Direct-driven permanent magnet motor speed optimization control method based on multi-strategy self-evolution perception fusion and toughness control

Abstract

The invention relates to a direct-drive permanent magnet motor speed optimization control method based on multi-strategy self-evolution perception fusion and toughness control, and solves the problems of poor model adaptability, sensitive sensor fault, weak disturbance suppression and the like in the prior art. The adopted technical scheme is that the method comprises the following steps: establishing a self-evolution motor modeling system; on the basis of a motor modeling system, credibility evaluation is carried out on collected multi-source motor operation data, and a fusion state estimation result is generated in combination with a multi-scale residual weighting mechanism; constructing a cognitive main controller, and outputting a motor speed optimization control signal; and setting a light fault threshold value and a heavy fault threshold value, constructing a three-level toughness control mechanism, and performing graded response on faults to realize fault recovery. The method has the advantages that the motor speed control precision, the system stability and the self-adaptive capacity are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent control and state estimation of electric machines, and particularly relates to a direct-drive permanent magnet motor speed optimization control method based on multi-strategy self-evolution perception fusion and resilience control. BACKGROUND

[0002] As the core driving unit of modern industrial servo, robots, electric vehicles, etc., the operation control performance of direct-drive permanent magnet motor will directly affect the energy efficiency and safety of the entire system. However, in a complex operating environment, the existing control system mainly has the following technical bottlenecks: the static filter estimator is slow in responding to parameter changes and multi-source heterogeneous data; the controller parameters are fixed or the fuzzy rules are simple, lacking intelligent adjustment linked with working conditions; fault detection is lagging, lacking a robust control structure integrating diagnosis, switching and recovery. Therefore, there is an urgent need for an electric machine control system that integrates multi-dimensional perception, self-evolution control strategy and closed-loop fault self-repairing capability to significantly improve the operation stability, self-adaptability and control robustness. SUMMARY

[0003] The present application aims to solve the above-mentioned problems existing in the prior art and provides a direct-drive permanent magnet motor speed optimization control method based on multi-strategy self-evolution perception fusion and resilience control, which integrates multi-scale weighted Kalman fusion, Bayesian regularization neural network self-learning, fuzzy parameter adjustment PID and disturbance observer cooperative control, and a sliding mode-model prediction-redundancy structure fault resilience switching mechanism, realizing high-robust control of permanent magnet motor under multi-interference, multi-sensor redundancy and complex load.

[0004] The above technical purpose of the present application is mainly solved by the following technical scheme: a direct-drive permanent magnet motor speed optimization control method based on multi-strategy self-evolution perception fusion and resilience control, characterized by comprising the following steps:

[0005] Establishing a self-evolving motor modeling system, which includes a temperature-adaptive spatial model and an auxiliary model with anti-overfitting capability;

[0006] Based on the spatial model and the auxiliary model, performing credibility evaluation on the collected multi-source motor operation data, and then generating a fusion state estimation result in combination with a multi-scale residual weighted mechanism;

[0007] Building a cognitive main controller, taking the fusion state estimation result as input, dynamically adjusting the control parameters and compensating the disturbance, and outputting the motor speed optimization control signal;

[0008] Setting a light fault threshold and a heavy fault threshold, building a three-level resilience control mechanism of the main controller as the foundation-sliding mode observer assisting light fault-model predictive control (MPC) taking over heavy fault, responding to faults in stages, and realizing fault recovery.

[0009] The technical scheme is used for solving the problems of poor model uncertainty response, strong sensor fault sensitivity, static controller parameters, weak disturbance suppression ability and lag fault processing means in traditional permanent magnet motor control. The technical scheme greatly enhances the stability, flexibility and intelligent level of the system while improving the control performance of the direct-drive permanent magnet motor, and is suitable for modern complex industrial control scenes with high reliability and high responsiveness requirements.

[0010] As a further improvement and supplement to the above technical scheme, the present application adopts the following technical measures:

[0011] As preferred, after the fault recovery, the process data is fed back to the motor modeling system, the model parameters are dynamically corrected, and the main controller parameters are updated, so as to realize self-evolution optimization and continuously output optimized speed control signals.

