Intelligent servo motor control method and device and computer equipment

Abstract

The invention relates to the technical field of servo motor control, and discloses an intelligent servo motor control method and device and computer equipment. The method comprises the following steps: carrying out multi-dimensional data acquisition and preprocessing on a position loop, a speed loop and a current loop of the intelligent servo motor to obtain standard structured parameter data; identifying a working condition characteristic parameter set of the intelligent servo motor; establishing a two-stage optimization architecture comprising a motor parameter layer and a control algorithm layer, and generating a dynamic parameter adjustment mapping table; a depth deterministic strategy gradient network is trained based on the characteristic parameter set and the dynamic parameter adjustment mapping table, and an optimal control strategy model for different working conditions is generated; and the optimal control strategy model is utilized, control parameters are automatically switched according to the working condition types recognized in real time, and smooth transition control of the intelligent servo motor among different working conditions is achieved. Smooth switching among different working conditions is achieved, system oscillation caused by sudden change of control parameters is avoided, and the disturbance influence in the working condition switching process is greatly reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of servo motor control, and in particular to an intelligent servo motor control method, device and computer equipment. Background Art

[0002] Traditional servo motor control methods mainly use PID control algorithms. Although they are simple in structure and easy to implement, they have many limitations under actual complex working conditions. These control methods usually use fixed parameter designs and cannot be adaptively adjusted for different working conditions, making it difficult to simultaneously meet the requirements of high precision and high dynamic performance under variable working conditions such as high-speed motion, precise positioning, and load mutations.

[0003] The key problem faced by existing intelligent servo motor control methods is the disconnection between working condition identification and control parameter optimization. On the one hand, there is a lack of effective automatic working condition identification mechanism, which makes it impossible to accurately identify the working state of the motor; on the other hand, controller parameter optimization mainly relies on manual experience adjustment or single-target optimization, and it is difficult to perform collaborative optimization at both the motor parameter layer and the control algorithm layer. Especially under the influence of uncertain factors such as large changes in load inertia ratio, encoder error, temperature drift and magnetic field nonlinearity, the position tracking error is large, the speed fluctuation is significant, the energy consumption is high and the temperature rise is too large, which seriously affects the overall performance of the intelligent servo system. Summary of the invention

[0004] The main purpose of the present invention is to provide an intelligent servo motor control method, device and computer equipment, which realizes smooth switching between different working conditions, avoids system oscillation caused by sudden changes in control parameters, and greatly reduces the disturbance effect during the working condition switching process.

[0005] To achieve the above object, the present invention provides an intelligent servo motor control method, comprising the following steps: Perform multi-dimensional data acquisition and preprocessing on the position loop, speed loop and current loop of the intelligent servo motor to obtain standard structured parameter data; Based on the standard structured parameter data, identifying the working condition of the intelligent servo motor, and obtaining a working condition characteristic parameter set; According to the operating condition characteristic parameter set, a two-level optimization architecture including a motor parameter layer and a control algorithm layer is established, and a dynamic parameter adjustment mapping table is generated; Training a deep deterministic policy gradient network based on the characteristic parameter set and the dynamic parameter adjustment mapping table to generate an optimal control strategy model for different working conditions; By utilizing the optimal control strategy model, control parameters are automatically switched according to the working condition type identified in real time, so as to realize smooth transition control of the intelligent servo motor between different working conditions.

[0006] The present invention also provides an intelligent servo motor control device, comprising: The acquisition module is used to collect and preprocess multi-dimensional data of the position loop, speed loop and current loop of the intelligent servo motor to obtain standard structured parameter data; An identification module, used to identify the working condition of the intelligent servo motor based on the standard structured parameter data, and obtain a working condition characteristic parameter set; An establishment module is used to establish a two-level optimization architecture including a motor parameter layer and a control algorithm layer according to the operating condition characteristic parameter set, and generate a dynamic parameter adjustment mapping table; A training module, used for training a deep deterministic policy gradient network based on the characteristic parameter set and the dynamic parameter adjustment mapping table to generate an optimal control strategy model for different working conditions; The switching module is used to utilize the optimal control strategy model to automatically switch control parameters according to the working condition type identified in real time, so as to realize smooth transition control of the intelligent servo motor between different working conditions.

[0007] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0008] In summary, the technical solution provided by the present invention collects multi-dimensional data of position loop, speed loop and current loop, and applies different filtering algorithms for targeted processing, effectively suppresses various signal noise interferences, improves the accuracy of state quantity estimation, and realizes accurate identification and classification of different working conditions such as high-speed motion, precision positioning, and load mutation based on K-means clustering algorithm and working condition feature extraction. Compared with traditional methods, the accuracy of working condition identification is significantly improved. A two-level optimization architecture of motor parameter layer and control algorithm layer is constructed to achieve comprehensive optimization from PWM parameters to control algorithm, solves the limitation of traditional methods that only optimize at a single level, and enables the system to maintain consistent dynamic response characteristics under different load inertia ratios. The control strategy model obtained by deep deterministic policy gradient network training has strong environmental adaptability and uncertainty processing capabilities, automatically generates optimal control parameters for different working conditions, and overcomes the defects of traditional fixed parameter control. The parameter gradient update algorithm and multi-working condition fusion control based on fuzzy logic are adopted to achieve smooth switching between different working conditions, avoid system oscillation caused by sudden changes in control parameters, and greatly reduce the disturbance effect during working condition switching. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 is a schematic diagram of the steps of an intelligent servo motor control method in one embodiment of the present invention; Figure 2is a structural block diagram of an intelligent servo motor control device in one embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0010] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0011] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0012] Reference Figure 1 , this embodiment provides an intelligent servo motor control method, comprising the following steps: S1, multi-dimensional data acquisition and preprocessing of the position loop, speed loop and current loop of the intelligent servo motor to obtain standard structured parameter data; Among them, a three-level cascade acquisition structure of position loop, speed loop and current loop is established for the intelligent servo motor to ensure comprehensive monitoring and analysis of multi-loop data. In order to ensure the accuracy and synchronization of the data, a high-precision data acquisition card is used to synchronously sample each loop, where the sampling frequency of the current loop is set to 20kHz to ensure the accurate capture of the high-frequency characteristics of the current signal, and the sampling frequency of the speed loop is set to 5kHz to ensure reasonable feedback of the dynamic change of speed. At the same time, the sampling frequency of the position loop is set to 1kHz to achieve high-precision acquisition of position information. The original multi-loop time series data is obtained through the multi-cascade acquisition structure. The current signal, speed signal and position signal in the original multi-loop time series data are respectively subjected to targeted noise reduction and filtering processing to eliminate the noise and unstable factors in the signal. For the current signal, a Butterworth low-pass filter is used for filtering processing, and the cutoff frequency is set to 2kHz to effectively suppress the high-frequency noise caused by the switching frequency and obtain a smooth and stable current signal after noise reduction. For the speed signal, the adaptive window median filter method is used for smoothing. The window width is dynamically adjusted according to the speed change rate, so that the window size varies in the range of 3 to 15 sampling points, thereby effectively removing the discrete noise in the speed signal and maintaining its true dynamic characteristics. At the same time, the improved Kalman filter algorithm is used to optimize the position signal. This method sets the process noise covariance to 0.01mm² and the measurement noise covariance to 0.005mm² to ensure effective denoising of the position signal during the filtering process, while enhancing the signal stability under the influence of mechanical resonance to obtain an accurate position signal. The data mutation point is identified for the denoised current signal, the smoothed speed signal and the accurate position signal. The sliding window method is used for mutation point detection. The window length is set to 200ms, and 3 times the standard deviation is used as the mutation threshold. The mutation point of the signal is identified by calculating the change of the data in the sliding window, and the collected data is divided into different working conditions such as the startup section, steady-state section, load change section and stop section to ensure that the subsequent extraction of characteristic parameters can accurately reflect the actual operating status of the motor. Through this mutation point identification method, the continuity and validity of the data are ensured, so that the feature parameter extraction can focus on the specific operating state. The key feature parameters are extracted from the segmented and marked signal data to form standardized structured parameter data. Among them, in order to reflect the control performance and dynamic response characteristics of the motor, the position overshoot is calculated, which refers to the maximum deviation value of the motor exceeding the target position when executing the position command. This indicator is crucial for evaluating the control effect of the position loop. At the same time, the speed fluctuation rate is calculated, that is, the speed change amplitude per unit time. This parameter can effectively evaluate the stability and anti-disturbance ability of the speed control loop. The current THD value, that is, the total harmonic distortion rate of the current, is calculated. This parameter reflects the quality of the motor current signal and the degree of influence of electromagnetic noise.In order to measure the stability margin of the system, the phase margin is calculated. This parameter represents the phase variation range allowed before the system becomes unstable and is an important indicator for evaluating the stability of the servo system. By extracting these characteristic parameters, standard structured parameter data is formed.

