An AI-based intelligent servo motor driver and control method

Abstract

This invention discloses an AI-based intelligent servo motor driver and control method, belonging to the fields of servo motor control and industrial artificial intelligence technology. The method includes the following steps: S1, system initialization, completing hardware self-test, parameter identification, digital twin modeling, and AI model deployment; S2, acquiring and preprocessing multi-physical quantity signals to achieve twin mapping, residual whitening, and feature weighting enhancement; S3, performing AI multi-loop adaptive control to complete trajectory planning, closed-loop adjustment, and global energy efficiency optimization; S4, executing edge incremental learning, constructing a dataset according to trigger conditions, and completing model updates with stability constraints; S5, achieving AI fault prediction and RUL evaluation, performing hierarchical early warning and adaptive fault-tolerant control; S6, achieving multi-axis state synchronization and high-precision collaborative control through multi-agent reinforcement learning. This invention improves the response speed, control accuracy, anti-disturbance capability, operational reliability, and energy efficiency of servo systems.

Description

Technical Field

[0001] This invention relates to the fields of servo motor control and industrial artificial intelligence technology, and in particular to an AI-based intelligent servo motor driver and control method. Background Technology

[0002] Servo motor drivers are core control components of high-end equipment, and their control performance directly determines the equipment's motion accuracy, dynamic response capability, and operational reliability. With the rapid development of the high-end manufacturing industry, increasingly higher requirements are being placed on the servo driver's adaptability, disturbance rejection performance, reliability, energy efficiency, and multi-axis collaborative accuracy.

[0003] In existing technologies, traditional servo drives generally adopt a fixed-parameter PID three-closed-loop control architecture. However, due to the nonlinear, strongly coupled, and time-varying inherent characteristics of permanent magnet synchronous motors, as well as disturbances such as load changes, nonlinear friction, and external shocks, their adaptive adjustment capabilities are severely insufficient. It is difficult to achieve optimal control across the entire operating range. In high-speed, high-dynamic, and wide-load-fluctuation scenarios, problems such as overshoot, steady-state error, and response lag are prone to occur, limiting control accuracy and dynamic performance.

[0004] Most existing servo drive solutions incorporating AI technology employ offline-trained neural network models, which can only cover a limited set of preset operating conditions. When the actual operating conditions, load characteristics, motor parameters, and training scenarios differ from those of the training scenarios, the model's adaptability drops sharply, and control performance deteriorates significantly. At the same time, existing solutions cannot achieve secure and stable online learning at the edge, and the model update process lacks strict stability theory constraints, which can easily affect control real-time performance and even lead to the risk of system instability, making it difficult to achieve full lifecycle iterative optimization of the model.

[0005] Furthermore, existing servo drives have a single dimension of state perception, only collecting basic current, voltage, and encoder position signals. The coupling characteristics of multiple physical quantity signals are not fully utilized, and high-value features are easily drowned out by noise. They cannot comprehensively perceive the full-dimensional operating status and health characteristics of the motor and drive. Fault diagnosis is mostly a post-fault alarm after the fault occurs, which cannot achieve early prediction and health management. Moreover, there is a lack of effective adaptive fault-tolerant control mechanisms. Single point faults can easily lead to sudden system shutdowns, which cannot meet the continuous and high-reliability operation requirements of high-end manufacturing scenarios. Summary of the Invention

[0006] The purpose of this invention is to provide an AI-based intelligent servo motor driver and control method to overcome the shortcomings of existing technologies, such as weak adaptive anti-disturbance capability of servo drivers, poor online adaptability of AI models, lack of stability guarantee for online learning, insufficient fault prediction and fault tolerance capabilities, and poor energy efficiency optimization.

[0007] To achieve the above objectives, the present invention provides an AI-based intelligent servo motor driver and control method, comprising the following steps: S1. System initialization and offline pre-trained model deployment, completing hardware self-check, motor body parameter identification, digital twin baseline model construction and lightweight AI model edge deployment; S2, Multi-physical quantity full-dimensional state perception and real-time feature extraction, complete multi-source signal synchronous acquisition, signal preprocessing and feature engineering, real-time state mapping of digital twin model based on coupled mutual information entropy, residual whitening and feature weighting enhancement; S3. AI-based multi-loop composite adaptive real-time control, which sequentially completes outer loop intelligent trajectory planning, speed / position loop adaptive composite control, current loop model prediction considering multiple loss coupling, and AI global energy efficiency optimization control; S4. Edge-end online incremental self-learning and dynamic iterative optimization of the model, completing the judgment of learning trigger conditions, construction of incremental datasets, incremental learning with Lyapunov stability hard constraints and safe model updates; S5. AI-based full lifecycle fault prediction and adaptive fault-tolerant control, completing fault feature identification and RUL prediction, fault level classification and early warning, and graded fault-tolerant control strategy execution. S6. Multi-axis cooperative intelligent synchronous control based on multi-agent reinforcement learning to achieve multi-axis state synchronization interaction and cooperative control optimization.

[0008] Preferably, S1 specifically includes the following steps: S11. Hardware self-test and offline calibration of motor body parameters: Complete the power-on self-test of power unit, sampling unit, encoder unit and dual-core control unit, and complete the offline identification of motor stator resistance, quadrature and direct axis inductance, back electromotive force coefficient, moment of inertia and viscous friction coefficient through recursive least squares method to obtain the basic parameters of motor body. S12. Construction of digital twin baseline model for motor-driver-load: Based on the identified basic parameters, a high-precision digital twin model including electromagnetic model, mechanical dynamics model, thermal loss model and fault evolution model is built as the baseline for system simulation verification and data generation. S13. Offline pre-training and lightweight deployment of AI composite control model system: A training dataset covering all operating conditions is generated through a digital twin model to complete the offline pre-training of the AI ​​composite control model system. The AI ​​composite control model system includes a DRL trajectory planning model, an LSTM-attention mechanism disturbance compensation model, an adaptive PID parameter optimization model, a DQN energy efficiency optimization model, a Transformer-PHM fault prediction model, and a MARL multi-axis collaborative model. The pre-trained model is pruned and lightweighted using INT8 quantization to adapt to edge computing power and deployed to a dedicated AI inference core.

