss is less than 2W, and the mean square error of magnetic field intensity distribution is less than 10A / m 2 When , the algorithm stops iterating and outputs the optimal solution.

[0167] In terms of harmonic spatial distribution optimization, the high-frequency priority area and the low-frequency priority area are divided according to the distance between the winding and the load end, and the spatial distribution of the harmonic injection signal is optimized: High-frequency priority injection area: close to the load end (within 5cm), the injection frequency range is 5-10kHz. Low-frequency priority injection area: far away from the load end (more than 5cm), the injection frequency range is 100Hz to 1kHz. For the optimized injection signal, the root mean square value (RMS) of the winding current is monitored in real time to ensure that its value is within 90%-110% of the rated value.

[0168] In terms of dynamic phase adjustment, the phase delay of the harmonic injection signal is dynamically adjusted according to the torque change rate: when the torque change rate is lower than 200rad / s 2 When the adjustment coefficient is set to 0.1; when it is higher than 200rad / s 2 When the phase adjustment is set to 1, the adjustment coefficient is set to 1.0. The phase adjustment amplitude is limited to 1° / ms, and the total phase adjustment angle does not exceed 30° to ensure system stability.

[0169] In terms of deep learning model prediction optimization, the deep learning model is used to predict and adjust the harmonic signal: input features: including harmonic frequency (Hz), amplitude (A), ambient temperature (℃) and torque change rate (rad / s 2 ). Model design: Convolutional neural network (CNN) extracts frequency and amplitude features collected in real time; recurrent neural network (LSTM) predicts the harmonic parameters at the next moment. The model inference time is optimized to 0.5ms to ensure that the real-time requirements of the harmonic signal update cycle of 5ms are met.

[0170] In terms of safety protection mechanism, during the harmonic injection process, the winding temperature and magnetic field strength distribution are monitored in real time: when the winding temperature exceeds 125°C, or the magnetic field strength deviation exceeds 10A / m for 10 consecutive sampling cycles, the protection mechanism is immediately triggered to stop harmonic injection and reduce the motor speed to a safe range. If the temperature is close to the threshold (such as 120±2°C), the amplitude of the high-frequency harmonic signal is preferentially reduced to 50% to reduce thermal stress.

[0171] Embodiment 3:

[0172] During the startup of the permanent magnet motor, this embodiment further optimizes the harmonic injection control method, wherein the harmonic parameter extraction and model input, for example, real-time acquisition of the three-phase current signal of the stator winding by a high-precision current sensor (sampling frequency of 20kHz), and the use of an improved adaptive filter (based on fast Fourier transform and least mean square algorithm) to decompose the signal, extract the harmonic frequency (range of 100Hz to 10kHz), amplitude (0 to 60% of the rated current) and phase information (0 to 360°). At the same time, frequency characteristics, amplitude, phase and sampled ambient temperature signals (sampling frequency of 1kHz, temperature range of -40°C to 85°C) are used as input features for training deep learning models to ensure coverage of complex working conditions.

[0173] In addition, in the dynamic optimization of genetic algorithms, an improved genetic algorithm can be used to optimize the amplitude and phase of the harmonic injection signal and clarify the logic of dynamic adjustment of the fitness function weight: that is, in the initial stage (torque fluctuation error>1Nm 2): The torque fluctuation error weight ratio is set to 70% first, and the power loss and magnetic field strength deviation weights are 15% each to ensure the stability of starting in the low-speed zone. 2 ): The power loss weight is increased to 50%, the torque fluctuation weight is reduced to 30%, and the magnetic field strength deviation weight is maintained at 20%.

[0174] The optimization objective is precisely controlled through the fitness function of multi-objective constraints:

[0175]

[0176] Where, f: fitness function, used to measure the quality of harmonic injection parameters; ΔT: torque fluctuation error; T max : Maximum target torque; P loss : Actual power loss caused by harmonic injection; P max : Maximum allowable power loss; σ B : The mean square error of the magnetic field intensity distribution, defined as the degree to which the magnetic field intensity in the winding deviates from the mean value; B max : Maximum target magnetic field strength; ω1, ω2, ω3 are dynamic weight coefficients, satisfying ω1+ω2+ω3=1.