[0013] S2, based on the standard structured parameter data, identify the working condition of the intelligent servo motor and obtain the working condition characteristic parameter set; Specifically, principal component analysis is performed on the standard structured parameter data to reduce data dimensions and remove redundant information, thereby improving computational efficiency and the accuracy of working condition identification. By calculating the covariance matrix of the standard structured parameter data, solving the eigenvalues ​​and eigenvectors, and selecting the principal components with a cumulative contribution rate of more than 95%, the feature data after dimensionality reduction is obtained. This process can retain the main information of the data to the greatest extent while reducing computational complexity. The K-means clustering algorithm is applied to the feature data after dimensionality reduction. The data is classified, and the Euclidean distance of samples within the class is minimized and the distance between classes is maximized to ensure that data of different working conditions can be effectively distinguished. In the clustering process, several cluster centers are randomly initialized, and the Euclidean distance of each data point to each center is calculated, and it is assigned to the nearest cluster center. Then the new cluster center is calculated, and it is continuously iterated and optimized until the convergence condition is met, that is, when the change of the cluster center is less than the set threshold or the maximum number of iterations reaches 500 times, a stable cluster center and clustering result are finally obtained. Based on the obtained clustering results, it is mapped with the predefined high-speed motion conditions, precise positioning conditions, load mutation conditions, anti-disturbance conditions and no-load operation conditions to ensure that the clustering category corresponds to the actual motor working state. By calculating the similarity between the cluster center and the standard working condition feature vector, and setting the similarity threshold to 0.85, when the cosine similarity between the feature vector of a cluster center and the feature vector of a predefined working condition exceeds this threshold, the working condition category corresponding to the class is determined, and a data set with working condition labels is obtained. The data set with working condition labels is decomposed to obtain multi-scale working condition signals. The original signal is decomposed into different frequency bands based on wavelet transform to extract features at different time scales. The db4 wavelet basis is selected and the number of decomposition layers is set to 5 to ensure the effective decomposition of different frequency components such as high speed, low speed, steady state and mutation, so as to capture the changes in the motor operating state at different time scales. Specific sensitive characteristic parameters are extracted from the multi-scale working condition signals for different working conditions to quantify the operating characteristics of each working condition. For high-speed motion conditions, the speed fluctuation rate, current THD value and phase margin are extracted, among which the speed fluctuation rate measures the speed stability of the motor, the current THD value reflects the degree of harmonic distortion of the motor current signal, and the phase margin is used to evaluate the stability margin of the system; for precise positioning conditions, the position overshoot, setting time and static error are extracted, among which the position overshoot represents the maximum deviation of the motor before reaching the target position, the setting time indicates the time required for the system to reach a stable state, and the static error is used to measure the accuracy of the final position of the motor; for load mutation conditions, the torque response time, speed drop rate and current rise time are extracted, among which the torque response time measures the motor's response to load changes. The speed drop rate reflects the impact of load mutation on speed, and the current rise time is used to evaluate the adaptability of motor current to load mutation; for anti-disturbance working conditions, the disturbance suppression rate, recovery time and maximum deviation are extracted, among which the disturbance suppression rate indicates the system's ability to suppress external disturbances, the recovery time is used to measure the time it takes for the system to recover from disturbances to a steady state, and the maximum deviation reflects the maximum fluctuation caused by disturbances; for no-load operating conditions, the no-load speed stability and current effective value are extracted, among which the no-load speed stability indicates the speed fluctuation of the motor under no-load state, and the current effective value is used to evaluate the current consumption of the motor under no-load state.By extracting specific sensitive characteristic parameters for different working conditions, a complete working condition characteristic vector is constructed. According to the working condition characteristic vector, the statistical characteristics are calculated to calculate its mean, standard deviation, kurtosis and skewness, and form a working condition characteristic parameter set. The mean is used to characterize the central trend of each characteristic parameter, the standard deviation is used to measure the degree of dispersion of the data, and the kurtosis and skewness are used to analyze the distribution form of the characteristic data.

[0014] S3, establishing a two-level optimization architecture including a motor parameter layer and a control algorithm layer according to the working condition characteristic parameter set, and generating a dynamic parameter adjustment mapping table; It should be noted that the data related to the load inertia ratio in the working condition characteristic parameter set is linearly mapped and calculated to determine the PWM basic parameter set of the motor parameter layer. The change of the load inertia ratio has an important impact on the dynamic performance and energy efficiency of the motor control. The PWM carrier frequency, dead time and filter coefficient are adjusted according to different inertia ratios. When the load inertia is relatively low, for example, less than 2, a higher carrier frequency is used to reduce torque pulsation and improve the dynamic response capability of the system. When the load inertia is relatively high, for example, greater than 10, the carrier frequency needs to be reduced to reduce switching losses and improve the operating stability of the system. At the same time, the adjustment of the dead time is nonlinearly mapped according to the current size. A larger current requires a smaller dead time to reduce the loss of voltage utilization, while a smaller current requires a larger dead time to reduce switching noise. In order to optimize the smoothness of the motor current signal, the setting of the current feedback filter coefficient is adaptively adjusted based on the bandwidth requirements of different working conditions. For example, the filter coefficient is set to 0.7 under high bandwidth conditions and to 0.9 under low bandwidth conditions, thereby ensuring the stability of the motor operation. Based on the load disturbance observation results in the working condition characteristic parameter set, the adaptive PI controller of the current loop is designed to form a current loop parameter optimization unit. Under working conditions with large load changes or frequent disturbances, high-gain PI parameters are used to improve the response speed of the current loop and reduce the current error. Under steady-state operating conditions, in order to reduce the oscillation of the control system, the gain of the PI parameters is appropriately reduced. An adaptive adjustment method based on the load disturbance observer is adopted to dynamically adjust the proportional coefficient and the integral coefficient according to the results of the disturbance estimation. For example, the range of the proportional coefficient Kp is set between 0.5 and 5.0, the range of the integral coefficient Ki is set between 50 and 500, and the sampling period is set to 50μs to ensure the real-time regulation. In order to reduce the delay of disturbance estimation, the estimation delay is controlled within 100μs to improve the control accuracy and the anti-disturbance ability of the system, thereby ensuring the efficient operation of the current loop. On the basis of completing the current loop optimization, the speed loop controller is constructed according to the speed command change rate data in the working condition characteristic parameter set, and the feedforward compensation structure based on the disturbance observer is adopted to form a speed loop parameter optimization unit. In the case of a large speed command change rate, such as a change rate higher than 500rpm / s, the feedforward compensation coefficient is increased to reduce the impact of the speed command on the control system, and the speed feedforward coefficient is set to 0.8. In the case of a low speed command change rate, such as less than 100rpm / s, the speed feedforward coefficient is set to 0.3 to reduce the overshoot and oscillation of the system. In the PI parameter setting of the speed loop, the range of the proportional coefficient is set between 0.1 and 2.0, and the range of the integral coefficient is set between 5 and 50 to ensure that the system can take into account both dynamic response and steady-state accuracy.At the same time, under high-speed motion conditions, the bandwidth of the speed loop is appropriately increased to improve the responsiveness of the system, while under precise positioning conditions, the bandwidth is reduced to improve the speed stability and control accuracy, thereby realizing dynamic optimization of the speed loop. On the basis of speed loop optimization, based on the position command characteristics in the characteristic parameter set of the working condition, a PI controller with feedforward compensation of the position loop is designed to form a position loop parameter optimization unit. Under point-to-point motion conditions, the position feedforward coefficient is increased to reduce the positioning error of the system, and the position feedforward coefficient is set to 0.8, while under trajectory tracking conditions, the position feedforward coefficient is set to 0.9 to improve the tracking accuracy of the system. At the same time, in order to optimize the dynamic response of the system, the acceleration feedforward compensation coefficient is set, for example, it is set to 0.4 under point-to-point motion conditions and 0.3 under trajectory tracking conditions to reduce the speed fluctuation of the system. The PI control parameters of the position loop also need to be adaptively adjusted, where the range of the proportional coefficient is set between 5 and 50, and the range of the integral coefficient is set between 0 and 10, to ensure that the system can achieve high-precision positioning control under different conditions. After completing the construction of the motor layer parameter adjustment scheme, current loop parameter optimization unit, speed loop parameter optimization unit and position loop parameter optimization unit, these parameters are subjected to multivariate regression analysis and parameter relationship coupling calculation to establish a dynamic parameter adjustment mapping table. By collecting a large amount of experimental data under typical working conditions and using the multivariate regression analysis method, the mapping relationship between the parameters is constructed to ensure the coordinated optimization of the motor layer parameters and the control algorithm layer parameters. For example, when the load inertia ratio changes greatly, the PWM parameters and the current loop PI parameters are adjusted synchronously to avoid the degradation of the system's dynamic performance. At the same time, when the speed command change rate is high, the adjustment of the speed loop feedforward compensation and the position loop feedforward compensation is coordinated to improve the stability of the system. In order to ensure the accuracy of the parameter coupling relationship, based on more than 100 sets of experimental data, the least squares method is used for parameter fitting, and the correlation coefficient between the parameters is calculated to establish an accurate mapping table. Through this dynamic parameter adjustment mapping table, it is ensured that the intelligent servo motor can achieve the optimal control performance under different working conditions and maintain stable operation under complex load environments. Among them, the correlation coefficient is calculated using the Pearson correlation coefficient method to quantify the mutual dependence between the parameters of the motor parameter layer and the control algorithm layer. The specific calculation process is as follows: First, collect more than 100 sets of experimental data under different working conditions. Each set of data contains PWM parameters (carrier frequency, dead time, filter coefficient) and control parameters (position loop, speed loop, current loop PI parameters and feedforward compensation coefficient). For any two parameters X and Y, calculate their correlation coefficient r(X,Y), the formula is r(X,Y)=cov(X,Y) / (σX·σY), where cov(X,Y) is the covariance of X and Y, σX and σY are the standard deviations of X and Y respectively. The covariance is calculated as cov(X,Y)=Σ[(X. i-X̄)(Y i -Ȳ)] / n, where X̄ and Ȳ are sample means and n is the number of samples. The relationship curve between parameters is fitted by the least squares method, and the determination coefficient R is used. 2 Evaluate the goodness of fit, where R 2 =1-SSres / SStot, SSres is the residual sum of squares, SStot is the total deviation sum of squares. For example, when the load inertia ratio changes from 2 to 15, the correlation coefficient between the carrier frequency and the current loop proportional gain is calculated to be -0.87, indicating that the two are strongly negatively correlated, that is, when the load inertia ratio increases, the carrier frequency should be reduced and the current loop proportional gain should be increased simultaneously; the correlation coefficient between the speed loop feedforward coefficient and the position loop feedforward coefficient is 0.92, indicating that the two are strongly positively correlated and need to be coordinated to maintain system stability; the correlation coefficient between the PWM dead time and the current loop integral gain is -0.76, indicating that when the load current increases and the dead time needs to be reduced, the current loop integral gain should be appropriately increased to enhance the anti-interference ability. The construction of the dynamic parameter adjustment mapping table adopts a method combining multivariate regression analysis with parameter space discretization. The specific construction process is as follows: First, the operating condition characteristic parameters (load inertia ratio JL / JM, speed command change rate dω / dt, position command characteristics) are used as independent variables, and the control parameters (PWM parameter set, each loop PI parameter and feedforward compensation coefficient) are used as dependent variables to establish a multiple regression model. For parameter pairs with obvious linear relationships, multiple linear regression Y=β0+β1X1+β2X2+...+β n X n +ε, where Y is the control parameter, X is the operating characteristic parameter, β is the regression coefficient, and ε is the random error; for parameter pairs with significant nonlinear relationships, polynomial regression Y=β0+β1X+β2X is used. 2 +...+β n X n +ε or exponential regression Y=β0·e β1x +ε. Based on the regression model, the operating condition characteristic parameter space is discretized into multiple interval points, the optimal control parameters corresponding to each interval point are calculated, and a lookup table is generated. To improve real-time performance, linear interpolation is used between adjacent discrete points for parameter calculation. For example, under the operating conditions of load inertia ratio JL / JM=5 and speed command change rate dω / dt=300rpm / s, the mapping table gives PWM carrier frequency fc=8kHz, dead time td=2.5μs, current loop proportional gain Kp i =3.2, integral gain Ki i =280, speed loop proportional gain Kp v =0.85, integral gain Ki v =25, speed feedforward coefficient Kff v =0.6, position loop proportional gain Kp p=35, integral gain Ki p =5, position feedforward coefficient Kff p =0.85, acceleration feedforward coefficient Kff a =0.5; when the load inertia ratio increases to JL / JM=10, the mapping table automatically adjusts to fc=6kHz, td=2.8μs, Kp i =4.5, Ki i =320, Kp v =0.65, Ki v =18, Kff v =0.7, Kp p =28, Ki p =3, Kff p =0.9, Kff a =0.65, achieving collaborative optimization of parameters to maintain the dynamic performance of the system.