[0009] Preferably, S2 specifically includes the following steps: S21. Multi-source signal high-frequency synchronous acquisition: The real-time core uses a current loop sampling frequency of 16kHz or higher to synchronously acquire three-phase stator current, bus voltage, encoder position / speed / acceleration signals, IGBT junction temperature, motor stator winding temperature, driver and motor triaxial vibration acceleration signals, noise signals, and bus ripple voltage. All acquired signals are bound to a synchronous timestamp, and the time synchronization error is ≤1μs. S22. Signal preprocessing and multi-dimensional feature extraction: Kalman filtering and amplitude normalization are performed on the original acquired signal to extract time-domain features, frequency-domain features and time-frequency-domain features respectively. The extracted features are divided into a control feature set for real-time control and a health feature set for health management. S23. Real-time mapping of digital twin model, adaptive whitening of residuals and feature weighting enhancement: The actual operating state parameters collected are input into the digital twin baseline model, and the simulation state values ​​of the corresponding working conditions are output. The state residuals between the actual values ​​and the simulation values ​​are calculated. The weight coefficients of each state quantity are calculated based on the mutual information entropy of multi-physical quantity coupling, and the residuals are adaptively whitened to eliminate autocorrelation interference. The control feature set and health feature set are weighted and enhanced based on the weight coefficients. At the same time, the parameters of the digital twin model are updated based on the whitened residuals, and the whitened residuals are used as the core basis for disturbance judgment and fault precursor identification.

[0010] Preferably, S3 specifically includes the following steps: S31, Outer Loop Intelligent Trajectory Planning and Setpoint Optimization: Receives target position / velocity commands from the host computer, and based on the DRL trajectory planning model, combined with the current motor operating status, load status, and system constraint thresholds, outputs a smooth S-shaped acceleration / deceleration setpoint curve without impact, avoiding overshoot and mechanical shock caused by step commands. S32, Adaptive Composite Control of Velocity Loop and Position Loop: Based on the deviation between the given value and the actual value, the rate of change of the deviation, and the weighted enhanced control feature set, the proportional, integral, and derivative parameters of the PID are adjusted online in real time through an adaptive PID parameter optimization model; at the same time, through an LSTM-attention mechanism disturbance compensation model, the total disturbance value caused by load mutation, nonlinear friction, and parameter drift is predicted in real time, and the disturbance compensation amount is fed forward to the control output to offset the disturbance effect; the real-time core is responsible for the hard real-time calculation of the PID closed loop, and the AI ​​inference core is responsible for the parallel calculation of parameter optimization and disturbance feedforward amount, with a total control delay ≤10μs; S33. Current Loop Model Predictive Control and AI-Enhanced Global Optimal Torque Control: Model predictive control (MPC) is used as the core controller of the current loop. Based on the motor electromagnetic model, current setpoint, actual current value, and bus voltage, it predicts the current trajectory for the next three control cycles and outputs the optimal PWM duty cycle signal. At the same time, a global loss model considering magnetic saturation cross-coupling and multi-loss coupling is constructed. Through the DQN energy efficiency optimization model, combined with the current speed, load torque, winding temperature, and magnetic saturation state, the optimal AC and DC axis current setpoints and IGBT switching frequencies are solved in real time. Dynamic optimal control of the maximum torque-current ratio and maximum torque-voltage ratio is achieved across the entire operating range, minimizing copper losses, iron losses, inverter switching losses, and dead-zone losses.

[0011] Preferably, S4 specifically includes the following steps: S41. Online learning trigger condition judgment: The system monitors the running status in real time and starts online incremental learning when any trigger condition is met. The trigger conditions include: control performance index exceeding the threshold, digital twin model whitening residual exceeding the threshold, significant change in operating conditions, and running time reaching the preset self-learning cycle. S42. Construction of Incremental Learning Dataset: Collect system input and output data, weighted and enhanced state features, and control effect data under the current working conditions. Combine these with the extended simulation data generated by the digital twin model under the same working conditions to construct an incremental learning dataset. Complete data cleaning, deduplication, and normalization. S43. Edge-end Incremental Learning and Model Safety Update with Lyapunov Stability Constraints: Construct a global Lyapunov function for the closed-loop control of the servo system, derive the stability boundary constraints of the AI ​​control model output, and use this constraint as the regularization term of the incremental learning loss function; the AI ​​inference kernel uses the PWM dead time and the system idle time slots without sampling periods to incrementally fine-tune the AI ​​model based on the incremental dataset, freezes the model's backbone feature extraction network, and only fine-tunes the fully connected output layer; the updated model is first simulated and verified in a digital twin model. After verifying that the control performance is better than the current running model and meets the stability constraints, a disturbanceless switching mechanism is used to smoothly replace the online running model. The switching process has no control shock or system instability risk.