[0012] As preferred, the space model dynamically corrects the temperature-sensitive parameters of the motor based on the initial values of the parameters at normal temperature, the temperature drift coefficients and the reference temperature, the temperature-sensitive parameters including resistance values and back electromotive force constants, and the temperature drift coefficients are obtained through offline identification or online recursive estimation.

[0013] As preferred, the auxiliary model is based on a long short-term memory neural network (LSTM) with Bayesian regularization, the loss function of the auxiliary model includes a deviation term of the predicted motor speed and the actual motor speed, and a penalty term of the change of the network parameters in the current iteration or the initial training, and the total loss function value is adjusted by adjusting the weights of the deviation term and the penalty term through a regularization coefficient, so as to improve the training stability and prevent overfitting.

[0014] As preferred, in the step of credibility evaluation, a credibility weight is assigned to each data source, the credibility weight integrates the residual variance and the sensor health factor, the residual variance reflects the stability of the deviation between the measured value of the data source and the predicted value of the model, and the sensor health factor reflects the health degree of the running state of the sensor, which is used to dynamically adjust the observation noise covariance matrix of Kalman filtering.

[0015] As preferred, the step of the multi-scale residual weighted mechanism includes: setting a plurality of different time scales, outputting state estimation results at each time scale through an estimator; calculating the observation residual of each scale; dynamically assigning a weighting coefficient according to the norm of the residual vector; and fusing the state estimation results of each scale according to the weighting coefficient to generate a fused state estimation result, the observation residual being the difference between the observation and the mapped quantity based on the state estimation result of the scale.

[0016] As preferred, the fuzzy self-tuning PID controller integrated by the cognitive main controller takes the speed error in the fusion state estimation result and the error change rate as input, adaptively adjusts the proportional coefficient, integral coefficient and differential coefficient, outputs the motor speed optimization control signal through the control law, the speed error is the difference between the target rotating speed and the actual rotating speed, and the error change rate is the difference between the current and last time speed error divided by the sampling period.

[0017] As preferred, the cognitive main controller integrates a disturbance observer (DOB) and a feedback compensator, the disturbance observer (DOB) estimates the external disturbance based on the residual error between the system actual output in the fusion state estimation result and the predicted output based on the system transfer function, and extracts the external disturbance, modeling error and un-modeled dynamic through a band-pass or low-pass filter; the feedback compensator compensates the estimation result of the external disturbance in real time.

[0018] As preferred, in the step of fault grading response:

[0019] When the light fault occurs, the sliding mode observer is activated, and the sliding mode estimation response strength is determined based on the state transition matrix, control input matrix, observation matrix, control input signal, actual output signal, gain matrix and sign function of the space model, so as to suppress the disturbance;

[0020] When the heavy fault occurs, the MPC controller is switched to, the deviation between the system state and the reference state trajectory and the control input energy consumption are balanced based on the optimization objective function, and the operation stability is ensured.

[0021] As preferred, the light fault threshold and the heavy fault threshold are set based on the statistical result of the operation residual error under the normal working condition of the system, the light fault threshold is the normal residual error mean plus twice the standard deviation, and the heavy fault threshold is the normal residual error mean plus four times the standard deviation.

[0022] The application has the beneficial effects: through the adaptive mechanism data fusion and the advanced control strategy, the high precision and the fast response of the direct-drive permanent magnet motor speed control are realized, so that the performance of the motor is optimized and the stability and the reliability of the system are improved. It can adapt to the variable working condition, prolong the service life of the motor, reduce the maintenance cost, and is easy to implement in the existing system. In addition, the application can also improve the production efficiency and the energy saving effect, and for industrial application, it can reduce the energy consumption and improve the overall efficiency of the production line. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 It is a flow chart of the application from the operation steps.

[0024] Figure 2 It is a flow chart of the application from the data input and output angle. DETAILED DESCRIPTION

[0025] The technical solutions of the present application will be further described in detail below by examples and in conjunction with the drawings.