[0015] S4, training a deep deterministic policy gradient network based on the feature parameter set and the dynamic parameter adjustment mapping table to generate the optimal control strategy model for different working conditions; Specifically, the target state vector and target action vector are constructed to form the basic data representation of reinforcement learning. Among them, the target state vector is composed of the core state information in the working condition characteristic parameter set, including position error, speed error, current error, load estimation value and temperature state, etc. These parameters reflect the real-time operating state of the motor and can effectively characterize the dynamic characteristics of the servo motor under different working conditions. The target action vector is composed of the key control parameters in the dynamic parameter adjustment mapping table, including the PI parameters of the position loop, speed loop and current loop, as well as the feedforward compensation coefficient, etc. These control parameters directly determine the control effect of the servo motor under different working conditions. Therefore, it is necessary to continuously optimize through reinforcement learning methods to obtain the optimal control strategy. On the basis of constructing the target state vector and target action vector, a Markov decision process model is established to form a theoretical framework of reinforcement learning. In this model, the state vector represents the current operating state of the system, the action vector represents the adjustment of the motor parameters by the controller, and the state transition is jointly determined by the motor's dynamic equations and control algorithms. Through this model, the optimization problem of the intelligent servo motor is transformed into a reinforcement learning problem, in which the intelligent agent selects appropriate control parameters under different states to minimize error and energy consumption and maximize the stability and dynamic performance of the system. After constructing the Markov decision process model, a reasonable multi-objective combined reward function is designed to guide the policy optimization in the reinforcement learning process. The reward function needs to consider multiple objectives such as position accuracy, dynamic response, energy consumption and temperature rise at the same time. For precise positioning conditions, the weight of the position accuracy item should be high to ensure that the motor can accurately track the target position, while for high-speed motion conditions, the weight of the dynamic response item should be increased to ensure that the system has a faster response speed when running at high speed. Under long-term operation conditions, the weight of the energy consumption item needs to be increased accordingly to optimize the energy utilization efficiency of the system, while under high-load conditions, the weight of the temperature rise item should be increased to prevent the motor from overheating. Through the multi-objective weight adjustment mechanism, it is ensured that the reinforcement learning model can generate appropriate control strategies under different conditions. After building the reinforcement learning environment, the Actor network is designed to realize the network structure of the deep deterministic policy gradient algorithm. The Actor network adopts a four-layer fully connected structure to ensure that the model can effectively capture the complex mapping relationship between input state and output action. In the first layer of the network, 128 neurons are used to perform preliminary feature extraction on the input state vector, and the LeakyReLU activation function is used to enhance the nonlinear expression capability. In the second and third layers, 256 and 128 neurons are used respectively to further extract deep features and ensure that the network has sufficient learning ability. In the output layer, the Tanh function is used to normalize the output to ensure that the range of control parameters is within a reasonable range. This network structure can generate optimal control actions under different working conditions, and continuously optimize control parameters through deep learning methods to improve system performance.On the basis of constructing the Actor network, an experience replay buffer is designed and a priority sampling strategy is adopted to improve the training efficiency of reinforcement learning. The experience replay buffer is used to store historical interaction data and randomly sample during the training process to break the correlation between data, thereby improving the stability of learning. At the same time, a priority sampling strategy is adopted to make samples with larger TD errors more likely to be selected to speed up the learning speed of the model. In the actual training process, the capacity of the experience replay buffer is set to 100,000 data, and the priority parameter α is initialized to 0.6 to ensure that the model can effectively utilize high-value samples. In the early stage of training, a larger exploration noise is used to increase the diversity of strategies, and in the later stage of training, the noise level is gradually reduced to ensure that the model can converge to the optimal strategy. After completing the construction of the reinforcement learning environment, the network structure of the deep deterministic policy gradient algorithm is pre-trained to accelerate the convergence of the model and improve the initial control performance. In the pre-training stage, the model is run in a high-precision intelligent servo motor simulation environment, and the control strategy is continuously optimized through the reinforcement learning algorithm. At the beginning of training, the learning rate is set to 0.001, and gradually decays to 0.0001 as the training process progresses to ensure the stability of the model in the later stage. At the same time, the exploration noise is initially set to 0.3 and gradually decays to 0.05 to balance the relationship between exploration and utilization. During the entire pre-training process, the training cycle is set to 500,000 steps to ensure that the model can fully learn the dynamic characteristics of the motor in the offline stage and form a relatively stable control strategy. After completing the pre-training, the pre-trained model is applied to the intelligent servo motor and fine-tuned online to improve its control performance in the actual environment. In the online fine-tuning stage, the operating data of the motor is collected in real time, and the control parameters are continuously adjusted based on the reinforcement learning method to optimize the dynamic response and stability of the system. During the fine-tuning process, the learning rate is set to 0.0001 and remains unchanged to ensure the stability of parameter adjustment, while the exploration noise is fixed to 0.05 to reduce disturbances during online training. After every 100 steps of training, the model is evaluated for performance, and the policy network with the best test results is saved to ensure that the final control strategy has the best control effect. Select the optimal control strategy model for different working conditions and apply it to the intelligent servo motor to achieve efficient and intelligent adaptive control.