[0012] Preferably, S5 specifically includes the following steps: S51. Fault Feature Identification and Remaining Service Prediction: The weighted and enhanced health feature set and the whitened residual of the digital twin model are input into the Transformer-PHM fault prediction model to identify fault precursor features such as encoder abnormality, IGBT aging, winding insulation degradation, bearing wear, and permanent magnet demagnetization in real time, and simultaneously predict the remaining service life (RUL) and fault occurrence probability of core components. S52. Fault Level Classification and Early Warning: Based on fault probability, RUL and whitening residual, the system health status is divided into four levels: healthy status, minor anomaly, moderate fault and severe fault. Corresponding early warning signals are output for different levels. When there is a minor anomaly, online self-learning optimization of control parameters is triggered. S53, Hierarchical Adaptive Fault-Tolerant Control: For moderate faults, automatically switch to the corresponding fault-tolerant control mode: When the encoder signal is abnormal, start the AI ​​sensorless observer to estimate the rotor position and speed in real time based on current, voltage, and vibration signals to maintain normal system operation; when a single current sampling fault occurs, switch to the redundant sampling channel and compensate the sampling value through the model observer; when a single IGBT tube is aging, execute the derating operation strategy to optimize the switching frequency and reduce thermal stress; for severe faults, trigger the safety shutdown procedure to avoid equipment damage.

[0013] Preferably, S6 specifically includes the following steps: S61, Multi-axis status information synchronization and interaction: Real-time industrial Ethernet is used to complete the synchronous acquisition of status information of each drive in the multi-axis system, with a global time synchronization accuracy of ≤100ns. S62, MARL Multi-Axis Cooperative Synchronization Control: Employing a multi-agent reinforcement learning MARL model, each axis's driver is treated as an independent agent. The global optimization objectives are minimizing multi-axis trajectory synchronization error, minimizing system vibration, and minimizing overall energy consumption. The control output of each axis is optimized in real time to compensate for synchronization errors caused by communication delays, mechanical lag, and differences in inter-axis characteristics, achieving high-precision multi-axis nanosecond-level cooperative control.

[0014] The present invention also provides an AI-based intelligent servo motor driver, including a heterogeneous dual-core control unit, a multi-physical quantity signal acquisition unit, a power drive unit, a safety redundancy unit, a multi-protocol real-time communication unit, and a storage unit; The heterogeneous dual-core control unit includes a hard real-time control core and a dedicated AI inference core. The hard real-time control core is used to perform hard real-time closed-loop control of the current loop, speed loop, and position loop. The dedicated AI inference core is used to perform online inference, incremental learning, and data processing of the AI ​​model. The two cores communicate in parallel through a high-speed internal bus. The multi-physical quantity signal acquisition unit includes an electrical signal acquisition module, a position sensing module, a temperature acquisition module, and a vibration and noise acquisition module, which are used to complete the synchronous acquisition of the full-dimensional status signals of the driver and the motor. The power drive unit includes a rectifier module, an IPM intelligent power module, and a buffer and protection circuit, which are used to perform power drive and hardware protection of the motor. The safety redundancy unit includes dual-channel redundant sampling channels, dual-channel encoder interfaces, and emergency stop and safety torque shutdown (STO) circuits, which are used to realize system redundancy protection and fault-tolerant control hardware support. The multi-protocol real-time communication unit supports EtherCAT, PROFINET IRT, and EtherNet / IP real-time industrial Ethernet protocols, enabling high-speed synchronous communication with host computers and multi-axis systems.

[0015] Preferably, the hard real-time control core adopts an ARM Cortex-R5F dual-core architecture with a main frequency of ≥800MHz and a control cycle of up to 62.5μs; the dedicated AI inference core adopts a RISC-V architecture with NPU acceleration, with an AI computing power of ≥4TOPS, and supports model pruning, quantization and incremental learning.

[0016] Preferably, the multi-physical quantity signal acquisition unit adopts a 16-bit or higher synchronous sampling ADC with a sampling frequency ≥16kHz, all acquisition channels are hardware synchronized, and the time synchronization error is ≤1μs.

[0017] Therefore, the beneficial effects of the above-mentioned AI-based intelligent servo motor driver and control method of the present invention are as follows: (1) The present invention significantly improves the control performance and anti-disturbance capability of the servo system: through the AI ​​composite adaptive control architecture and feature weighting enhancement mechanism, compared with the traditional fixed parameter PID control, the system response speed and anti-load disturbance capability are significantly improved, and the control without overshoot and without steady-state error is achieved in the entire working range, meeting the high precision and high dynamic control requirements of high-end equipment.

[0018] (2) This invention solves the industry pain points of poor online adaptability and lack of stability guarantee of traditional AI servo models: through digital twin baseline model, edge incremental online learning and Lyapunov stability hard constraint mechanism, the full life cycle safe iteration of AI model is realized. It can adapt to different loads, working conditions and motor parameter drift scenarios, theoretically guaranteeing the system closed-loop stability of the entire online learning process, and completely solving the risk of instability of industrial AI control.

[0019] (3) The present invention achieves optimal energy efficiency control under all operating conditions: through the multi-loss coupled optimal torque control algorithm, compared with the traditional fixed MTPA control, the overall energy consumption of the system is reduced, the torque control accuracy is improved, the operating cost of the equipment is significantly reduced, and it is in line with the development trend of green manufacturing.

[0020] (4) The present invention significantly improves the system's operational reliability and continuous operation capability: through full-dimensional state perception, residual whitening processing and Transformer-PHM fault prediction model, the fault prediction time can be effectively advanced and the fault identification accuracy can be improved; through the adaptive fault-tolerant control mechanism, smooth degradation operation under moderate faults is realized, meeting the continuous operation requirements of high-end manufacturing scenarios.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the structure of an AI-based intelligent servo motor driver according to the present invention; Figure 2 This is a schematic diagram of a control method for an AI-based intelligent servo motor driver according to the present invention. Detailed Implementation

[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0025] Example 1: like Figure 1 As shown, this embodiment provides an AI-based intelligent servo motor driver, including a heterogeneous dual-core control unit, a multi-physical quantity signal acquisition unit, a power drive unit, a safety redundancy unit, a multi-protocol real-time communication unit, and a storage unit.

