Self-adaptive control method and system for high-frequency response current loop of servo driving system

Abstract

The invention relates to the field of servo drive control, and provides a self-adaptive control method and system for a high-frequency response current loop of a servo drive system, and the method comprises the steps: collecting an instantaneous state vector set, constructing an input graph snapshot, and fusing a loss item to output a thermal coupling graph neural network model; a current loop total delay and coupling torque model is identified through a parameter and dynamics identification decoder, a next cycle parameter is predicted by combining a future motion instruction, and a composite disturbance voltage set is mapped and output; according to the reference current, the rotating speed, the motor parameter set and the composite disturbance voltage set, outputting feed-forward compensation voltage; calculating a current tracking error, comparing a stability margin and a safety threshold, calculating a feedback control voltage component, superposing a feedforward compensation voltage to synthesize a total instruction voltage, and applying dynamic amplitude limiting to output a safety voltage instruction set. According to the method, thermal-mechanical coupling identification prediction of physical prior and adaptive feedback amplitude limiting of stability margin are fused, and prediction feedforward and stability constraint feedback combined control is constructed.

Description

Technical Field

[0001] This invention relates to the field of servo drive control technology, and in particular to a high-frequency response current loop adaptive control method and system for servo drive systems. Background Technology

[0002] With the rapid application of servo drive systems in high-end CNC machine tools, precision robots, and automated production lines, high-frequency response control of the current loop has become a key technology for ensuring high-precision contour tracking and high dynamic response of the system. How to ensure the high performance and high stability of the control loop in real time under increasingly complex dynamic operating conditions and changing physical characteristics, and effectively compensate for the problems of poor adaptability, poor robustness, and model recognition lag of traditional control strategies, has become an important issue that urgently needs to be addressed in the field of high-performance servo control.

[0003] Chinese patent application CN116248006A discloses a servo driver and its overshoot-free current loop circuit. The overshoot-free current loop circuit includes three current loops disposed within the servo driver and corresponding to the U-phase, V-phase, and W-phase of the servo motor, respectively. Each current loop includes an actual current feedback module, a command current conversion module, a current comparison module, and a MOSFET switching module. The actual current feedback module is connected to the current output terminal of the corresponding phase and the first input terminal of the current comparison module. The command current conversion module is connected to the upper controller of the servo motor and the second input terminal of the current comparison module. The current comparison module is connected to the first input terminal of the MOSFET switching module and the voltage input terminal of the corresponding phase of the servo motor.

[0004] However, current technology still faces many challenges. In multi-axis high-precision contour machining, especially in quadrant switching scenarios, servo drive systems often simultaneously experience external coupling disturbances from adjacent axes and internal parameter drift caused by heat generation from their own power devices, particularly phase lag. If a high-gain fixed PI controller is forced to suppress external coupling disturbances, the system will be unable to detect the phase lag and stability margin reduction caused by temperature accumulation in real time. Once the PI controller encounters a critical phase margin due to drift in a high-gain state, the high-frequency gain will cross the 0dB line. This will cause the entire drive system to suddenly experience high-frequency self-excited oscillations, potentially causing permanent hardware damage to the driver power module. Furthermore, the severe current fluctuations will be transmitted to the motor output, ultimately leading to surface marks or even scrapping of the machined workpiece, threatening the yield rate and equipment safety in high-end manufacturing. Summary of the Invention

[0005] To achieve the above objectives, this invention provides a high-frequency current loop adaptive control method for servo drive systems, the specific technical solution of which is as follows:

[0006] A set of instantaneous state vectors is collected and combined with kinematic configuration parameters to construct a dynamic weighted adjacency matrix. A set of physical prior constraints is constructed using the thermal parameters of the driver hardware. The set of instantaneous state vectors, the dynamic weighted adjacency matrix, and the set of physical prior constraints are used to construct an input graph snapshot. A physical prior constraint loss term is defined, which is fused with the data-driven loss term and used to train a graph neural network. A thermally coupled graph neural network model is then output.

[0007] The input graph snapshot is input into the thermally coupled graph neural network model, and the output hidden state set is processed by the parameter identification decoder and the dynamic identification decoder to identify the system-level current loop total delay and coupling torque model from the hidden state set. Combined with future motion commands, the system-level current loop total delay and coupling torque of the next cycle are predicted. The next cycle composite disturbance voltage set is output by fusing through the mapping function.

[0008] Reference for the next cycle based on future motion instructions shaft current and reference The set of shaft current, motor speed and motor parameters is used to calculate the standard feedforward voltage using the standard field-oriented control model. The standard feedforward compensation voltage set is generated and output by combining the composite disturbance voltage set component by component.