[0016] The root mean square value of the position loop tracking error is normalized to obtain the position accuracy reward component. The real-time position error of the motor over a period of time is calculated and statistically analyzed using the root mean square formula so that a larger position error corresponds to a smaller reward value, while a smaller error corresponds to a larger reward value. By normalizing the error data, the reward component is ensured to vary within the range of [0,1], so that the reinforcement learning model can measure the control effect under different states with a relatively uniform scale. While constructing the position accuracy reward component, the dynamic response characteristics of the system are comprehensively considered, and the dynamic response reward component is constructed based on the speed response time, overshoot, and volatility. Among them, the speed response time reflects the time it takes for the system to reach the target speed from the initial state, the overshoot indicates the maximum amplitude of the system exceeding the target value, and the volatility measures the speed fluctuation of the system in steady-state operation. In order to construct a reasonable dynamic response reward, these parameters are normalized separately, and the overall reward value is calculated by weighted summation. For example, the standardized range of response time is set so that shorter response time gets higher reward and longer response time gets lower reward. At the same time, nonlinear mapping is performed on overshoot so that smaller overshoot brings higher reward value and larger overshoot reduces reward value. The normalization of fluctuation adopts statistical analysis method to calculate the standard deviation of speed over a period of time and use its reciprocal as reward value to ensure that the system can maintain low speed fluctuation during steady-state operation. While optimizing dynamic performance, the energy efficiency of the system is considered and the energy consumption reward component is calculated based on the input power and output mechanical power of the motor. The input power of the motor is determined by voltage, current and phase angle, while the output mechanical power is calculated by torque and angular velocity. By calculating the ratio of input power to output power, the energy efficiency ratio of the motor is obtained, and further converted into reward value through normalization, so that higher energy efficiency ratio brings higher reward and lower energy efficiency ratio reduces reward value. In order to prevent the system from excessively pursuing energy efficiency and ignoring dynamic performance, a balance factor is introduced in the reward function so that the energy efficiency reward occupies an appropriate weight under specific working conditions without negatively affecting other performance indicators. The temperature rise of the motor during operation is normalized to obtain the temperature rise reward component. The temperature rise is calculated based on the heat transfer model inside the motor. The temperature of the motor winding is monitored in real time and its increase relative to the ambient temperature is calculated. In order to ensure the rationality of this parameter in the reinforcement learning environment, the temperature rise data is normalized and mapped so that a lower temperature rise brings a higher reward, while a higher temperature rise leads to a decrease in the reward value. In order to avoid drastic temperature rise changes in the system in a short period of time, a time smoothing term is introduced in the reward calculation to reduce the impact of sudden temperature rise on the control strategy, thereby ensuring that the system can maintain a stable temperature control level during long-term operation. The combined weight of the reward function is dynamically adjusted based on the working condition identification results to ensure that the optimization objectives under different working conditions can be adaptively adjusted.A weight allocation strategy based on fuzzy rules is adopted to realize the adaptive weight adjustment scheme. According to the output of the working condition identification model, the current working condition is determined to belong to the categories of high-speed motion, precise positioning, load mutation, anti-disturbance or no-load operation, and the membership of each working condition is calculated based on fuzzy logic. For different working condition categories, corresponding weight adjustment rules are defined. For example, in the precise positioning working condition, the weight of the position accuracy reward should be higher, while the weight of the dynamic response reward should be appropriately reduced. In the high-speed motion working condition, the weight of the dynamic response reward should be increased, and the weight of the energy consumption reward should be appropriately reduced. In the load mutation working condition, the indicators related to the dynamic response reward and the anti-disturbance capability should dominate, while in the no-load operation working condition, the energy efficiency and temperature rise should be optimized first to ensure that the system can achieve the optimal control effect under different working conditions. Based on the above adaptive weight adjustment scheme, the position accuracy reward component, the dynamic response reward component, the energy consumption reward component and the temperature rise reward component are weighted and combined to finally construct a reinforcement learning environment. In each time step, the weight of each reward component is determined according to the current working condition identification result, and the total reward value is calculated by weighted summation to ensure that the model can automatically adjust the optimization target under different working conditions. During the reinforcement learning training process, the weight parameters are updated online to adapt to changes in the motor operating environment and optimize the robustness of the control strategy.

[0017] S5, using the optimal control strategy model, automatically switches the control parameters according to the real-time identified working condition type, and realizes the smooth transition control of the intelligent servo motor between different working conditions.

[0018] Among them, a working condition-strategy mapping table is established based on the optimal control strategy model to ensure that the control parameters under different working conditions can be automatically adapted. For the five working conditions of high-speed motion, precision positioning, load mutation, anti-disturbance and no-load operation, five groups of basic control strategies are preset, and a basic control strategy set is constructed to call the corresponding optimal control parameters under different working conditions. Under high-speed motion conditions, a control strategy with feedforward compensation as the main and PI adjustment as the auxiliary is adopted, in which the position feedforward coefficient is set to 0.9, the speed feedforward coefficient is set to 0.9, and the acceleration feedforward coefficient is set to 0.7 to ensure that the motor maintains accurate trajectory tracking capabilities during high-speed operation. Under precise positioning conditions, a control strategy with PI adjustment as the main and feedforward compensation as the auxiliary is adopted, in which the position loop proportional coefficient is increased by 50% and the integral coefficient is reduced by 30% to ensure that the system achieves stable positioning control under low-speed and high-precision conditions. Under load mutation conditions, a fast response control strategy based on a disturbance observer is adopted, in which the bandwidth of the disturbance observer is set to 500Hz and the feedforward compensation coefficient is set to 0.5, to ensure that the system can quickly adjust the current loop parameters when the load changes suddenly and reduce overshoot. Under anti-disturbance conditions, a robust control strategy based on a sliding mode observer is adopted, in which the sliding mode observer gain is set to 200 and the reaching law coefficient is set to 20, to improve the stability and anti-interference ability of the system under disturbance conditions. Under no-load operation conditions, a low-gain energy-saving control strategy is adopted, in which the current loop bandwidth is reduced by 30% and the position loop proportional coefficient is reduced by 40% to reduce energy consumption during no-load operation. A parameter gradual update algorithm is designed for the basic control strategy set to ensure that the switching of control parameters between different working conditions will not cause system oscillation or instability. A smooth switching mechanism is adopted to make the update of control parameters in a gradual manner, and the maximum change rate is controlled to not exceed 5% during the parameter update cycle to ensure that the system maintains a smooth transition during switching. In order to solve the problem of composite working conditions at the boundary of multiple working conditions, a fuzzy logic method based on Gaussian function is introduced. By calculating the fuzzy membership of each working condition, the degree of belonging of the current system in different working condition categories is determined, and a multi-condition fusion control strategy is generated by weighted fusion. For example, when the system is in the boundary area of ​​high-speed motion and precise positioning, the weights of the high-speed motion strategy and the precise positioning strategy are calculated based on the membership, and the final control parameters are calculated by weighted average method to ensure that the system can smoothly transition between different working conditions without causing system instability due to sudden parameter adjustments. After the multi-condition fusion control strategy is generated, it is evaluated in real time and a series of comprehensive performance indicators are calculated to judge the adaptability of the current control strategy. Among them, the main evaluation indicators include the root mean square value of position tracking error, speed fluctuation standard deviation, current THD, response time and steady-state error. According to the comparison results of these indicators with the preset thresholds, it is judged whether the current strategy needs to be adjusted. If the comprehensive performance index is lower than the preset optimal threshold, the fine-tuning process is triggered to further optimize the control parameters.For example, when the RMS value of the position tracking error exceeds the set range, the position loop PI parameters are adjusted to optimize the control accuracy; when the speed fluctuation standard deviation is high, the speed loop feedforward compensation parameters are increased to reduce the speed fluctuation of the system; when the current THD value is too high, the PWM parameters and filter coefficients are optimized to reduce the current harmonic distortion, thereby improving the stability and efficiency of the motor operation. In the fine-tuning process, a local optimization strategy for control parameters based on the gradient descent method is adopted to ensure that the control parameters can be adjusted in the optimal direction. The direction of parameter adjustment is determined by calculating the gradient change of the comprehensive performance index, and the optimization update is performed with a step size of 5%. At the same time, the maximum number of iterations is set to 10 times to ensure that the parameter adjustment converges within a reasonable range. In addition, in order to improve the adaptability of the control strategy, a dynamic step size adjustment mechanism is adopted, so that when the error is large, the step size is large, and when the error is small, the step size is reduced, so as to improve the optimization efficiency and avoid parameter oscillation. After the optimized control parameter calculation is completed, it is input into the control system of the intelligent servo motor, and the control signal is output to the IGBT power module through the DSP controller to achieve the final control execution. The DSP controller adjusts key control variables such as PWM duty cycle, current loop PI parameters, speed loop feedforward compensation coefficient, and position loop gain in real time according to the optimized parameters, and ensures the real-time and accuracy of the control signal through high-speed computing capabilities. The IGBT power module adjusts the current, voltage, and torque of the drive motor according to the output signal of the DSP controller to achieve precise control of the motor's operating state and ensure smooth transition between different working conditions. Through this complete control strategy, it is ensured that the intelligent servo motor always maintains high-performance operation under complex working conditions, and achieves smooth transition between different working conditions, avoiding oscillation and control instability caused by sudden changes, thereby improving the overall intelligence level and operating efficiency of the servo motor system.

[0019] In one example, multi-dimensional data collection and preprocessing are performed on the position loop, speed loop, and current loop of the intelligent servo motor to obtain standard structured parameter data, including: A three-level cascade acquisition structure of position loop, speed loop and current loop is established for the intelligent servo motor, and synchronous sampling is performed to obtain the original multi-loop timing data; Performing filtering processing on the current signal in the original multi-loop time series data to obtain a noise-reduced current signal, performing median filtering on the speed signal in the original multi-loop time series data to obtain a smoothed speed signal, and performing Kalman filtering on the position signal in the original multi-loop time series data to obtain an accurate position signal; The data mutation points are identified on the noise-reduced current signal, the smoothed speed signal and the precise position signal to obtain segmented marked signal data; Characteristic parameters of position overshoot, speed fluctuation rate, current THD value and phase margin are extracted from the segmented marked signal data to obtain standard structured parameter data.