[0026] The heterogeneous dual-core control unit comprises a hard real-time control core and a dedicated AI inference core. The hard real-time control core uses an ARM Cortex-R5F dual-core architecture with a clock speed ≥800MHz, used for hard real-time closed-loop control of the current loop, speed loop, and position loop, with a minimum control cycle of 62.5μs. The dedicated AI inference core uses a RISC-V architecture with NPU acceleration, providing AI computing power ≥4TOPS, supporting model pruning, quantization, and incremental learning, used for online inference, incremental learning, and data processing of AI models. The dual cores communicate in parallel via a high-speed internal bus, with a data interaction latency ≤1μs, balancing the hard real-time requirements of control with the computing power demands of AI computation.

[0027] Multi-physical quantity signal acquisition unit: including electrical signal acquisition module, position sensing module, temperature acquisition module, vibration and noise acquisition module, adopting 16-bit or higher synchronous sampling ADC, sampling frequency ≥16kHz, all acquisition channels are hardware synchronized, time synchronization error ≤1μs, used to complete the synchronous acquisition of all-dimensional status signals of driver and motor.

[0028] Power drive unit: includes rectifier module, IPM intelligent power module (using Infineon IPM intelligent power module FF450R12ME4, supporting 1200V / 450A rated parameters), buffer and protection circuit, supports overcurrent, overvoltage, overtemperature and short circuit hardware protection, used to perform motor power drive and hardware-level safety protection.

[0029] Safety redundancy unit: includes dual redundant sampling channels, dual encoder interfaces, and emergency stop and safety torque shutdown (STO) circuits, meeting the SIL3 safety level, and is used to provide hardware support for system redundancy protection and fault-tolerant control.

[0030] Multi-protocol real-time communication unit: Supports EtherCAT, PROFINET IRT, and EtherNet / IP real-time industrial Ethernet protocols, with a communication cycle of 125μs, used to achieve high-speed synchronous communication with host computers and multi-axis systems.

[0031] Storage unit: Includes 8MB of Flash and 2MB of RAM, used to store system programs, AI model parameters, running data and fault logs.

[0032] like Figure 2 As shown, the control method for the above-mentioned intelligent servo motor driver includes the following steps: S1. System initialization and offline pre-trained model deployment: This involves completing hardware self-checks, motor body parameter identification, digital twin baseline model construction, and lightweight AI model deployment at the edge. Specifically, this includes the following steps: S11. Hardware self-test and offline calibration of motor body parameters: After the system is powered on, it completes the full-link power-on self-test of the power drive unit, multi-physical quantity signal acquisition unit, encoder unit, and heterogeneous dual-core control unit. After the self-test is passed, the offline identification of the motor stator resistance, quadrature and direct axis inductance, back electromotive force coefficient, moment of inertia and viscous friction coefficient is completed by recursive least squares method to obtain the basic parameters of the motor body. At the same time, sampling channel offset calibration and encoder zero point calibration are completed.

[0033] S12. Construction of Digital Twin Baseline Model for Motor-Driver-Load: Based on the identified basic parameters of the motor body, combined with the hardware parameters of the driver and the mechanical characteristics of the load, a high-precision digital twin baseline model is built. The model includes the electromagnetic model of the motor, the mechanical dynamics model, the thermal loss model, and the fault evolution model. It can accurately simulate the operating state, loss characteristics and fault evolution law of the motor and driver under all working conditions, and serve as the core baseline for system simulation verification, training data generation and state residual calculation.

[0034] S13. Offline pre-training and lightweight deployment of AI composite control model system: Through the digital twin baseline model, a massive training dataset covering different load, speed, temperature, disturbance and fault scenarios is generated to complete the offline pre-training of the AI ​​composite control model system.

[0035] The AI ​​composite control model system includes a DRL deep reinforcement learning trajectory planning model, an LSTM-attention mechanism perturbation compensation model, a fuzzy neural network adaptive PID parameter optimization model, a DQN deep Q network energy efficiency optimization model, a Transformer-PHM fault prediction model, and a MARL multi-agent reinforcement learning multi-axis collaborative model. The pre-trained model is subjected to structured pruning and INT8 quantization to reduce its size and computational load, adapting to the computing power and memory limitations of the edge AI inference core. Finally, the lightweight model is deployed to the dedicated storage area of ​​the AI ​​inference core.

[0036] S2. Multi-physical quantity full-dimensional state perception and real-time feature extraction, completing multi-source signal synchronous acquisition, signal preprocessing and feature engineering, real-time state mapping of digital twin model based on coupled mutual information entropy, residual whitening and feature weighting enhancement, specifically including the following steps: S21. High-frequency synchronous acquisition of multi-source signals: The hard real-time control core triggers the multi-physical quantity signal acquisition unit to complete the hardware synchronous acquisition of multi-source signals at a current loop sampling frequency of 16kHz or higher. The acquired signals include: three-phase stator current, bus voltage, encoder position / speed / acceleration signal, IGBT junction temperature, motor stator winding temperature, driver and motor triaxial vibration acceleration signal, noise signal, and bus ripple voltage. All acquired signals are bound to a global synchronization timestamp, with a time synchronization error ≤1μs, ensuring the consistency of timing data.