[0009] calculate Axis current tracking error and The shaft current tracking error is calculated by dynamically outputting the proportional gain and integral gain through the gain scheduler based on the comparison results of the stability margin and the preset safety threshold. PI control operation is performed to calculate the set of feedback control voltage components. The set of feedback control voltage components and the set of feedforward compensation voltage components are superimposed to synthesize the total command voltage. Dynamic safety limiting is applied to the total command voltage using the comparison results, and a safe voltage command set is output.

[0010] Furthermore, the output method of the thermally coupled graph neural network model includes:

[0011] Synchronously collect all The electrical, mechanical, and thermal state data of each servo axis are used to construct an instantaneous state vector set;

[0012] The electrical state includes Axis current components and The shaft current component, the mechanical state includes motor speed and motor position, and the thermal state includes the physical temperature of the intelligent power unit and the temperature of the current sensor;

[0013] Based on the kinematic configuration parameters, the physical connection between servo axes is analyzed, a static topological adjacency matrix is ​​constructed, and the dynamic weight is calculated by using the real-time position vector in the instantaneous state vector set to drive the state-dependent edge weight function.

[0014] Instantaneous power loss is calculated based on electrical and thermal state data of instantaneous state vector set. A thermal dynamic unit is constructed in combination with driver hardware thermal parameters to estimate hardware temperature. Based on the hardware temperature, the total delay of current loop is calculated through delay drift unit to construct physical prior constraint set.

[0015] The instantaneous state vector set is used as node features, the dynamic weighted adjacency matrix is ​​used as edges, and the input graph snapshot is constructed by combining the physical prior constraint set. The physical prior constraint loss term is defined based on the physical prior constraint set and fused with the data-driven loss term to form a fusion loss. The fusion loss is used as the optimization objective and the input graph snapshot is used as the input to train the graph neural network and output the thermally coupled graph neural network model.

[0016] Furthermore, the element value rule in the static topological adjacency matrix is ​​that the element of the corresponding static topological adjacency matrix is ​​assigned a value of 1 if and only if the system kinematic configuration parameters determine the motion of one axis and have a dynamic effect on the other axis; otherwise, if the two axes are decoupled in dynamics and have no direct kinematic dependence, the element is assigned a value of 0.

[0017] Furthermore, the method for outputting the next cycle composite disturbance voltage set includes:

[0018] The input graph snapshot is fed into the thermally coupled graph neural network model, and execution is performed within the thermally coupled graph neural network model. The layered graph neural network computes by using the dynamic weighted adjacency matrix in the input graph snapshot to weight and aggregate neighborhood information in order to output a set of hidden states;

[0019] Using the parameter identification decoder in the thermally coupled graph neural network model, the total delay of the second current loop of each axis is identified from the set of hidden states, and the total delay of the system-level current loop is determined. Based on the total delay of the system-level current loop and the target closed-loop bandwidth of the servo current loop, the system stability margin is calculated.

[0020] By using the dynamic identification decoder in the thermo-coupled graph neural network model, the coupling torque model is decoded from the set of hidden states, and the estimated value of the coupling torque is calculated in real time.

[0021] By utilizing the future motion commands of the servo PLC motion controller and the total delay of the system-level current loop, the total delay of the system-level current loop in the next cycle is predicted. Based on the coupling torque model, the coupling torque in the next cycle is predicted. The total delay of the system-level current loop in the next cycle and the coupling torque in the next cycle are fused through a mapping function to output the composite disturbance voltage set for the next cycle.

[0022] Furthermore, the mapping function comprehensively considers torque disturbances to The effects of axis voltage conversion and delay on d / q axis voltage;

[0023] The torque disturbance to The conversion of shaft voltage cancels out the predicted coupling torque. External disturbances to the shaft current loop are applied to maintain the desired... shaft current, apply an additional The shaft voltage, the mapping function uses the set of motor parameters to calculate the equivalent of the predicted coupling torque to offset the prediction. Axis voltage disturbance component, and include the predicted Shaft equivalent disturbance voltage;

[0024] The effect of the delay on the d / q axis voltage is determined by the mapping function, which calculates the equivalent d / q axis voltage deviation dynamically generated by the predicted total system-level current loop delay acting on the controller, based on the predicted total system-level current loop delay, the controller's dynamic characteristics, and the current internal state of the controller. This deviation is then added to the predicted total system-level current loop delay. Shaft equivalent disturbance voltage and prediction In the shaft equivalent disturbance voltage.