[0020] In this example, a high-precision data acquisition system is designed to enable it to accurately collect signals from different loops and ensure that the sampling times of all loops are strictly aligned. The sampling frequency of the current loop is set to 20kHz and the sampling frequency of the speed loop is set to , the sampling frequency of the position loop is , to ensure high-precision acquisition of highly dynamic signals, while taking into account the system's computational burden and real-time performance. The synchronous trigger mechanism is used to strictly align the sampling times of all loops to ensure the timing consistency of the data. Assume that the collected current signal is The speed signal is , the position signal is , then the original multi-loop time series data is represented as a time series matrix: in, Indicates At each sampling moment, the corresponding current, speed and position signal values ​​are , and The original signal is filtered to ensure the reliability and accuracy of the signal. In order to process the current signal in the original multi-loop time series data, a Butterworth low-pass filter is used to filter it to remove high-frequency noise and smooth the signal. Assume that the filtered current signal is , the filtering process is described by the following equation: in, is the transfer function of the filter, is the cut-off angular frequency, is the order of the filter. By properly choosing 2000rad / s, so that the high-frequency components in the signal can be suppressed and the current signal after noise reduction can be obtained At the same time, in order to process the speed signal, the adaptive window median filtering method is used to eliminate the mutation noise and improve the smoothness of the signal. The window width is defined as , then the filtered speed signal Calculated by the following formula: Among them, the window width Adaptive adjustment based on speed change rate. For example, when the speed change rate is large, select a smaller (such as 3 sampling points), and when the speed change rate is small, choose a larger (such as 15 sampling points) to balance signal smoothness and dynamic response performance. For the position signal, Kalman filtering is used for optimization to reduce measurement errors and enhance the system's anti-interference ability. Assume that the true value of the position signal is , the measured value is , then the state update equation of the Kalman filter is: in, is the estimated position signal, is the Kalman gain, For the measurement matrix, the Kalman gain is calculated as follows: in, is the forecast error covariance matrix, To measure the noise covariance matrix, choose To ensure the system's good suppression ability for small position disturbances, thereby obtaining accurate position signals After completing the signal filtering process, data mutation point identification is performed to identify key state changes and perform data segmentation. The sliding window method is used to calculate the standard deviation change rate of the signal, and the mutation threshold is set to 3 times the standard deviation, that is: in, Represents the signal value at a certain moment, is the sliding window width, is the standard deviation within the window. When the set threshold is exceeded, it is determined that a data mutation has occurred, and the signal is segmented and marked to obtain the segmented signal data. After obtaining the segmented and marked signal data, characteristic parameters such as position overshoot, speed fluctuation rate, current THD value and phase margin are extracted to construct standard structured parameter data. Position overshoot Defined as: in, is the target position value, which reflects the dynamic response capability of the system. Defined as the standard deviation of velocity variation: in, is the steady-state speed average, which measures the stability of the system speed. The total harmonic distortion (THD) of the current is calculated as: in, The current signal The harmonic components, is the fundamental current, which is used to evaluate the quality of the current signal. It is defined as: in, The closed-loop transfer function of the system at the cut-off frequency The phase angle at which the system is stabilized is measured.

[0021] In one example, based on standard structured parameter data, the working condition of the intelligent servo motor is identified to obtain a working condition characteristic parameter set, including: Perform principal component analysis on the standard structured parameter data to obtain the feature data after dimension reduction; Apply K-means clustering algorithm to the feature data after dimension reduction, and obtain cluster centers and clustering results by minimizing the Euclidean distance of samples within a class and maximizing the distance between classes. The clustering results are mapped with the predefined high-speed motion conditions, precise positioning conditions, load mutation conditions, anti-disturbance conditions and no-load operation conditions to obtain a data set with condition labels. Perform signal decomposition on the data set with working condition labels to obtain multi-scale working condition signals; Specific sensitive characteristic parameters are extracted from multi-scale working condition signals for different working conditions: speed fluctuation rate, current THD value and phase margin are extracted for high-speed motion conditions; position overshoot, setting time and static error are extracted for precision positioning conditions; torque response time, speed drop rate and current rise time are extracted for load mutation conditions; disturbance suppression rate, recovery time and maximum deviation are extracted for anti-disturbance conditions; no-load speed stability and current effective value are extracted for no-load operation conditions to obtain working condition characteristic vectors; The statistical characteristics are calculated according to the operating condition characteristic vector to obtain the operating condition characteristic parameter set.

[0022] In this example, a standard data matrix is ​​constructed, which contains multiple key parameters during motor operation, such as position error, speed fluctuation rate, current THD value, phase margin, torque response time, steady-state error, etc. Assume that the original standard structured parameter data matrix is: in, Indicates The first eigenvalues, the number of rows in the matrix Indicates the number of samples and columns of different working conditions In order to remove redundant information between features and reduce computational complexity, principal component analysis is performed by calculating the covariance matrix: in, is the mean vector of all data. Then calculate the eigenvalues ​​of the covariance matrix and the eigenvector , and select the principal component with a cumulative contribution rate of 95% to reduce the data dimension. The feature data after dimensionality reduction is expressed as: in, is the selected eigenvector matrix, is the feature data after dimensionality reduction. After obtaining the feature data after dimensionality reduction, the K-means clustering algorithm is applied to classify the working conditions in a way that minimizes the Euclidean distance of samples within a class and maximizes the distance between classes. Define the cluster center And initialize Random center points are collected and then the cluster centers are updated using an iterative optimization method so that each sample is assigned to the nearest center: in, Indicates that it belongs to After completing the iteration, we get The clustering results are mapped with the predefined working condition categories to identify five working conditions: high-speed motion, precise positioning, load mutation, anti-disturbance and no-load operation. The standard feature vector for each working condition is defined. ,in, Represents the typical characteristic value of a certain working condition. By calculating the cluster center Vector with standard condition The similarity between: When the similarity threshold is higher than the set value (such as 0.85), the cluster category is mapped to the corresponding working condition category to obtain a data set with working condition labels. After the working condition label is determined, the data is decomposed to obtain multi-scale working condition signals. Wavelet transform is used for signal decomposition, using Daubechies 4 (db4) wavelet basis function and setting 5 layers of decomposition to convert the original signal Decomposed into different frequency bands: in, is the decomposition coefficient, is the wavelet basis function. By selecting decomposition signals of different scales, different time characteristics of working condition changes are extracted. Based on multi-scale signals, specific sensitive characteristic parameters are extracted for different working conditions. For example, under high-speed motion conditions, the speed fluctuation rate is calculated: in, For the The speed value at the moment, is the average speed, which is used to evaluate the speed stability of the system. At the same time, the total harmonic distortion (THD) of the current is calculated: in, For the The harmonic components, is the fundamental component, which is used to evaluate the quality of the current signal. For precise positioning conditions, calculate the position overshoot: in, is the filtered position signal, is the target position value. Calculate the settling time , which is the shortest time for the system to reach the steady-state error range. Under the load mutation condition, the torque response time is extracted: in, For the moment The torque, is the steady-state torque value, Measures the response speed of the motor when the load changes suddenly. Under the anti-disturbance condition, calculate the disturbance suppression rate: in, is the steady-state position value, which is used to measure the system's ability to resist external disturbances. Under no-load operation, calculate the no-load speed stability: in, The steady-state speed is used to evaluate the speed fluctuation of the system under no-load conditions. At the same time, the effective value of the no-load current is calculated. : in, is a current signal, which is used to evaluate the energy consumption characteristics of the system. All calculated operating condition characteristic parameters are integrated into the operating condition characteristic parameter set, and the mean standard deviation, kurtosis and skewness of each characteristic are calculated by statistical methods to characterize the characteristic distribution under different operating conditions.

[0023] In one example, according to the operating condition characteristic parameter set, a two-level optimization architecture including a motor parameter layer and a control algorithm layer is established, and a dynamic parameter adjustment mapping table is generated, including: Perform linear mapping calculation on the load inertia ratio data in the working condition characteristic parameter set to obtain the motor parameter layer PWM basic parameter set, and perform nonlinear mapping based on the motor parameter layer PWM basic parameter set and the current size data in the working condition characteristic parameter set to obtain the motor layer parameter adjustment plan; Based on the load disturbance observation results in the working condition characteristic parameter set, the current loop adaptive PI controller is designed to obtain the current loop parameter optimization unit; A speed loop controller is constructed based on the speed command change rate data in the working condition characteristic parameter set, and a speed loop parameter optimization unit is obtained by using a feedforward compensation structure based on a disturbance observer. Based on the position command characteristics in the working condition characteristic parameter set, a PI controller with feedforward compensation for the position loop is designed to obtain a position loop parameter optimization unit. Multivariate regression analysis and parameter relationship coupling calculation are performed on the motor layer parameter adjustment scheme, current loop parameter optimization unit, speed loop parameter optimization unit and position loop parameter optimization unit to obtain a dynamic parameter adjustment mapping table.

[0024] In this example, a linear mapping calculation is performed on the load inertia ratio data to determine the PWM basic parameter set of the motor parameter layer. Assume that the load inertia ratio is defined as in , represents the load inertia, Represents the motor inertia. In order to optimize the PWM control strategy of the motor drive, the carrier frequency is adjusted according to the inertia ratio. , Dead time and current feedback filter coefficient The adjustment of the carrier frequency needs to be lowered at high inertia ratios to reduce switching losses, and increased at low inertia ratios to reduce torque ripple, so it is expressed as: in, and are the maximum and minimum carrier frequencies, respectively, and are the maximum and minimum ranges of the load inertia ratio, respectively. Similarly, the dead time According to the current Nonlinear mapping is performed to ensure that the dead time is reduced in the case of high current and the voltage utilization is improved, while the dead time is increased in the case of low current to reduce the switching noise. The settings are as follows: in, and Represent the minimum and maximum dead time respectively, is the adjustment factor, current The actual load current mean value is taken from the operating condition characteristic parameter set. In order to optimize the filtering characteristics of the current signal, the current feedback filter coefficient It needs to be matched with the motor bandwidth requirement and calculated by the following formula: in, and are the minimum and maximum filter coefficients respectively. After completing the calculation of the basic PWM parameters, the adaptive PI controller of the current loop is designed to optimize the current control performance. The optimization of the current loop needs to be based on the load disturbance observation results. Assume that the disturbance observation value of the current loop is ,in The current loop proportional gain is calculated from the actual current fluctuation and integral gain Dynamic adjustment is required to suppress the impact of disturbances. The calculation formula is as follows: in, and are the maximum values ​​of the current loop proportional and integral gain, respectively. and It is the adjustment factor used to control the gain change amplitude with disturbance. After the current loop optimization is completed, based on the speed command change rate Construct a speed loop controller and adopt a feedforward compensation structure based on disturbance observer to improve the responsiveness of speed control. Set the proportional gain of the speed loop PI controller to and integral gain It needs to be adjusted according to the speed command change rate, and its optimization model is as follows: in, and is the adjustment range of the speed loop proportional gain, is the maximum value of the speed loop integral gain, is the adjustment factor. At the same time, the feedforward compensation term based on the disturbance observer The calculation is as follows: in, represents the velocity disturbance, Represents the speed time constant of the system. This feedforward compensation term can automatically adjust the control parameters of the speed loop when the disturbance is large to reduce the impact of the disturbance. After the speed loop is optimized, the PI controller with feedforward compensation of the position loop is designed according to the characteristics of the position command to optimize the precise positioning capability of the motor. Set the acceleration feedforward coefficient of the position command to and position feedforward coefficient It needs to be adjusted according to the working conditions, and the optimization formula is as follows: in, Represents the acceleration of the position command, is the maximum acceleration, is the target location, is the current position signal. The feedforward compensation structure can automatically adjust under different trajectory tracking conditions to ensure a balance between accuracy and response speed.