[0037] S22. Signal Preprocessing and Multi-Dimensional Feature Extraction: Kalman filtering is applied to the original acquired signal to eliminate sensor noise and electromagnetic interference; amplitude normalization is performed on the filtered signal to unify the data input range; time-domain features (mean, variance, peak value, kurtosis, margin), frequency-domain features (feature frequency and harmonic amplitude extracted by FFT), and time-frequency-domain features (transient disturbance and impact features extracted by wavelet transform) are extracted respectively; the extracted features are divided into control feature set and health feature set.

[0038] S23. Real-time mapping of digital twin model, adaptive whitening of residuals, and feature weighting enhancement: The actual operating state parameters collected are input into the digital twin baseline model in real time, and the simulated state values ​​under the corresponding operating conditions are output. The state residuals between the actual values ​​and the simulated values ​​are calculated. The weight coefficients of each state quantity are calculated based on the mutual information entropy of multi-physical quantity coupling. The degree of coupling between each physical quantity and the core output (position / torque) of the system is quantified based on the mutual information entropy, eliminating the weight deviation caused by the difference in dimensions. The calculation formula is as follows: In the formula, For the first i The coupling weights of each physical quantity For the first i A physical quantity signal With system output Mutual information entropy, Let be the total number of physical quantities collected, and let the weights satisfy . .

[0039] The formula for calculating mutual information entropy is: In the formula: for and The joint probability density, , They are respectively and The marginal probability density.

[0040] The weighted residual sequence is whitened to eliminate autocorrelation and avoid misjudgment caused by noise. The calculation formula is as follows: In the formula: for k The whitened residual vector at time step 1. For the weighted residual vector, For the first i The original residuals of each physical quantity; W The whitening matrix is ​​derived from the weighted residual covariance matrix. eigenvalue decomposition yields R =UΛU T ,but W =UΛ −1 / 2 This ensures that the covariance matrix of the residuals after whitening is the identity matrix, thus eliminating autocorrelation interference.

[0041] The control feature set and health feature set are weighted and enhanced based on weight coefficients to improve the AI ​​model's ability to focus on high-value features. The calculation formula is as follows: In the formula, This is the weighted and enhanced feature matrix. The original feature matrix, This is a diagonal weight matrix, where the diagonal elements are the coupling weights of the corresponding physical quantities. Simultaneously, the parameters of the digital twin model are updated based on the whitening residual, and the whitening residual is used as the core basis for disturbance judgment and fault precursor identification.

[0042] S3. AI-based multi-loop composite adaptive real-time control, which sequentially completes outer loop intelligent trajectory planning, speed / position loop adaptive composite control, current loop model prediction considering multiple loss coupling, and AI global energy efficiency optimization control, specifically including the following steps: S31, Outer Loop Intelligent Trajectory Planning and Setpoint Optimization: Receives target position / velocity commands from the host computer, and based on a pre-trained DRL trajectory planning model, combines the current motor operating status, load status, and system constraint thresholds (maximum acceleration, maximum torque, vibration threshold, temperature rise threshold) to output a smooth S-shaped acceleration / deceleration setpoint curve without impact. This avoids overshoot, mechanical shock, and noise caused by traditional step commands, while optimizing the acceleration / deceleration characteristics of the trajectory to reduce invalid travel and energy loss.

[0043] S32. Adaptive Composite Control of Velocity and Position Loops: Employing a composite control architecture of "adaptive PID + feedforward disturbance compensation," based on the deviation between the given and actual values, the rate of change of the deviation, and the weighted enhanced control feature set, a fuzzy neural network adaptive PID parameter optimization model is used to adjust the three core parameters of the PID—proportional, integral, and derivative—in real-time to adapt to the system's dynamic characteristics under different operating conditions. Simultaneously, an LSTM-attention mechanism disturbance compensation model focuses on disturbance features with high impact weights, predicting in real-time the total disturbance value caused by load mutations, nonlinear friction, parameter drift, and external shocks, and feeding the disturbance compensation amount forward to the control output to preemptively offset the impact of disturbances on the system. The hard real-time control core is responsible for the hard real-time calculation of the PID closed loop, while the AI ​​inference core is responsible for the parallel calculation of parameter optimization and disturbance feedforward, achieving a total control delay ≤10μs and realizing a fast response with no overshoot and no steady-state error under all operating conditions.

[0044] S33. Current Loop Model Predictive Control and AI-Enhanced Global Optimal Torque Control: Model predictive control (MPC) is used as the core controller of the current loop. Based on the motor electromagnetic model, current setpoint, actual current value, and bus voltage, it predicts the current trajectory for the next three control cycles, solves the optimal control sequence within control constraints, and outputs the optimal PWM duty cycle signal. At the same time, a global loss model considering magnetic saturation cross-coupling and multi-loss coupling is constructed. Through the DQN energy efficiency optimization model, combined with the current speed, load torque, winding temperature, and magnetic saturation state, the optimal quadrature and direct axis current setpoints and IGBT switching frequencies are solved in real time. Dynamic and smooth switching and optimal control of maximum torque-current ratio, maximum torque-voltage ratio, and field weakening control are achieved across the entire operating range, minimizing motor copper losses, iron losses, driver switching losses, and dead zone losses.

[0045] S4. Online incremental self-learning and dynamic iterative optimization of the model at the edge: This involves determining learning trigger conditions, constructing incremental datasets, and performing incremental learning and safe model updates with hard constraints on Lyapunov stability. Specifically, it includes the following steps: S41. Online learning trigger condition judgment: The system monitors the operating status and control performance in real time. When any trigger condition is met, the online incremental learning process is automatically started. The trigger conditions include: control performance indicators exceeding the threshold (overshoot > 5%, static error > 2% of the rated value, response time exceeding the threshold), digital twin model whitening residual exceeding the threshold, significant changes in operating conditions (load change > 30%, winding temperature change > 20℃), and continuous running time reaching the preset self-learning cycle.