[0025] Furthermore, the method for generating and outputting the feedforward compensation voltage set includes:

[0026] Utilizing the next cycle reference of future motion commands shaft current and reference The shaft current, the collected motor speed, and the set of motor parameters are used to calculate, using a standard field-oriented control model, including standard... Shaft feedforward voltage components and standard Standard feedforward voltage for the shaft feedforward voltage component;

[0027] The standard feedforward voltage and the composite disturbance voltage set are added component by component and combined to obtain a result containing... Shaft feedforward compensation voltage and The set of feedforward compensation voltages for shaft feedforward compensation voltages.

[0028] Furthermore, the standard Shaft feedforward voltage components include compensation Two compensating properties: cross-coupled electromotive force and compensated back electromotive force;

[0029] The compensation The cross-coupled electromotive force is when the axis is Axial flow through reference When the shaft current is low, due to the motor speed and nominal... Shaft inductance, in A cross-coupled electromotive force is induced in the shaft;

[0030] The aforementioned compensating back electromotive force is a back electromotive force generated by the permanent magnet flux linkage when the motor rotates at a certain speed, which is proportional to the speed and mainly acts on... axis.

[0031] Furthermore, the method for outputting the safety voltage command set includes:

[0032] Calculate the reference for the next period shaft current and reference The axis currents are respectively compared with the data collected at the current time. Axis current components and The difference between the shaft current components generates Axis current tracking error and Shaft current tracking error;

[0033] The stability margin is compared with the preset safety threshold. If the stability margin is higher than the preset safety threshold, the proportional gain and integral gain of the output are set to the first proportional gain and the first integral gain respectively by the gain scheduler. If the stability margin is lower than the preset safety threshold, the proportional gain and integral gain of the output are set to the second proportional gain and the second integral gain respectively.

[0034] in accordance with Axis current tracking error and The shaft current tracking error is investigated, and discrete-time PI control is performed using proportional gain and integral gain to calculate the set of feedback control voltage components for the next cycle.

[0035] The feedforward compensation voltage set and the feedback control voltage component set are superimposed to synthesize the total command voltage. The stability margin is compared with the preset margin threshold. If the stability margin is higher than the preset margin threshold, a first limiting strategy is applied through the dynamic safety limiting unit. If the stability margin is lower than the preset margin threshold, a second limiting strategy is applied through the dynamic safety limiting unit, and a safe voltage command set is output.

[0036] A high-frequency response current loop adaptive control system for a servo drive system, which is used to implement the above-mentioned high-frequency response current loop adaptive control method for the servo drive system, includes a model building module, a discrimination and prediction module, a feedforward compensation module, and an adaptive feedback module.

[0037] The model building module is used to collect a set of instantaneous state vectors, combine them with kinematic configuration parameters, construct a dynamic weighted adjacency matrix, construct a set of physical prior constraints using the driver hardware thermal parameters, construct an input graph snapshot using the set of instantaneous state vectors, the dynamic weighted adjacency matrix, and the set of physical prior constraints, define a physical prior constraint loss term, fuse it with a data-driven loss term, train a graph neural network, and output a thermally coupled graph neural network model.

[0038] The discrimination and prediction module is used to input the input graph snapshot into the thermally coupled graph neural network model, output a set of hidden states, and then, through the parameter identification decoder and the dynamic identification decoder, identify the system-level current loop total delay and coupling torque model from the set of hidden states, respectively. Combined with future motion commands, it predicts the system-level current loop total delay and coupling torque of the next cycle, and outputs the composite disturbance voltage set of the next cycle through mapping function fusion.

[0039] The feedforward compensation module is used to reference the next cycle of future motion commands. shaft current and reference The set of shaft current, motor speed, and motor parameters is used to calculate the standard feedforward voltage using a standard field-oriented control model. This is then combined with the composite disturbance voltage set and superimposed component by component to generate and output the feedforward compensation voltage set.

[0040] The adaptive feedback module is used to calculate Axis current tracking error and The shaft current tracking error is calculated by dynamically outputting the proportional gain and integral gain through the gain scheduler based on the comparison results of the stability margin and the preset safety threshold. PI control operation is performed to calculate the set of feedback control voltage components. The set of feedback control voltage components and the set of feedforward compensation voltage components are superimposed to synthesize the total command voltage. Dynamic safety limiting is applied to the total command voltage using the comparison results, and a safe voltage command set is output.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] This invention uses a dynamically weighted adjacency matrix representing external coupling disturbances as the model input of a graph neural network, and simultaneously uses a set of physical prior constraints representing internal parameter drift as the physical prior constraint loss term of the graph neural network for fusion training. This achieves simultaneous modeling and decoupling of the two physical effects of external coupling and internal drift, thereby avoiding the dilemma of traditional fixed PI controllers being unable to balance high bandwidth and high stability under a single gain.