[0177] In terms of spatial distribution optimization, the spatial distribution of the optimized harmonic signal is further optimized: for example, in the area close to the load end (within 5 cm), a high-frequency harmonic signal (5-10kHz) is injected, and the total power of the high-frequency injection signal is controlled within 35% of the total power of the winding to reduce eddy current losses. In the area far away from the load end (more than 5 cm), a low-frequency harmonic signal (100 Hz to 1 kHz) is injected, and the root mean square value (RMS) of the winding current is ensured to be between 95% and 105% of the rated value by adjusting the signal amplitude. Add a spatial distribution optimization feedback mechanism: monitor the RMS value of the winding current in real time. If it exceeds the target range (such as >105%), dynamically reduce the injection amplitude of the high-frequency harmonic signal to 80% of the original value until the RMS value is restored.

[0178] Also, for deep learning model prediction, in terms of model architecture optimization, for example, an improved multi-scale convolutional neural network (MS-CNN) is used in combination with a bidirectional LSTM (Bi-LSTM) for modeling, and an attention mechanism is added to enhance the harmonic feature perception capability under complex working conditions. Its input feature expansion includes, for example, real-time extracted harmonic frequency, amplitude, phase, sampled ambient temperature signal (sampling period 1ms), and torque change rate (sampling frequency 500Hz). And in terms of model training data enhancement, for example, random noise and amplitude disturbance (±10%) are introduced to historical signal samples to generate an enhanced data set, the sample size is increased by 2 times, and the model robustness is improved. And in terms of model output, for example, predicting the amplitude, frequency, and phase of the harmonic signal at the next moment, the inference time is optimized to 0.8ms to meet the signal update cycle (5ms).

[0179] Furthermore, for dynamic phase adjustment, when the torque change rate is low (<100rad / s 2 ), the phase adjustment coefficient K is fixed at 0.1 to ensure the smoothness of the adjustment; when the torque change rate is high (>200rad / s 2 ), dynamically increase K to 1.5 to improve phase response capability. To avoid sudden changes in the phase adjustment process, the phase angle change after each adjustment is limited to no more than 20°, and the maximum phase adjustment rate is set to 0.8° / ms to ensure system stability.

[0180] In addition, in terms of safety protection mechanisms, for example, the real-time monitoring mechanism of temperature and magnetic field strength can be enhanced: the sampling period of thermistor sensors is optimized to 5ms, and the deviation range of Hall effect sensors is reduced to 3A / m. When the winding temperature is close to 125°C (such as 122±1°C), the amplitude of high-frequency harmonic signals is preferentially reduced by 50%; when the magnetic field strength deviation is close to 10% (such as 9.5%-10%), the low-frequency harmonic injection ratio is increased to 60% of the total injection amount. In addition, dynamic adjustment feedback records can be added, such as recording the last 5 optimization parameters for subsequent optimization analysis and fault recovery.

[0181] Embodiment 4:

[0182] During the startup process of the permanent magnet motor, this embodiment describes in detail the dynamic adjustment strategy, deep learning model architecture and specific implementation method of hardware support during the harmonic injection signal optimization process to improve the dynamic response capability of harmonic injection, model prediction accuracy and overall optimization efficiency.

[0183] For example, in terms of dynamic adjustment and feedback mechanism, error comparison and dynamic correction: the torque change signal in the startup phase is collected in real time, and the error is compared with the torque change signal predicted by the deep learning model. When the error exceeds 5%, the amplitude of the harmonic signal is dynamically adjusted, and the high-frequency harmonic signal injection amount is preferentially reduced by 10% (the adjustment amount is not less than 5% of the total injection amount). If the torque error exceeds the positive and negative range of the predicted value for 5 consecutive sampling cycles (each cycle is 10 milliseconds), further signal correction is triggered. The adjusted harmonic injection signal is recorded in real time, and the optimization results are used to update the parameters of the deep learning model in the next stage to ensure that the model can dynamically adapt to changes in working conditions.