[0025] In one example, a deep deterministic policy gradient network is trained based on a feature parameter set and a dynamic parameter adjustment mapping table to generate an optimal control strategy model for different working conditions, including: The working condition characteristic parameter set is used to construct a target state vector including position error, speed error, current error, load estimation value and temperature state, and the parameters in the dynamic parameter adjustment mapping table are used to construct a target action vector including PI parameters of the position loop, speed loop and current loop and feedforward compensation coefficient; Construct a Markov decision process model based on the target state vector and the target action vector; Design a multi-objective combined reward function for the Markov decision process model to obtain a reinforcement learning environment; Based on the reinforcement learning environment, the Actor network is constructed. The Actor network adopts a four-layer fully connected structure to obtain the deep deterministic policy gradient algorithm network structure; Design an experience replay buffer, adopt a priority sampling strategy, obtain a training data sampling scheme, and pre-train the deep deterministic policy gradient algorithm network structure on the intelligent servo motor to obtain a pre-trained model; The pre-trained model is applied to the intelligent servo motor for online fine-tuning, and the optimal control strategy model for different working conditions is selected.

[0026] In this example, a target state vector is constructed based on the operating condition characteristic parameter set. The state vector contains key variables that can describe the operating state of the servo motor, including position error. , speed error , Current Error , load estimation And temperature status Among them, the position error Reflects the deviation between the current position of the motor and the target position, speed error Describes the speed deviation of the motor, while the current error, while the current error It is used to evaluate the difference between the current motor current and the target current. Load estimation Reflects the load-to-inertia ratio of the motor, while the temperature state Represents the temperature of the motor winding or drive to prevent overheating from affecting the control performance. Therefore, the target state vector is expressed as: While constructing the target state vector, the parameters in the mapping table are adjusted according to the dynamic parameters to construct the target action vector, which contains the PI parameters and feedforward compensation coefficients of the position loop, speed loop, and current loop to optimize the motor control performance. and integral gain Dynamic adjustment is required to optimize the position tracking accuracy and the proportional gain of the velocity loop and integral gain Need to adapt to different speed change rates, and the proportional gain of the current loop and integral gain It is necessary to ensure the stability of the motor current. Feedforward compensation coefficient Dynamic optimization is also required to enhance the responsiveness of the system. Therefore, the target action vector is expressed as: After constructing the target state vector and target action vector, a Markov decision process model is constructed to describe the state transition process of the servo motor. Assume that the motor is in state Take Action Then, transfer to the next state , and get rewards , then the state transition process is expressed as: in, Represents the dynamic equation of the motor, is the noise term of the system. In order to optimize the control strategy, a multi-objective combined reward function is designed, which includes position accuracy reward, dynamic response reward, energy consumption reward and temperature rise reward. Position accuracy reward Calculated by the normalized RMS value of the position error to ensure that the motor can accurately track the target position; dynamic response reward Calculated by speed fluctuation, overshoot and response time to optimize the dynamic characteristics of the motor; energy consumption bonus Calculated by input electrical power and output mechanical power to reduce energy consumption; Temperature rise bonus It is calculated by the temperature change of the motor winding to prevent overheating from affecting the stability of the system. Therefore, the total reward function is expressed as: in, Fuzzy logic is used to adaptively adjust the weights of each reward to ensure that the optimization objectives under different working conditions can change adaptively. After building the reinforcement learning environment, the Actor network is further constructed to realize the training of the deep deterministic policy gradient (DDPG) algorithm. The Actor network adopts a four-layer fully connected structure, in which the input layer is the target state vector, the first hidden layer and the second hidden layer contain 128 and 256 neurons respectively, and the LeakyReLU activation function is used to enhance the nonlinear expression ability, while the third hidden layer contains 128 neurons to further extract deep features. The final output layer uses the Tanh function to ensure that the range of the output action meets the physical constraints. Therefore, the calculation of the Actor network is expressed as: in, Represents the strategy function of the Actor network, parameters Optimization is performed using the gradient descent method. In order to improve the training efficiency of reinforcement learning, an experience replay buffer is designed and a priority sampling strategy is adopted. The experience replay buffer is used to store historical interaction data and perform random sampling during training to break the correlation between data. The priority sampling strategy makes samples with larger TD errors more likely to be selected to speed up the learning of the model. Assume that the sample set in the experience buffer is , where each sample contains state, action, reward and next state information, then the probability of priority sampling is calculated as: in, Represents the priority of the sample, is the importance parameter of control priority. This strategy can effectively improve the training efficiency and reduce the oscillation problem during the training process. After completing the construction of the reinforcement learning environment, the deep deterministic policy gradient algorithm network is pre-trained to accelerate the model convergence and improve the initial control performance. In the pre-training stage, the model is first run in a high-precision servo motor simulation environment, and the control strategy is continuously optimized through the reinforcement learning algorithm. In the early stage of training, the learning rate is set to 0.001, and gradually decays to 0.0001 as the training process proceeds to ensure the stability of the model in the later stage. At the same time, the exploration noise is initially set to 0.3 and gradually decays to 0.05 to balance the relationship between exploration and utilization. During the entire pre-training process, the training cycle is set to 500,000 steps to ensure that the model can fully learn the dynamic characteristics of the motor in the offline stage and form a relatively stable control strategy. After completing the pre-training, the pre-trained model is applied to the intelligent servo motor and fine-tuned online to improve its control performance in the actual environment. In the online fine-tuning stage, the motor's operating data is collected in real time, and the control parameters are continuously adjusted based on the reinforcement learning method to optimize the dynamic response and stability of the system. During the fine-tuning process, the learning rate is set to 0.0001 and remains unchanged to ensure the stability of parameter adjustment, while the exploration noise is fixed to 0.05 to reduce disturbances during online training. After every 100 steps of training, the model is evaluated for performance and the policy network with the best test results is saved to ensure that the final control strategy has the best control effect. The optimal control strategy model for different working conditions is selected and applied to the intelligent servo motor to achieve efficient and intelligent adaptive control. For example, in a high-speed motion condition, if the speed error of the initial control strategy is large, the speed loop PI parameters and feedforward compensation coefficients will be automatically adjusted during the fine-tuning process to improve the tracking accuracy of the system, while reducing energy consumption and ensuring that the motor always maintains high performance under complex load environments.

[0027] In one example, a multi-objective combined reward function is designed for a Markov decision process model to obtain a reinforcement learning environment, including: The root mean square value of the position loop tracking error is normalized to obtain the position accuracy reward component, and the dynamic response reward component is constructed based on the speed response time, overshoot and fluctuation; The energy consumption bonus component is calculated based on the motor input power and output mechanical power, and the motor operating temperature rise is normalized to obtain the temperature rise bonus component; The combined weight of the reward function is dynamically adjusted according to the working condition identification results, and a weight allocation strategy based on fuzzy rules is adopted to obtain an adaptive weight adjustment scheme; Based on the adaptive weight adjustment scheme, the position accuracy reward component, dynamic response reward component, energy consumption reward component and temperature rise reward component are weightedly combined to obtain a reinforcement learning environment.