[0026] Embodiment: As shown in Figure 1 and Figure 2 A direct-drive permanent magnet motor speed optimization control method based on multi-strategy self-evolution perception fusion and resilience control includes the following steps:

[0027] A self-evolution motor modeling system is established, which includes a temperature-adaptive spatial model and an auxiliary model with anti-overfitting capability;

[0028] Based on the spatial model and the auxiliary model, the credibility of the collected multi-source motor operation data is evaluated, and a fusion state estimation result is generated in combination with a multi-scale residual weighted mechanism;

[0029] A cognitive main controller is constructed to take the fusion state estimation result as input, dynamically adjust the control parameters and compensate for disturbances, and output motor speed optimization control signals;

[0030] Set light fault threshold and heavy fault threshold, build a three-level resilience control mechanism of main controller-based-sliding mode observer-assisted light fault-model predictive control (MPC) to take over heavy fault, respond to fault classification, and realize fault recovery.

[0031] The technical solution is used to solve the problems of poor model uncertainty response, strong sensor fault sensitivity, static controller parameters, weak disturbance suppression capability and lag fault handling means in traditional permanent magnet motor control. The technical solution greatly enhances the stability, flexibility and intelligent level of the system while improving the control performance of the direct-drive permanent magnet motor, and is suitable for modern complex industrial control scenarios with high reliability and high responsiveness requirements.

[0032] Next, the above technical solution is further expanded and described in detail:

[0033] In actual application, after realizing fault recovery, the process data is fed back to the motor modeling system, the model parameters are dynamically corrected, and the main controller parameters are updated to realize self-evolution optimization and continuously output optimized speed control signals. (The fault recovery process data dynamically corrects the parameters of the temperature-adaptive state space model (such as temperature drift coefficient) and the network parameters of the Bayesian regularization LSTM auxiliary model in S1; the update of the main controller includes adjusting the parameters of the fuzzy self-tuning PID, the gain of the disturbance observer, etc. to realize continuous optimization of the control signal.)

[0034] In practical applications, the space model is used to dynamically correct temperature-sensitive parameters of the motor based on initial values of the parameters at normal temperature, temperature drift coefficients and a reference temperature, the temperature-sensitive parameters including resistance values and back electromotive force constants, and the temperature drift coefficients being obtained through offline identification or online recursive estimation.

[0035] In the technical solution, when the space model is constructed:

[0036] In the space model, resistance R and back electromotive force K e are dynamically corrected according to the following relationship with temperature change:

[0037] R(T)=R0(1+α R (T-T0)), K e (T)=K e0 0(1+α K (T-T0)) (1)

[0038] Wherein, R0, K e0 are values at normal temperature; α R , α K are temperature drift coefficients, obtained through offline identification or online recursive estimation; and T0 is a reference temperature.

[0039] In practical applications, the auxiliary model is based on a long short-term memory neural network (LSTM) with Bayesian regularization, the loss function of the auxiliary model includes a deviation term of predicted motor speed and actual motor speed and a penalty term of changes in network parameters in the current iteration or initial training, and the total loss function value is adjusted by adjusting the weights of the deviation term and the penalty term through a regularization coefficient, so as to improve the training stability and prevent overfitting.

[0040] In the technical solution, the loss function is:

[0041]

[0042] Wherein, L is a total loss function value, used to measure the deviation between predicted output and true value, and simultaneously includes a model complexity penalty term; λ is a regularization coefficient, a non-negative hyperparameter, used to adjust the trade-off between the prediction error term and the parameter change penalty term; θ is a current LSTM network parameter vector (such as weights and biases, etc.), representing a set of network structure parameters obtained through current training; θ old is an LSTM network parameter vector in the last iteration or initial training; ω t is an actual motor speed (true value) at the t th moment, obtained from sensor measurement; is a predicted motor speed at the t th moment, output by the LSTM network; and ‖θ-θ old ‖ 2L2 norm (Euclidean norm square) for parameter variation, used to punish the drastic update of model parameters, improve the training stability, and prevent overfitting.