[0184] In terms of deep learning model architecture, such as model architecture design: the deep learning model adopts a hybrid architecture, including: Convolutional neural network (CNN) part: use multi-scale convolution kernels (sizes are 3×3, 5×5 and 7×7 respectively) to extract harmonic features of different frequency bands, and enhance the perception of multi-frequency characteristics of harmonic signals. Recurrent neural network (RNN) part: combined with bidirectional long short-term memory network (Bi-LSTM), the time series characteristics of harmonic signals are modeled. Attention mechanism module: by dynamically allocating weights, the harmonic features of different time steps are weighted to improve the perception of transient signal changes. The modules in the model architecture cooperate with each other in function. CNN is responsible for feature extraction, Bi-LSTM realizes time series modeling, and the attention mechanism dynamically optimizes feature weighting to form an efficient and stable harmonic signal prediction system.

[0185] As well as model training and enhancement: the model is trained using a historical harmonic data set covering multiple working conditions, and the data sample size is expanded to 1.5 times the original data through data enhancement methods (such as introducing random noise and amplitude disturbances) to improve the robustness of the model. The model prediction error target is controlled within 5% to meet real-time requirements. In addition, in terms of input and output features, for example, input features include: real-time extracted harmonic frequency, amplitude, phase, sampled ambient temperature signal (sampling period 1 millisecond), torque change rate (sampling frequency 500Hz). The output is the predicted harmonic injection signal parameters at the next moment, including amplitude, frequency and phase.

[0186] In terms of hardware support, hardware performance requirements are as follows: Digital signal processor (DSP): processing speed not less than 1GHz, ensuring that the dynamic optimization calculation time is less than 5 milliseconds to meet real-time requirements. Current sensor and torque sensor: sampling resolution not less than 0.01A and 0.1Nm respectively, to ensure the accuracy of optimization input data. Parallel computing support: The hardware control module adopts a parallel computing architecture to support the synchronous processing of harmonic signal dynamic optimization and spatial distribution optimization. In addition, the hardware guarantees the realization of the technology, and the hardware performance directly determines the real-time optimization calculation and the dynamic response capability of harmonic injection. The high processing speed of the DSP and the combination of high-precision sensors are used to ensure that the wave optimization system runs without delay and error during the rapid startup process, which are all extended implementation methods known to ordinary technicians in this field.

[0187] Embodiment 5:

[0188] See also Figure 4 In order to verify the superiority of the technical solution of the present invention over the prior art, a comparative experiment was designed. The experiment was carried out during the startup process of the permanent magnet synchronous motor to compare the performance of the following three optimization methods:

[0189] Prior art 1: Fixed model optimization strategy based on PI control.

[0190] Prior art 2: Harmonic signal analysis and feedback control method combined with Fast Fourier Transform (FFT).

[0191] The present invention is a dynamic optimization system based on the combination of deep learning (CNN+Bi-LSTM+attention mechanism), genetic algorithm and real-time feedback mechanism.

[0192] The experimental data was collected from a permanent magnet synchronous motor with a rated power of 15kW, and the motor performance was tested under load change, startup transient and sudden operating condition change (load change rate up to ±50%). The evaluation indicators include:

[0193] Torque ripple RMS value (Nm): measures the smoothness of motor starting.

[0194] Power loss reduction rate (%): Evaluates the energy efficiency improvement after harmonic optimization.

[0195] Dynamic response time (ms): The response speed of the optimization algorithm.