[0028] In this example, the RMS value of the position loop tracking error is normalized to obtain the position accuracy bonus component. Position Loop Tracking Error is the actual position of the motor With target location The root mean square (RMS) value of the deviation between in, is the length of an evaluation cycle. In order to ensure that the reward values ​​under different working conditions have a uniform scale, the root mean square error is normalized so that its value is always in the range of [0,1], so that the position accuracy reward component It is expressed as: in, is the preset maximum allowable position error. Approaching zero, Close to 1, it means the position control accuracy is high, and when When larger, The decrease indicates that the positioning accuracy of the servo motor is poor. After obtaining the position accuracy bonus component, the dynamic response bonus component is constructed based on the speed response time, overshoot and fluctuation. Speed ​​response time It represents the time required for the system to reach the steady-state speed from the initial state, and the overshoot It measures the maximum overshoot amplitude of the system before reaching the target speed, the speed fluctuation Reflects the speed fluctuation degree of the system in steady-state operation. Dynamic response reward component It is necessary to ensure fast response while minimizing overshoot and fluctuation, so it is expressed as: in, They are the preset maximum response time, maximum overshoot and maximum speed fluctuation. is the weight coefficient to ensure that the contribution of different indicators to the reward meets the control requirements. For example, in high-speed motion conditions, should be smaller, and and It needs to be increased appropriately to suppress overshoot and fluctuation. In addition to the dynamic response, the energy consumption bonus component needs to be calculated to ensure that the system reduces energy consumption while achieving good control performance. Motor input power By bus voltage and phase current Calculate the output mechanical power The motor torque and angular velocity Calculate, so the energy efficiency is expressed as: In order to enable the reinforcement learning model to optimize energy consumption, an energy consumption reward component is constructed , which is expressed as follows: in, is the theoretical maximum energy conversion efficiency. If the motor operates in the high efficiency range, then Close to 1, it means the system energy efficiency is good, while when the motor energy loss is high, The decrease indicates that the system's operating efficiency is low and the control strategy needs to be optimized. The motor operating temperature rise is normalized to obtain the temperature rise bonus component. The motor winding temperature With ambient temperature In order to ensure that the motor maintains a stable temperature during long-term operation, the following temperature rise bonus components are constructed: in, is the maximum temperature rise allowed by the motor. If the temperature rise is lower, If the temperature is close to 1, it indicates that the system is operating at a normal temperature. If the temperature is close to or exceeds the limit, When the load drops to zero, the reinforcement learning model will adjust the control strategy to reduce the motor load or optimize the PWM parameters to reduce heat generation. In order to enable reinforcement learning to adapt to different working conditions and dynamically adjust the contribution ratio of each reward component, an adaptive weight adjustment scheme is designed. This scheme uses fuzzy logic to allocate weights based on the working condition identification results, so that the optimization target can be adaptively adjusted under different working conditions. Define the working condition vector ,in Respectively represent the attention to position accuracy, dynamic response, energy consumption and temperature rise under different working conditions, and the weights are calculated by fuzzy rules , so that: in, , the sum of weights always satisfies . In the case of precision positioning, is set to the maximum value, and Appropriately reduce, so that the control strategy mainly optimizes the position accuracy; under high-speed motion conditions, is set to the maximum value, and Appropriately reduce to optimize the dynamic response characteristics; under long-term operation conditions, Improve to ensure temperature rise control; in energy-saving conditions, Improve to optimize energy efficiency. Combined with the adaptive weight adjustment scheme, all reward components are weighted combined to obtain the comprehensive reward function in the reinforcement learning environment: This reward function guides the agent to adjust the control strategy during the reinforcement learning process so that the servo motor can achieve optimal performance under different working conditions.

[0029] In one example, the optimal control strategy model is used to automatically switch control parameters according to the type of working condition identified in real time, so as to achieve smooth transition control of the intelligent servo motor between different working conditions, including: Based on the optimal control strategy model, a working condition-strategy mapping table is established, and five groups of basic control strategies are preset for five working conditions: high-speed motion, precise positioning, load mutation, anti-disturbance and no-load operation, to obtain a basic control strategy set; A parameter gradient update algorithm is designed for the basic control strategy set to obtain a smooth switching mechanism. The fuzzy membership of each working condition is calculated through a Gaussian function for the composite working condition at the boundary of multiple working conditions to obtain a multi-condition fusion control strategy. The multi-condition fusion control strategy calculates comprehensive performance indicators including the RMS value of position tracking error, speed fluctuation standard deviation, current THD, response time and steady-state error. When the comprehensive performance indicator is lower than the preset threshold, the fine-tuning process is triggered to obtain the performance evaluation result. Based on the performance evaluation results, the control parameters are locally optimized by the gradient descent method to obtain the optimized control parameters; The optimized control parameters are input into the control system of the intelligent servo motor, and the control signal is output to the IGBT power module through the DSP controller to achieve smooth transition control of the intelligent servo motor between different working conditions.

[0030] In this example, a working condition-strategy mapping table is established based on the optimal control strategy model, and basic control strategies are preset for five working conditions: high-speed motion, precise positioning, load mutation, anti-disturbance, and no-load operation. The characteristic parameters of different working conditions are defined, and the working condition-strategy correspondence is established in combination with the optimal control strategy model. Assume that the working condition characteristic parameter set is ,in Represents position error represents the velocity fluctuation rate, represents the current harmonic distortion (THD), represents the temperature rise, Represents the load inertia ratio. Then, by calculating the similarity between the working condition characteristic parameters and the basic control strategy, a working condition-strategy mapping table is established, so that: in, represents the optimal control strategy for the current working conditions, Representative The characteristic vector of the preset basic control strategy. When the calculated optimal strategy has the highest matching degree with the preset working condition, the strategy is selected as the initial control parameter and the basic control strategy set is constructed. After the basic control strategy is determined, in order to avoid the sudden switching between different working conditions causing unstable motor operation, the parameter gradual update algorithm is adopted to ensure the smooth transition of the control parameters. Assume that the current control parameter vector is , the target control parameter vector is , then the parameter update process is expressed as: in, is the gradient step coefficient, whose size determines the adjustment rate of the parameters, and is set to To ensure a smooth transition. When the system is in the boundary area of ​​multiple working conditions, the fuzzy logic based on Gaussian function is used to calculate the fuzzy membership of each working condition. : in, Representative The standard deviation of the same working condition is used to measure the discrete degree of the working condition characteristics. The multi-condition fusion control strategy is calculated by weighted average : This method ensures that when the system is at the boundary of different working conditions, the control parameters can be adjusted smoothly to avoid instability caused by mutations. After executing the multi-condition fusion control strategy, the comprehensive performance index of the system is calculated in real time to judge the effectiveness of the current control strategy. Assume that the comprehensive performance index vector is ,in is the RMS value of the position error, is the speed fluctuation standard deviation, is the current harmonic distortion, is the response time, is the steady-state error, then the comprehensive performance evaluation formula is: in, Represents the weights of different performance indicators, and uses fuzzy logic method to adaptively adjust to meet the needs of different working conditions. Exceeding the preset threshold When , the fine-tuning process is triggered to optimize the control parameters. The fine-tuning process uses the gradient descent method to perform local optimization of the control parameters. Assuming that the optimization goal is to minimize the comprehensive performance error, the optimization problem is expressed as: The control parameters are updated using the gradient descent method: in, is the learning rate, set to , is the gradient of the performance index with respect to the control parameter, calculated by numerical differentiation: This method ensures that the control parameters can be optimized in the optimal direction when adjusted, avoiding system instability caused by blind adjustment. The optimized control parameters are input into the control system of the intelligent servo motor, and the control signal output is executed through the DSP controller. The DSP controller adjusts key variables such as PWM duty cycle, current loop Pl parameters, speed loop feedforward compensation coefficient and position loop gain in real time according to the optimized parameters, and ensures the real-time and accuracy of the control signal with high-speed computing power. The IGBT power module adjusts the current, voltage and torque of the drive motor according to the output signal of the DSP controller, thereby realizing smooth transition control between different working conditions.

[0031] The present embodiment also includes: according to the working condition characteristic parameter set, the temperature drift characteristics of the intelligent servo motor are monitored in real time, the dynamic response curves at different temperatures are obtained through step test and frequency sweep test, a temperature-parameter mapping model is established, and a temperature drift characteristic matrix is ​​obtained; the nonlinear characteristics of the magnetic field of the intelligent servo motor are identified, and the degree of magnetic flux saturation and magnetic resistance torque fluctuation are identified by injecting excitation signals of different frequencies and recording the current-torque curve, and the magnetic field nonlinear compensation coefficient is obtained by using an adaptive particle filter algorithm; based on the temperature drift characteristic matrix and the magnetic field nonlinear compensation coefficient, the encoder error of the intelligent servo motor is accurately measured, the system error and the random error are distinguished by the autocorrelation analysis method, and an error compensation function including position periodic error terms, speed-related error terms and acceleration-related error terms is constructed to obtain the encoder error compensation amount; the joint flexibility characteristics of the intelligent servo motor are modeled, the elastic deformation coefficient is identified through the resonant frequency test, a joint flexibility mathematical model including a stiffness matrix, a damping matrix and an inertia matrix is ​​established, and a flexible vibration suppression parameter set is obtained; based on the temperature drift characteristic matrix, the magnetic field nonlinear compensation coefficient, the encoder error compensation amount and the flexible vibration The dynamic suppression parameter set is used to build a comprehensive model of the individual characteristics of the intelligent servo motor. The Bayesian method is used to fuse the control data collected under high-speed cutting, precision positioning and load mutation conditions to obtain a comprehensive characteristic compensation model; the Markov chain Monte Carlo method is used to estimate the parameters of the comprehensive characteristic compensation model, and the sampling chain length is set to 10000, the convergence threshold is 0.001, and the transfer probability matrix of different working conditions is calculated to obtain the posterior distribution of the parameters, and the maximum posterior probability estimate is extracted as the optimal parameter set; the optimal parameter set is combined with the working condition characteristic parameter set, and the compensation control module is used to calculate the parameters of the comprehensive characteristic compensation model. The servo motor control signal is corrected in real time, the dynamic adjustment range of temperature drift compensation is ±3%, the adjustment range of magnetic field nonlinear compensation is ±5%, the encoder error compensation resolution reaches 0.1 arc second, and the joint flexible vibration suppression bandwidth is set to 10Hz-100Hz; a sub-micron positioning accuracy predictor is constructed based on the comprehensive characteristic compensation model, and the recursive least squares method is used to update the model parameters online. The prediction window length is set to 200ms, and the prediction step is 5ms. The temperature, magnetic field, encoder and flexible characteristic parameters are combined to predict the future position deviation and perform feedforward compensation.