[0043] In practical applications, in the step of credibility evaluation, a credibility weight is assigned to each data source, which comprehensively considers residual variance and sensor health factor, the residual variance reflects the stability of the deviation between the measured value of the data source and the predicted value of the model, and the sensor health factor reflects the health degree of the running state of the sensor, and is used for dynamically adjusting the observation noise covariance matrix of Kalman filtering.

[0044] In the technical solution, a credibility weight w i :

[0045]

[0046] wherein, is the residual variance, and p i is the sensor health factor, used for dynamically adjusting the observation noise covariance matrix of Kalman filtering.

[0047] In practical applications, the step of the multi-scale residual weighted mechanism comprises: setting a plurality of different time scales, outputting state estimation results at each time scale through an estimator; calculating observation residuals at each scale; dynamically assigning a weighting coefficient according to the norm of the residual vector; and fusing the state estimation results at each scale according to the weighting coefficient to generate a fused state estimation result, wherein the observation residual is the difference between the observation and the mapped quantity based on the state estimation result at the scale.

[0048] In the technical solution, the state estimation fusion adopts a multi-scale Kalman filter, and the weighting manner is as follows:

[0049] For each scale i, the observation residual at time k represents the difference between the model prediction and the actual measurement, and is calculated as follows:

[0050]

[0051] wherein, is the state estimation result of the estimator at the kth sampling time under the ith time scale (such as short-term, high-frequency or long-term, low-frequency), which is usually obtained from the output of the ith Kalman filter; z k is the observation or measurement data (such as motor speed sensor output) obtained at the kth sampling time, used for residual calculation with the estimated value; and C is an observation matrix in the state space model, used for mapping the measurable quantity (such as motor speed) from the system state vector x k .

[0052] The dynamic weighting coefficient is:

[0053]

[0054] wherein, is the fusion weight coefficient of the i-th scale at the k-th time, reflecting the relative importance of the scale in the fusion estimation; is the norm (usually L2 norm) of the residual vector of the i-th estimated scale, used to measure the reliability of the scale estimation, the smaller the residual, the higher the reliability.

[0055] The fusion estimation is:

[0056]

[0057] The final state estimation result after fusion, representing the output of weighted superposition of each scale estimation result, used as the input of the subsequent controller.

[0058] In actual application, the fuzzy self-tuning PID controller integrated in the cognitive main controller takes the speed error and error change rate in the fusion state estimation result as input, and adaptively adjusts the proportional coefficient, integral coefficient and differential coefficient, and outputs the motor speed optimization control signal through the control law.

[0059] In the technical solution,

[0060] The fuzzy self-tuning PID controller takes the error e(t) and error change rate Δe(t) as input, and outputs the parameters of the adaptively adjusted proportional coefficient K p , integral coefficient K i and differential coefficient K d , and the control law is:

[0061]

[0062] The speed error at the current time, defined as the difference between the target speed and the actual motor speed, is calculated as:

[0063] e(t) = ω ref (t) - ω(t) (8)

[0064] The error change rate is calculated as:

[0065]

[0066] wherein, T s represents the sampling period of the controller.

[0067] In practical applications, the cognitive main controller integrates a disturbance observer (DOB) and a feedback compensator, the disturbance observer (DOB) estimates external disturbances based on the residual between the actual output in the fusion state estimation result and the predicted output based on the system transfer function, and extracts external disturbances, modeling errors and unmodeled dynamics through a band-pass or low-pass filter; the feedback compensator compensates the estimation result of external disturbances in real time.

[0068] In the technical solution,

[0069] The disturbance observer (DOB) estimates external disturbances (i.e. real-time estimation of external disturbances, modeling errors or unmodeled dynamics by the DOB) as follows:

[0070]

[0071] Wherein, G(s) is a system transfer function, representing a dynamic relationship model of the motor from the control input (voltage) to the output variable (such as speed); Q(s) is a band-pass or low-pass filter, usually selected as a low-pass or band-pass structure, used to extract disturbance components from the residual signal to avoid high-frequency noise misestimation; y(s) is the actual output (such as speed) signal of the system after Laplace transformation; u(s) is the Laplace transform form of the controller output signal, i.e. the motor driving voltage.