[0196] The present invention extracts the frequency domain features of harmonic signals through CNN with multi-scale convolution kernels, combines Bi-LSTM to realize dynamic modeling of time series, and dynamically weights key transient signals through the attention mechanism, so that the adjustment of harmonic signal injection parameters is more accurate. Experimental results show that under complex working conditions, the RMS value of torque fluctuation is only 42.8% of that of prior art 1 and 50% of that of prior art 2, which significantly improves the stability of motor starting. However, the prior art is mostly based on fixed optimization models, which makes it difficult to adapt to load changes and transient disturbances in real time. Through the global optimization of genetic algorithms and the dynamic adjustment of real-time feedback mechanisms, the power loss reduction rate of the present invention is increased to 27%, which is 2.7 times that of prior art 1 and 1.93 times that of prior art 2, effectively reducing the energy waste caused by high-frequency harmonics and showing excellent energy efficiency optimization capabilities.

[0197] The present invention adopts high-performance DSP (processing speed not less than 1GHz) and parallel computing architecture to shorten the optimization calculation time to 5 milliseconds, which is much lower than the prior art 1 (50 milliseconds) and the prior art 2 (30 milliseconds). This millisecond-level response capability significantly improves the real-time and adaptability of the system under complex dynamic conditions. In addition, the present invention realizes accurate modeling and optimization of complex dynamic conditions through the collaborative mechanism of deep learning and genetic algorithm, and solves the bottleneck problem that it is difficult to balance real-time and accuracy in the prior art. In the experiment, in the scenario where the load change rate reaches ±50%, both the prior art 1 and the prior art 2 show a large torque fluctuation, while the present invention can complete dynamic adjustment within 5 milliseconds and control the torque fluctuation RMS within 0.9Nm, which exceeds expectations. At the same time, the present invention is based on real-time feedback optimization closed-loop system, which can quickly adjust the harmonic injection parameters when the working condition changes suddenly, and maintain the long-term stability of the optimization results. In the long-term test simulating the load fluctuation scenario, the optimization performance fluctuation range of the present invention is only ±2%, while the optimization performance fluctuation ranges of the prior art 1 and the prior art 2 reach ±12% and ±8%, respectively. Based on this, it can be seen from the experimental data and technical effects that the present invention is significantly superior to the existing technology in terms of real-time, accuracy and adaptability. The specific improvements include: the torque fluctuation reduction rate is increased to more than 50%; the power loss optimization capability is significantly improved, the dynamic response time is shortened to milliseconds, and the optimization efficiency is also significantly improved.

[0198] Embodiment 6:

[0199] The deep learning model predicts the harmonic signal adjustment suggestions for the next moment based on the real-time collected harmonic parameter set (including frequency, amplitude, phase, etc.) as the initial parameter seed for genetic algorithm optimization. During the genetic algorithm optimization process, the feedback value of the motor operation error (torque fluctuation, power loss and magnetic field strength distribution) is introduced in real time to dynamically correct the optimization target. When the deviation between the prediction result of the deep learning model and the optimization result of the genetic algorithm exceeds the preset range (such as 5%), the dynamic adjustment of the genetic algorithm is prioritized for correction. The number of iterations of the genetic algorithm is automatically reduced when the convergence rate is low to shorten the optimization time and save computing resources.

[0200] In addition, for the dynamic phase adjustment of harmonic signals, a sensitivity adjustment mechanism based on real-time feedback is designed: when the torque changes slightly, the adjustment amplitude is kept at a low speed to ensure the stability of the system; when the torque changes significantly, the adjustment sensitivity is dynamically increased to quickly respond to changes in external working conditions; to avoid excessive adjustment causing system oscillation, the upper limit of the maximum adjustment amplitude and response speed is set. The dynamic adjustment strategy ensures that the optimization results are stable and real-time, and adapts to the needs of different operating stages.