[0032] Reference Figure 2 , this embodiment provides an intelligent servo motor control device, including: The acquisition module 1 is used for performing multi-dimensional data acquisition and preprocessing on the position loop, speed loop and current loop of the intelligent servo motor to obtain standard structured parameter data; Identification module 2, used for identifying the working condition of the intelligent servo motor based on standard structured parameter data, and obtaining a working condition characteristic parameter set; Establishing module 3, for establishing a two-level optimization architecture including a motor parameter layer and a control algorithm layer according to the working condition characteristic parameter set, and generating a dynamic parameter adjustment mapping table; Training module 4, used for training a deep deterministic policy gradient network based on a feature parameter set and a dynamic parameter adjustment mapping table to generate an optimal control strategy model for different working conditions; The switching module 5 is used to utilize the optimal control strategy model to automatically switch the control parameters according to the working condition type identified in real time, so as to realize the smooth transition control of the intelligent servo motor between different working conditions.

[0033] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0034] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0035] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0036] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0037] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An intelligent servo motor control method, characterized in that: The following steps are involved: Perform multi-dimensional data acquisition and preprocessing on the position loop, speed loop and current loop of the intelligent servo motor to obtain standard structured parameter data; Based on the standard structured parameter data, identifying the working condition of the intelligent servo motor, and obtaining a working condition characteristic parameter set; According to the operating condition characteristic parameter set, a two-level optimization architecture including a motor parameter layer and a control algorithm layer is established, and a dynamic parameter adjustment mapping table is generated; Training a deep deterministic policy gradient network based on the characteristic parameter set and the dynamic parameter adjustment mapping table to generate an optimal control strategy model for different working conditions; By utilizing the optimal control strategy model, control parameters are automatically switched according to the working condition type identified in real time, so as to realize smooth transition control of the intelligent servo motor between different working conditions.

2. The intelligent servo motor control method according to claim 1, characterized in that: The multi-dimensional data collection and preprocessing of the position loop, speed loop and current loop of the intelligent servo motor are performed to obtain standard structured parameter data, including: A three-level cascade acquisition structure of position loop, speed loop and current loop is established for the intelligent servo motor, and synchronous sampling is performed to obtain original multi-loop timing data; Performing filtering processing on the current signal in the original multi-loop time series data to obtain a noise-reduced current signal, performing median filtering on the speed signal in the original multi-loop time series data to obtain a smoothed speed signal, and performing Kalman filtering on the position signal in the original multi-loop time series data to obtain an accurate position signal; Performing data mutation point identification on the noise-reduced current signal, the smoothed speed signal and the precise position signal to obtain segmented marked signal data; Characteristic parameters of position overshoot, speed fluctuation rate, current THD value and phase margin are extracted from the segmented marked signal data to obtain standard structured parameter data.

3. The intelligent servo motor control method according to claim 1, characterized in that: The method of identifying the working condition of the intelligent servo motor based on the standard structured parameter data and obtaining a working condition characteristic parameter set includes: Performing principal component analysis on the standard structured parameter data to obtain feature data after dimension reduction; Applying the K-means clustering algorithm to the feature data after dimension reduction, and obtaining cluster centers and clustering results by minimizing the Euclidean distance of samples within a class and maximizing the distance between classes; Mapping the clustering results with predefined high-speed motion conditions, precise positioning conditions, load mutation conditions, anti-disturbance conditions, and no-load operation conditions to obtain a data set with condition labels; Performing signal decomposition on the data set with the working condition labels to obtain a multi-scale working condition signal; Specific sensitive characteristic parameters are extracted from the multi-scale working condition signals for different working conditions: speed fluctuation rate, current THD value and phase margin are extracted for high-speed motion conditions; position overshoot, setting time and static error are extracted for precise positioning conditions; torque response time, speed drop rate and current rise time are extracted for load mutation conditions; disturbance suppression rate, recovery time and maximum deviation are extracted for anti-disturbance conditions; no-load speed stability and current effective value are extracted for no-load operation conditions to obtain working condition characteristic vectors; Statistical characteristics are calculated according to the operating condition characteristic vector to obtain an operating condition characteristic parameter set.

4. The intelligent servo motor control method according to claim 1, characterized in that: According to the operating condition characteristic parameter set, a two-level optimization architecture including a motor parameter layer and a control algorithm layer is established, and a dynamic parameter adjustment mapping table is generated, including: Performing linear mapping calculation on the load inertia ratio data in the operating condition characteristic parameter set to obtain a motor parameter layer PWM basic parameter set, and performing nonlinear mapping according to the motor parameter layer PWM basic parameter set and the current magnitude data in the operating condition characteristic parameter set to obtain a motor layer parameter adjustment scheme; Designing a current loop adaptive PI controller based on the load disturbance observation results in the operating condition characteristic parameter set to obtain a current loop parameter optimization unit; A speed loop controller is constructed according to the speed command change rate data in the operating condition characteristic parameter set, and a speed loop parameter optimization unit is obtained by adopting a feedforward compensation structure based on a disturbance observer; Design a PI controller with feedforward compensation for the position loop based on the position command characteristics in the working condition characteristic parameter set to obtain a position loop parameter optimization unit; The motor layer parameter adjustment scheme, the current loop parameter optimization unit, the speed loop parameter optimization unit and the position loop parameter optimization unit are subjected to multivariate regression analysis and parameter relationship coupling calculation to obtain a dynamic parameter adjustment mapping table.

5. The intelligent servo motor control method according to claim 1, characterized in that: The training of a deep deterministic policy gradient network based on the characteristic parameter set and the dynamic parameter adjustment mapping table to generate an optimal control strategy model for different working conditions includes: The working condition characteristic parameter set is used to construct a target state vector including position error, speed error, current error, load estimation value and temperature state, and the parameters in the dynamic parameter adjustment mapping table are used to construct a target action vector including PI parameters of position loop, speed loop and current loop and feedforward compensation coefficient; Constructing a Markov decision process model based on the target state vector and the target action vector; Designing a multi-objective combined reward function for the Markov decision process model to obtain a reinforcement learning environment; An Actor network is constructed based on the reinforcement learning environment, wherein the Actor network adopts a four-layer fully connected structure to obtain a deep deterministic policy gradient algorithm network structure; Designing an experience replay buffer, adopting a priority sampling strategy, obtaining a training data sampling scheme, and pre-training the deep deterministic policy gradient algorithm network structure on an intelligent servo motor to obtain a pre-trained model; The pre-trained model is applied to the intelligent servo motor for online fine-tuning, and the optimal control strategy model for different working conditions is selected.

6. The intelligent servo motor control method according to claim 5, characterized in that: The step of designing a multi-objective combined reward function for the Markov decision process model to obtain a reinforcement learning environment includes: The root mean square value of the position loop tracking error is normalized to obtain the position accuracy reward component, and the dynamic response reward component is constructed based on the speed response time, overshoot and fluctuation; The energy consumption bonus component is calculated based on the motor input power and output mechanical power, and the motor operating temperature rise is normalized to obtain the temperature rise bonus component; The combined weight of the reward function is dynamically adjusted according to the working condition identification results, and a weight allocation strategy based on fuzzy rules is adopted to obtain an adaptive weight adjustment scheme; Based on the adaptive weight adjustment scheme, the position accuracy reward component, the dynamic response reward component, the energy consumption reward component and the temperature rise reward component are weightedly combined to obtain a reinforcement learning environment.

7. The intelligent servo motor control method according to claim 1, characterized in that: The optimal control strategy model is used to automatically switch control parameters according to the working condition type identified in real time to achieve smooth transition control of the intelligent servo motor between different working conditions, including: Based on the optimal control strategy model, a working condition-strategy mapping table is established, and five groups of basic control strategies are preset for five working conditions: high-speed motion, precise positioning, load mutation, anti-disturbance and no-load operation, to obtain a basic control strategy set; A parameter gradual update algorithm is designed for the basic control strategy set to obtain a smooth switching mechanism, and the fuzzy membership of each working condition is calculated by a Gaussian function for a composite working condition at the boundary of multiple working conditions to obtain a multi-condition fusion control strategy; The multi-operating condition fusion control strategy calculates comprehensive performance indicators including the root mean square value of position tracking error, speed fluctuation standard deviation, current THD, response time and steady-state error, and when the comprehensive performance indicator is lower than a preset threshold, a fine-tuning process is triggered to obtain a performance evaluation result; Based on the performance evaluation results, the control parameters are locally optimized by a gradient descent method to obtain optimized control parameters; The optimized control parameters are input into the control system of the intelligent servo motor, and the control signal is output to the IGBT power module through the DSP controller to achieve smooth transition control of the intelligent servo motor between different working conditions.

8. An intelligent servo motor control device, characterized in that: For implementing the steps of the method according to any one of claims 1 to 7, the intelligent servo motor control device comprises: The acquisition module is used to collect and preprocess multi-dimensional data of the position loop, speed loop and current loop of the intelligent servo motor to obtain standard structured parameter data; An identification module, used to identify the working condition of the intelligent servo motor based on the standard structured parameter data, and obtain a working condition characteristic parameter set; An establishment module is used to establish a two-level optimization architecture including a motor parameter layer and a control algorithm layer according to the operating condition characteristic parameter set, and generate a dynamic parameter adjustment mapping table; A training module, used for training a deep deterministic policy gradient network based on the characteristic parameter set and the dynamic parameter adjustment mapping table to generate an optimal control strategy model for different working conditions; The switching module is used to utilize the optimal control strategy model to automatically switch control parameters according to the working condition type identified in real time, so as to realize smooth transition control of the intelligent servo motor between different working conditions.

9. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

Images