[0046] S42. Incremental Learning Dataset Construction: When online learning is triggered, the system automatically collects system input and output data, weighted and enhanced state features, and control effect data within the current operating condition window. At the same time, it combines the extended simulation data of the same operating condition generated by the digital twin model to construct an incremental learning dataset. The dataset is cleaned of outliers, deduplicated, and normalized, and regularization constraints are added to avoid model overfitting.

[0047] S43. Edge-end Incremental Learning and Safe Model Update with Lyapunov Stability Constraints: Construct a global Lyapunov function for the closed-loop control of the servo system, derive the stability boundary constraints of the AI ​​control model output, and use this constraint as the regularization term of the incremental learning loss function; the AI ​​inference kernel uses the PWM dead time and the system idle time slots of the non-sampling period to incrementally fine-tune the AI ​​model based on the incremental dataset, freezes the model backbone feature extraction network, and only fine-tunes the fully connected output layer; the updated model is first simulated and verified in a digital twin model for more than 100 cycles. After verifying that the control performance is better than the current running model and meets the stability constraints, a disturbanceless switching mechanism is adopted to smoothly replace the online running model in the gap of the control cycle. The switching process has no control shock and no output jump, realizing the safe and stable online iteration of the model.

[0048] S5. AI-based full lifecycle fault prediction and adaptive fault-tolerant control, completing fault feature identification and RUL prediction, fault level classification and early warning, and execution of graded fault-tolerant control strategies, specifically including: S51. Fault Feature Identification and Remaining Service Prediction: The weighted and enhanced health feature set and the whitened residual of the digital twin model are input into the Transformer-PHM fault prediction model. The attention mechanism focuses on the fault-related temporal features to identify fault precursor features such as encoder signal abnormalities, IGBT aging, winding insulation degradation, bearing wear, and permanent magnet demagnetization in real time. The remaining service life (RUL) and fault occurrence probability of the core components of the driver and motor are predicted simultaneously.

[0049] S52. Fault Level Classification and Early Warning: Based on the fault occurrence probability, RUL and whitening residual, the system health status is divided into four levels: Level 1 healthy status, Level 2 minor abnormality, Level 3 moderate fault, and Level 4 severe fault. Corresponding early warning signals are output to the host computer for different levels. When there is a Level 2 minor abnormality, online self-learning is automatically triggered to optimize control parameters to offset the performance degradation caused by the minor abnormality.

[0050] S53, Hierarchical Adaptive Fault-Tolerant Control: For Level 3 moderate faults, the system automatically switches to the corresponding fault-tolerant control mode to achieve smooth degraded operation: When the encoder signal is abnormal, the AI ​​sensorless observer is activated to estimate the rotor position and speed in real time based on current, voltage, and vibration signals to maintain normal system operation; when a single current sampling fault occurs, the system automatically switches to the redundant sampling channel and compensates the sampling value through the model observer to ensure normal control of the current loop; when a single IGBT tube ages, a derating operation strategy is implemented to optimize the switching frequency, reduce thermal stress, and extend the service life of components; for Level 4 severe faults, the safe torque shutdown (STO) and safe shutdown procedures are immediately triggered to avoid equipment damage and safety accidents.

[0051] S6. Multi-axis cooperative intelligent synchronization control based on multi-agent reinforcement learning, achieving multi-axis state synchronization interaction and cooperative control optimization, specifically including: S61, Multi-axis status information synchronization and interaction: Through a multi-protocol real-time communication unit, based on EtherCAT / PROFINETIRT real-time industrial Ethernet, the status information of each drive in the multi-axis system is synchronously acquired and interacted, with a global time synchronization accuracy of ≤100ns, ensuring the timing consistency of multi-axis data.

[0052] S62, MARL Multi-Axis Collaborative Synchronization Control: Employing a multi-agent reinforcement learning (MARL) model, each axis's actuator in the multi-axis system is treated as an independent agent. The global optimization objectives are minimizing multi-axis trajectory synchronization error, minimizing system vibration, and minimizing overall energy consumption. Each agent shares state information in real time, combining its own operating state with the global collaborative objectives to optimize the control output of each axis in real time. It automatically compensates for synchronization errors caused by communication delays, mechanical lag, and differences in dynamic characteristics between axes, achieving nanosecond-level high-precision collaborative control for multi-axis systems. This is suitable for the collaborative motion requirements of high-end equipment such as multi-joint robots and five-axis CNC machine tools.

[0053] Therefore, the present invention adopts the above-mentioned AI-based intelligent servo motor driver and control method, which significantly improves the response speed, control accuracy, anti-interference ability, operational reliability and energy efficiency of the servo system, and is suitable for high-end manufacturing scenarios such as industrial robots, high-end CNC machine tools and precision transmission equipment.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A control method for an AI-based intelligent servo motor driver, characterized in that, Includes the following steps: S1. System initialization and offline pre-trained model deployment, completing hardware self-check, motor body parameter identification, digital twin baseline model construction and lightweight AI model edge deployment; S2, Multi-physical quantity full-dimensional state perception and real-time feature extraction, complete multi-source signal synchronous acquisition, signal preprocessing and feature engineering, real-time state mapping of digital twin model based on coupled mutual information entropy, residual whitening and feature weighting enhancement; S3. AI-based multi-loop composite adaptive real-time control, which sequentially completes outer loop intelligent trajectory planning, speed / position loop adaptive composite control, current loop model prediction considering multiple loss coupling, and AI global energy efficiency optimization control; S4. Edge-end online incremental self-learning and dynamic iterative optimization of the model, completing the judgment of learning trigger conditions, construction of incremental datasets, incremental learning with Lyapunov stability hard constraints and safe model updates; S5. AI-based full lifecycle fault prediction and adaptive fault-tolerant control, completing fault feature identification and RUL prediction, fault level classification and early warning, and graded fault-tolerant control strategy execution. S6. Multi-axis cooperative intelligent synchronous control based on multi-agent reinforcement learning to achieve multi-axis state synchronization interaction and cooperative control optimization.