[0201] In addition, according to the physical layout of the motor stator winding, the signal injection area can be clearly divided into a high-frequency priority area and a low-frequency priority area, such as the high-frequency priority area: close to the load end (within 5cm), high-frequency harmonic signals are injected first to enhance transient response capabilities, while controlling the high-frequency injection power within the range of 30%-35% of the total winding injection power. Low-frequency priority area: far away from the load end (outside 5cm), low-frequency harmonic signals are injected first to reduce high-frequency electromagnetic interference and keep the current root mean square value (RMS) within the rated value range. The winding current RMS value is monitored in real time. When it deviates from the target range, the proportion of high-frequency harmonic injection is automatically adjusted to ensure the stability of the signal energy distribution.

[0202] In addition, the prediction capability of the model can be enhanced through multiple input features (such as frequency, amplitude, phase, ambient temperature and torque change rate), and the feature range can be expanded to adapt to complex working conditions, such as enhancing the diversity of data sets: perturb historical signal data (such as randomly changing amplitude or frequency), generating more samples, and improving the robustness of the model under different working conditions. Optimize the model architecture: Combine multi-scale convolution kernels to extract features of different frequency bands, and dynamically weight key transient signals through the attention mechanism to enhance the model's response to complex dynamic signals.

[0203] And, the optimization calculation is controlled within 5ms throughout the process to ensure the millisecond-level response capability of harmonic signal adjustment. For example, a multi-core digital signal processor (DSP) is used to ensure the efficient collaborative calculation of the genetic algorithm and the deep learning model; high-precision current sensors and torque sensors are equipped to ensure the accuracy of the optimized input data. And in the safety protection mechanism, a two-level early warning mechanism is added, such as a first-level early warning: when the winding temperature is close to 125°C (such as 120±2°C) or the magnetic field strength deviation is close to 10%, the injection amplitude of the high-frequency harmonic signal is preferentially reduced. Second-level early warning: When the temperature exceeds 125°C or the magnetic field strength exceeds the threshold continuously, the emergency protection mode is triggered, the harmonic signal injection is stopped, and the motor speed is gradually reduced to a safe range. In this protection mode, the last optimization parameters are recorded in real time to facilitate subsequent recovery and fault analysis, which are all extended implementation methods known to ordinary technicians in this field.

[0204] In the present invention, unless otherwise clearly stipulated and limited, the terms such as "installation", "setting", "connection", "fixation" and "screw-on" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral one; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0205] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be considered exemplary and non-restrictive in all respects, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims be included in the present invention.