[0072] In practical applications, in the step of fault classification response:

[0073] When a light fault occurs, a sliding mode observer is activated, and the sliding mode estimation response strength is determined based on the state transition matrix, the control input matrix, the observation matrix, the control input signal, the actual output signal, the gain matrix and the sign function of the space model, to suppress disturbances;

[0074] When a heavy fault occurs, switch to the MPC controller, balance the deviation of the system state and the reference state trajectory and the control input energy consumption based on the optimization objective function, and ensure the operation stability.

[0075] In the technical solution,

[0076] Light fault: δ1<‖r k ‖<δ2: enter the sliding mode state estimation + DOB compensation working mode,

[0077] The sliding mode observer is enabled in the light fault state, and its form is:

[0078]

[0079] Wherein, is the system state vector estimated by the sliding mode observer; A, B, and C are the state transition matrix, control input matrix, and aforementioned observation matrix in the system state-space model, respectively; u is the system control input signal (such as drive voltage); y is the actual output of the system (such as measured motor speed); P is the gain matrix of the sliding mode observer, which determines the response strength of the sliding mode estimation; sgn(·) is the sign function, which depends on the sign of the measurement error.

[0080] Major fault: ||r k ‖>δ2; Switch the MPC controller, with the optimization objective being:

[0081]

[0082] Where, x ref (k) is the reference state trajectory (ideal operating target) set in MPC control; x(k) is the current actual or predicted state of the system at time k; u(k) is the control input variable at time k; η is the energy consumption weighting factor, the trade-off between control accuracy and energy consumption; N p The prediction step size of the predictive controller is the length of the future prediction time domain.

[0083] In practical applications, the minor fault threshold and the major fault threshold are set based on the statistical results of the operating residuals under normal system conditions. The minor fault threshold is the mean of the normal residuals plus two standard deviations, and the major fault threshold is the mean of the normal residuals plus four standard deviations.

[0084] In this technical solution,

[0085] The settings for δ1 and δ2 are based on the following:

[0086] By collecting the system's operating residual ||r| under normal, non-faulty operating conditions k The mean μ is calculated. r and standard deviation σ r The lower limit for judging minor faults is set as: δ1 = μ r +2σ r This indicates that the fluctuation exceeds 95% of the normal range; the upper limit for judging serious faults is set as: δ2 = μ r +4σ r It is used to identify severe anomalies or structural mismatches.

[0087] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Various modifications and variations can be made to the above embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A direct-drive permanent magnet motor speed optimization control method based on multi-strategy self-evolution perception fusion and resilience control, characterized in that, The method comprises the following steps: establishing a self-evolution motor modeling system including a temperature self-adaptive spatial model and an auxiliary model with anti-overfitting capability; based on the spatial model and the auxiliary model, performing credibility evaluation on the collected multi-source motor operation data, and combining a multi-scale residual weighted mechanism to generate a fusion state estimation result; constructing a cognitive main controller to take the fusion state estimation result as input, dynamically adjust the control parameters and compensate for disturbances, and output motor speed optimization control signals; setting light fault threshold and heavy fault threshold, constructing a three-level resilience control mechanism of main controller-based-sliding mode observer-assisted light fault-model predictive control (MPC) to take over heavy fault, responding to fault classification, and realizing fault recovery.

2. The multi-strategy self-evolution based on perception fusion and resilience control of direct-drive permanent magnet motor speed optimization control method according to claim 1, characterized in that: After realizing fault recovery, the process data is fed back to the motor modeling system to dynamically correct the model parameters and update the main controller parameters, realizing self-evolution optimization and continuously outputting optimized speed control signals.