2. The control method for an AI-based intelligent servo motor driver according to claim 1, characterized in that, S1 specifically includes the following steps: S11. Hardware self-test and offline calibration of motor body parameters: Complete the power-on self-test of power unit, sampling unit, encoder unit and dual-core control unit, and complete the offline identification of motor stator resistance, quadrature and direct axis inductance, back electromotive force coefficient, moment of inertia and viscous friction coefficient through recursive least squares method to obtain the basic parameters of motor body. S12. Construction of digital twin baseline model for motor-driver-load: Based on the identified basic parameters, a high-precision digital twin model including electromagnetic model, mechanical dynamics model, thermal loss model and fault evolution model is built as the baseline for system simulation verification and data generation. S13. Offline pre-training and lightweight deployment of AI composite control model system: A training dataset covering all operating conditions is generated through a digital twin model to complete the offline pre-training of the AI ​​composite control model system. The AI ​​composite control model system includes a DRL trajectory planning model, an LSTM-attention mechanism disturbance compensation model, an adaptive PID parameter optimization model, a DQN energy efficiency optimization model, a Transformer-PHM fault prediction model, and a MARL multi-axis collaborative model. The pre-trained model is pruned and lightweighted using INT8 quantization to adapt to edge computing power and deployed to a dedicated AI inference core.

3. The control method for an AI-based intelligent servo motor driver according to claim 1, characterized in that, S2 specifically includes the following steps: S21. Multi-source signal high-frequency synchronous acquisition: The real-time core uses a current loop sampling frequency of 16kHz or higher to synchronously acquire three-phase stator current, bus voltage, encoder position / speed / acceleration signals, IGBT junction temperature, motor stator winding temperature, driver and motor triaxial vibration acceleration signals, noise signals, and bus ripple voltage. All acquired signals are bound to a synchronous timestamp, and the time synchronization error is ≤1μs. S22. Signal preprocessing and multi-dimensional feature extraction: Kalman filtering and amplitude normalization are performed on the original acquired signal to extract time-domain features, frequency-domain features and time-frequency-domain features respectively. The extracted features are divided into a control feature set for real-time control and a health feature set for health management. S23. Real-time mapping of digital twin model, adaptive whitening of residuals and feature weighting enhancement: The actual operating state parameters collected are input into the digital twin baseline model, and the simulation state values ​​of the corresponding working conditions are output. The state residuals between the actual values ​​and the simulation values ​​are calculated. The weight coefficients of each state quantity are calculated based on the mutual information entropy of multi-physical quantity coupling, and the residuals are adaptively whitened to eliminate autocorrelation interference. The control feature set and health feature set are weighted and enhanced based on the weight coefficients. At the same time, the parameters of the digital twin model are updated based on the whitened residuals, and the whitened residuals are used as the core basis for disturbance judgment and fault precursor identification.

4. The control method for an AI-based intelligent servo motor driver according to claim 1, characterized in that, S3 specifically includes the following steps: S31, Outer Loop Intelligent Trajectory Planning and Setpoint Optimization: Receives target position / velocity commands from the host computer, and based on the DRL trajectory planning model, combined with the current motor operating status, load status, and system constraint thresholds, outputs a smooth S-shaped acceleration / deceleration setpoint curve without impact, avoiding overshoot and mechanical shock caused by step commands. S32, Adaptive Composite Control of Velocity Loop and Position Loop: Based on the deviation between the given value and the actual value, the rate of change of the deviation, and the weighted enhanced control feature set, the proportional, integral, and derivative parameters of the PID are adjusted online in real time through an adaptive PID parameter optimization model; at the same time, through an LSTM-attention mechanism disturbance compensation model, the total disturbance value caused by load mutation, nonlinear friction, and parameter drift is predicted in real time, and the disturbance compensation amount is fed forward to the control output to offset the disturbance effect; the real-time core is responsible for the hard real-time calculation of the PID closed loop, and the AI ​​inference core is responsible for the parallel calculation of parameter optimization and disturbance feedforward amount, with a total control delay ≤10μs; S33. Current Loop Model Predictive Control and AI-Enhanced Global Optimal Torque Control: Model predictive control (MPC) is used as the core controller of the current loop. Based on the motor electromagnetic model, current setpoint, actual current value, and bus voltage, it predicts the current trajectory for the next three control cycles and outputs the optimal PWM duty cycle signal. At the same time, a global loss model considering magnetic saturation cross-coupling and multi-loss coupling is constructed. Through the DQN energy efficiency optimization model, combined with the current speed, load torque, winding temperature, and magnetic saturation state, the optimal AC and DC axis current setpoints and IGBT switching frequencies are solved in real time. Dynamic optimal control of the maximum torque-current ratio and maximum torque-voltage ratio is achieved across the entire operating range, minimizing copper losses, iron losses, inverter switching losses, and dead-zone losses.