Claims

1. A permanent magnet motor starting control method based on harmonic injection, characterized in that: The following steps are involved: Step 1, harmonic parameter extraction: During the starting process of the permanent magnet motor, the real-time starting current signal is collected, and the current signal is decomposed in real time using an adaptive filter based on the least mean square algorithm to extract the harmonic frequency, amplitude and phase information in the current signal to generate a harmonic parameter set; The harmonic parameter set includes frequency characteristics, amplitude and phase information collected in real time, wherein the frequency characteristics will be used as one of the input features of the deep learning model and used in subsequent optimization steps; Step 2, multi-frequency harmonic dynamic optimization injection: According to the harmonic parameter set, the amplitude and phase of the harmonic injection signal are optimized and adjusted using a genetic algorithm, wherein: The optimization goal is to minimize the torque fluctuation in the low-speed area during the motor startup phase and reduce the energy loss caused by the harmonic injection signal; The optimization objective function is defined as: Among them, M actual,i is the actual torque after the i-th harmonic injection, M target,i is the set target torque, P loss,i is the power loss caused by harmonic injection, w i To optimize the weight coefficient; The weight coefficient w i The calculation formula is: If ΔM actual，i =0 or ΔM actual，i =0, then the weight coefficient w i Set to the default value of 0.1 to avoid calculation anomalies; The result of the dynamic optimization generates an adjusted harmonic injection signal, which includes an optimized amplitude, frequency and phase, and is used as an input for subsequent steps; Step 3, optimization of spatial distribution of harmonic injection: according to the harmonic injection signal generated in step 2, the spatial distribution of the signal is optimized, wherein: Divide the motor stator winding into multiple spatial injection areas: The area close to the load end is defined as the high-frequency priority injection area, into which high-frequency harmonic signals are preferentially injected, with a frequency range of 1 kHz to 10 kHz to enhance the transient electromagnetic response capability; The area far away from the load end is defined as the low-frequency priority injection area, into which low-frequency harmonic signals are preferentially injected, with a frequency range of 100 Hz to 1 kHz to reduce high-frequency electromagnetic interference; The goals of spatial distribution optimization include: the mean square error of magnetic field intensity distribution is less than 10A / m 2 ; The high-frequency harmonic injection amount does not exceed 30% of the total winding injection amount; When optimizing the spatial distribution of harmonic signals, the root mean square value RMS of the winding current is calculated, and the RMS value is calculated according to the following formula: Where i(t) is the function of the winding current changing with time, T is the period of the current signal; the RMS value is used to evaluate the stability of the energy distribution injected by the harmonic signal; Step 4, harmonic injection based on time dynamic phase delay: by real-time monitoring of the torque sensor output or based on the dynamic analysis of the current signal, the influence of the harmonic injection signal on the transient torque change is obtained, and the phase delay angle of the harmonic injection signal is dynamically adjusted using the transient torque tracking control algorithm; The phase delay angle is calculated based on the following formula: Among them, θ optimal is the initial phase angle; ΔT is the transient torque change, Δt is the time interval, K is the dynamic adjustment coefficient, and the specific adjustment rules are as follows: When ΔT / Δt≤50rad / s 2 ,K=0.1; When 50rad / s 2 <ΔT / Δt≤200rad / s 2 ,K=0.5; When ΔT / Δt>200rad / s 2 ,K=1.0; The total adjustment angle of dynamic phase delay shall not exceed 180 rad / s 2 To avoid winding magnetic field mismatch or reverse harmonic interference; Step 5: Intelligent predictive control of harmonics: Combined with the deep learning model, intelligent predictive adjustment of harmonic injection signals is realized, where: The deep learning model includes: Convolutional neural network (CNN) extracts the frequency characteristics of real-time current signals, including high-frequency and low-frequency harmonic components; The recurrent neural network (RNN) combines real-time features and historical harmonic data through the long short-term memory (LSTM) network to predict the amplitude, frequency, and phase of the harmonic injection signal at the next moment. The input features of the deep learning model further include ambient temperature signals and torque change rates collected in real time to improve the model's prediction ability for complex working conditions; The inference time of the deep learning model is controlled within 1 millisecond, meeting the real-time requirement of the harmonic signal update period of 5 milliseconds; Step 6, safety protection mechanism: During the harmonic injection process, safety protection is achieved through the following steps: The thermistor sensor samples the stator winding temperature in real time with a period of 10 milliseconds, and the temperature threshold is no more than 125 degrees Celsius; The Hall effect sensor array collects the magnetic field intensity distribution in the winding in real time at a frequency of 100 times per second, and the optimization goal is to control the magnetic field intensity deviation within 5A / m; When the winding temperature or magnetic field strength exceeds the threshold for 10 consecutive sampling cycles, the emergency protection mode is triggered, the harmonic signal injection is stopped and the motor speed is gradually reduced to a safe range.

2. A permanent magnet motor starting control method and system based on harmonic injection according to claim 1, characterized in that: The fitness function of the genetic algorithm is further defined as follows: The square error of the motor starting torque fluctuation is: Actual power loss caused by harmonic injection signal: Mean square error of magnetic field intensity distribution: Among them, B j is the magnetic field strength at the jth point in the winding, is the average value of the magnetic field strength.

3. A permanent magnet motor starting control method and system based on harmonic injection according to claim 2, characterized in that: The optimizer of the deep learning model adopts Adam optimizer with a learning rate of 0.

001. During the model training process, the prediction error target of frequency features and time series data is controlled within 5%.