3. The direct-drive permanent magnet motor speed optimization control method based on multi-strategy self-evolution perception fusion and resilience control according to claim 1 or 2, characterized in that: the spatial model dynamically corrects the temperature-sensitive parameters of the motor, including resistance value and back electromotive force constant, based on the initial value of the parameters at normal temperature, temperature drift coefficient and reference temperature, and the temperature drift coefficient is obtained through offline identification or online recursive estimation.

4. The direct-drive permanent magnet motor speed optimization control method based on multi-strategy self-evolution perception fusion and resilience control according to claim 1 or 2, characterized in that: the auxiliary model is based on a long short-term memory neural network (LSTM) with Bayesian regularization, the loss function of the auxiliary model includes a deviation term of predicted motor speed and actual motor speed, and a penalty term of current network parameter change from the last iteration or initial training, and the total loss function value is adjusted by adjusting the weight of the deviation term and the weight of the penalty term through a regularization coefficient, to improve the training stability and prevent overfitting.

5. The direct-drive permanent magnet motor speed optimization control method based on multi-strategy self-evolution perception fusion and resilience control according to claim 1 or 2, characterized in that: in the step of credibility evaluation, a credibility weight is assigned to each data source, which combines residual variance and sensor health factor, the residual variance reflects the stability of the deviation between the measured value of the data source and the predicted value of the model, and the sensor health factor reflects the health degree of the sensor operating state, which is used to dynamically adjust the Kalman filter observation noise covariance matrix.

6. The direct-drive permanent magnet motor speed optimization control method based on multi-strategy self-evolution perception fusion and resilience control according to claim 1 or 2, characterized in that: the step of the multi-scale residual weighted mechanism includes: setting multiple different time scales, outputting state estimation results at each time scale through an estimator; calculating the observation residual of each scale; dynamically assigning a weighting coefficient according to the norm of the residual vector; and fusing the state estimation results of each scale according to the weighting coefficient to generate a fusion state estimation result, the observation residual being the difference between the observation and the mapped quantity based on the state estimation result of the scale.

7. The direct-drive permanent magnet motor speed optimization control method based on multi-strategy self-evolution perception fusion and resilience control according to claim 1 or 2, characterized in that: the fuzzy self-tuning PID controller integrated in the cognitive main controller takes the speed error and error change rate in the fusion state estimation result as input, adaptively adjusts the proportional coefficient, integral coefficient and differential coefficient, outputs the motor speed optimization control signal through the control law, the speed error is the difference between the target speed and the actual speed, and the error change rate is the difference between the current and the last time speed error divided by the sampling period.

8. The direct-drive permanent magnet motor speed optimization control method based on multi-strategy self-evolution perception fusion and resilience control according to claim 1 or 2, characterized in that: the cognitive main controller integrates a disturbance observer (DOB) and a feedback compensator, the disturbance observer (DOB) estimates external disturbances based on the residual between the actual output of the system and the predicted output based on the system transfer function in the fusion state estimation result, and extracts external disturbances, modeling errors and unmodeled dynamics through a band-pass or low-pass filter; the feedback compensator compensates the estimation result of external disturbances in real time.

9. The direct-drive permanent magnet motor speed optimization control method based on multi-strategy self-evolution perception fusion and resilience control according to claim 1 or 2, characterized in that: in the step of fault classification response, when a light fault occurs, a sliding mode observer is activated, and the sliding mode estimation response strength is determined based on the state transition matrix, control input matrix, observation matrix, control input signal, actual output signal, gain matrix and sign function of the space model to suppress disturbances; when a heavy fault occurs, switch to the MPC controller, balance the deviation of system state and reference state trajectory and control input energy consumption based on the optimization objective function, and ensure the stability of operation.

10. The direct-drive permanent magnet motor speed optimization control method based on multi-strategy self-evolution perception fusion and resilience control according to claim 1 or 2, characterized in that: the light fault threshold and the heavy fault threshold are set based on the statistical results of the residual under normal operating conditions of the system, the light fault threshold is the mean value of the normal residual plus two times the standard deviation, and the heavy fault threshold is the mean value of the normal residual plus four times the standard deviation.

Images