5. The control method for an AI-based intelligent servo motor driver according to claim 1, characterized in that, S4 specifically includes the following steps: S41. Online learning trigger condition judgment: The system monitors the running status in real time and starts online incremental learning when any trigger condition is met. The trigger conditions include: control performance index exceeding the threshold, digital twin model whitening residual exceeding the threshold, significant change in operating conditions, and running time reaching the preset self-learning cycle. S42. Construction of Incremental Learning Dataset: Collect system input and output data, weighted and enhanced state features, and control effect data under the current working conditions. Combine these with the extended simulation data generated by the digital twin model under the same working conditions to construct an incremental learning dataset. Complete data cleaning, deduplication, and normalization. S43. Edge-end Incremental Learning and Model Safety Update with Lyapunov Stability Constraints: Construct a global Lyapunov function for the closed-loop control of the servo system, derive the stability boundary constraints of the AI ​​control model output, and use this constraint as the regularization term of the incremental learning loss function; the AI ​​inference kernel uses the PWM dead time and the system idle time slots without sampling periods to incrementally fine-tune the AI ​​model based on the incremental dataset, freezes the model's backbone feature extraction network, and only fine-tunes the fully connected output layer; the updated model is first simulated and verified in a digital twin model. After verifying that the control performance is better than the current running model and meets the stability constraints, a disturbanceless switching mechanism is used to smoothly replace the online running model. The switching process has no control shock or system instability risk.

6. The control method for an AI-based intelligent servo motor driver according to claim 1, characterized in that, S5 specifically includes the following steps: S51. Fault Feature Identification and Remaining Service Prediction: The weighted and enhanced health feature set and the whitened residual of the digital twin model are input into the Transformer-PHM fault prediction model to identify fault precursor features such as encoder abnormality, IGBT aging, winding insulation degradation, bearing wear, and permanent magnet demagnetization in real time, and simultaneously predict the remaining service life (RUL) and fault occurrence probability of core components. S52. Fault Level Classification and Early Warning: Based on fault probability, RUL and whitening residual, the system health status is divided into four levels: healthy status, minor anomaly, moderate fault and severe fault. Corresponding early warning signals are output for different levels. When there is a minor anomaly, online self-learning optimization of control parameters is triggered. S53, Hierarchical Adaptive Fault-Tolerant Control: For moderate faults, automatically switch to the corresponding fault-tolerant control mode: When the encoder signal is abnormal, start the AI ​​sensorless observer to estimate the rotor position and speed in real time based on current, voltage, and vibration signals to maintain normal system operation; when a single current sampling fault occurs, switch to the redundant sampling channel and compensate the sampling value through the model observer; when a single IGBT tube is aging, execute the derating operation strategy to optimize the switching frequency and reduce thermal stress; for severe faults, trigger the safety shutdown procedure to avoid equipment damage.

7. The control method for an AI-based intelligent servo motor driver according to claim 1, characterized in that, S6 specifically includes the following steps: S61, Multi-axis status information synchronization and interaction: Real-time industrial Ethernet is used to complete the synchronous acquisition of status information of each drive in the multi-axis system, with a global time synchronization accuracy of ≤100ns. S62, MARL Multi-Axis Cooperative Synchronization Control: Employing a multi-agent reinforcement learning MARL model, each axis's driver is treated as an independent agent. The global optimization objectives are minimizing multi-axis trajectory synchronization error, minimizing system vibration, and minimizing overall energy consumption. The control output of each axis is optimized in real time to compensate for synchronization errors caused by communication delays, mechanical lag, and differences in inter-axis characteristics, achieving high-precision multi-axis nanosecond-level cooperative control.

8. An AI-based intelligent servo motor driver for executing the control method according to any one of claims 1-7, characterized in that: It includes a heterogeneous dual-core control unit, a multi-physical quantity signal acquisition unit, a power drive unit, a safety redundancy unit, a multi-protocol real-time communication unit, and a storage unit; The heterogeneous dual-core control unit includes a hard real-time control core and a dedicated AI inference core. The hard real-time control core is used to perform hard real-time closed-loop control of the current loop, speed loop, and position loop. The dedicated AI inference core is used to perform online inference, incremental learning, and data processing of the AI ​​model. The two cores communicate in parallel through a high-speed internal bus. The multi-physical quantity signal acquisition unit includes an electrical signal acquisition module, a position sensing module, a temperature acquisition module, and a vibration and noise acquisition module, which are used to complete the synchronous acquisition of the full-dimensional status signals of the driver and the motor. The power drive unit includes a rectifier module, an IPM intelligent power module, and a buffer and protection circuit, which are used to perform power drive and hardware protection of the motor. The safety redundancy unit includes dual-channel redundant sampling channels, dual-channel encoder interfaces, and emergency stop and safety torque shutdown (STO) circuits, which are used to realize system redundancy protection and fault-tolerant control hardware support. The multi-protocol real-time communication unit supports EtherCAT, PROFINET IRT, and EtherNet / IP real-time industrial Ethernet protocols, enabling high-speed synchronous communication with host computers and multi-axis systems.

9. The AI-based intelligent servo motor driver according to claim 8, characterized in that: The hard real-time control core adopts an ARM Cortex-R5F dual-core architecture with a main frequency of ≥800MHz and a minimum control cycle of 62.5μs; the dedicated AI inference core adopts a RISC-V architecture with NPU acceleration, with AI computing power of ≥4TOPS, and supports model pruning, quantization and incremental learning.

10. The AI-based intelligent servo motor driver according to claim 8, characterized in that: The multi-physical quantity signal acquisition unit adopts a 16-bit or higher synchronous sampling ADC with a sampling frequency of ≥16kHz. All acquisition channels are hardware synchronized with a time synchronization error of ≤1μs.

Images