4. A permanent magnet motor starting control method and system based on harmonic injection according to claim 3, characterized in that: The dynamic optimization model meets the following constraints when performing optimization calculation: the stator winding temperature shall not exceed 125 degrees Celsius; the magnetic field intensity deviation is controlled within 5%; and the optimization calculation time does not exceed 10% of the dynamic adjustment period.

5. A permanent magnet motor starting control method and system based on harmonic injection according to claim 1, characterized in that: The spatial distribution optimization of harmonic injection in step 3 also includes the following objectives: in the high-frequency priority injection area, the injection power of the harmonic signal accounts for no more than 40% of the total injection power to avoid eddy current losses caused by excessive high-frequency signals; in the low-frequency priority injection area, the root mean square value (RMS) of the winding current after the harmonic signal is injected is maintained between 90% and 110% of the rated value to ensure the stability of the winding energy distribution.

6. A permanent magnet motor starting control method and system based on harmonic injection according to claim 5, characterized in that: The dynamic adjustment rules of the fitness function of the genetic algorithm include: in the initial stage of starting, the weight ratio of the torque fluctuation error is preferentially set to 60%-80% to reduce the starting oscillation in the low-speed zone; in the middle and late stages of starting, the weight ratio of the power loss is dynamically adjusted to 40%-60% to optimize the motor starting efficiency; the weight ratio of the magnetic field uniformity deviation is always kept within 20% to avoid the influence of excessive weight on other optimization objectives.

7. A permanent magnet motor starting control method and system based on harmonic injection according to claim 1, characterized in that: The input features of the deep learning model further include: a torque change rate signal with a sampling frequency of 500 Hz; a real-time waveform of the winding current caused by harmonic injection with a sampling frequency of 1 kHz; and an ambient temperature signal during the startup phase with a sampling period of 10 milliseconds.

8. A permanent magnet motor starting control method and system based on harmonic injection according to claim 1, characterized in that: The training process of the deep learning model includes the following enhancement mechanisms: Data enhancement: By randomly perturbing the historical harmonic signals, an extended data set of no less than 1.5 times the original data volume is generated to improve the robustness of the model; Dynamic adjustment of model training accuracy: When the training error convergence rate is lower than 1% / epoch, the learning rate is automatically reduced to 50% of the original learning rate; Overfitting protection: During the training process, the model performance is monitored in real time through cross-validation. When the error of the validation set is greater than twice the error of the training set, the model early stopping mechanism is triggered.

9. A permanent magnet motor starting control method and system based on harmonic injection according to claim 1, characterized in that: The security protection mechanism also includes: When the winding temperature is close to 125 degrees Celsius, such as 120 ± 2 degrees Celsius, it is preferred to reduce the injection amplitude of the high-frequency harmonic signal to within 50% of the original amplitude to reduce the thermal stress of the winding; When the magnetic field intensity deviation is close to 10%, such as 8%-10%, the injection ratio of the low-frequency harmonic signal is increased in real time to balance the magnetic field distribution in the winding; If the temperature or magnetic field strength exceeds the threshold and cannot be restored, the emergency protection mode is triggered, the harmonic signal injection is stopped and the motor speed is gradually reduced to a safe range.

10. A permanent magnet motor starting control method and system based on harmonic injection according to claim 1, characterized in that: The dynamic phase adjustment in step 4 further includes the following rules: When the amplitude of the harmonic injection signal exceeds 20% of the rated amplitude, the adjustment rate of the dynamic phase delay angle is limited to within 1° / ms to avoid motor oscillation caused by phase mutation; When the transient torque change rate is greater than 200 rad / s 2 When , the dynamic adjustment coefficient K can be dynamically increased to 1.5 to improve the sensitivity of phase adjustment; During the dynamic phase delay adjustment process, the phase angle change after each adjustment shall not exceed 30° to ensure the stability of the control system. The steps 2 and 5 are combined with a real-time feedback mechanism to further optimize the harmonic injection signal through the following steps.

Images