[0043] This invention utilizes two dedicated decoders in a thermodynamically coupled graph neural network model—a parameter identification decoder and a dynamic identification decoder—to decouple and identify the internal parameter drift that leads to the dilemma from a unified set of hidden states. This is quantified as the total delay of the system-level current loop and the external coupled disturbance, which is quantified as a coupled torque model. The results are then combined with future motion commands to fuse the predicted effects of both into a single composite disturbance voltage. This provides subsequent adaptive control with a forward-looking basis for simultaneously compensating for the two conflicting disturbances, avoiding the trade-off between stability and bandwidth that traditional fixed PI controllers cannot achieve.

[0044] This invention achieves forward-looking cancellation of known dynamics and predicted disturbances by actively superimposing the standard feedforward voltage and non-ideal composite disturbance voltage of the compensation ideal model dynamics. This improves the problems of insufficient bandwidth, inability to effectively suppress coupling disturbances and reduced contour tracking accuracy caused by traditional PI controllers that rely solely on hysteresis feedback.

[0045] This invention uses the real-time stability margin as the adaptive scheduling basis for the feedback gain of the PI controller, namely the proportional gain and integral gain, and as the dynamic safety limiting basis for the final total command voltage. This enables the automatic reduction of gain and strict limiting when stability decreases, and the automatic increase of gain and relaxation of limiting when stability is sufficient. This avoids the dilemma that traditional fixed PI controllers face when internal parameter drift and external coupling disturbances are superimposed, as they cannot balance high stability and high bandwidth disturbance suppression. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart illustrating the principle of the high-frequency response current loop adaptive control method for the servo drive system of the present invention.

[0048] Figure 2 This is a functional block diagram of the high-frequency response current loop adaptive control system of the servo drive system of the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Example 1:

[0051] Please see Figure 1 As shown, this embodiment provides a high-frequency response current loop adaptive control method for a servo drive system, including:

[0052] S1000, collects the instantaneous state vector set. Combined with kinematic configuration parameters Construct a dynamic weighted adjacency matrix Utilizing the driver hardware thermal parameters Constructing a set of physical prior constraints The instantaneous state vector set The dynamic weighted adjacency matrix and the set of physical prior constraints Jointly construct input graph snapshot Define the physical prior constraint loss term , and data-driven loss terms Fuse and train graph neural networks to output a thermally coupled graph neural network model. .

[0053] Specifically, this step aims to construct a graphical neural network model capable of simultaneously describing the kinematic coupling and thermal-mechanical coupling of hardware thermal drift in a multi-axis computer numerical control (CNC) machine tool system. It also collects multimodal data from the multi-axis CNC machine tool system in real time as its input. This thermo-coupled graph neural network model This forms the basis for subsequent implementation of adaptive control of the high-frequency response current loop.

[0054] Further, step S1000 includes:

[0055] Step S1100: Synchronously collect electrical, mechanical, and thermal state data of all N servo axes to construct an instantaneous state vector set. .

[0056] Specifically, this step aims to synchronously collect electrical, mechanical, and thermal multimodal data of the multi-axis CNC machine tool system in each control cycle, using k as the current time index, and construct an instantaneous state vector for each axis i. The instantaneous state vector Simultaneously, it forms the dynamic weighted adjacency matrix for step S1200. Provide mechanical state input, which is the set of physical prior constraints for step S1300. Providing real-time inputs of electrical and thermal states forms the physical data foundation for subsequent perturbation attribution in Graph Neural Networks (GNNs).

[0057] In the specific implementation process, this step performs the following actions: In each control cycle, i.e., time k, the following state variables of all axes are synchronously acquired through the high-speed analog-to-digital converter (ADC) and encoder interface:

[0058] First, electrical status, acquiring the d-axis current component. and q-axis current component .

[0059] Second, mechanical status, collecting motor speed. and motor position .

[0060] Third, thermal status, collecting the physical temperature of the Intelligent Power Module (IPM). and current sensor temperature .

[0061] The collected state variables are combined into an instantaneous state vector of axis i at time k. The instantaneous state vector This constitutes the motion state of servo axis i at time k, and serves as a node in the time k graph. The feature inputs are then used. Finally, a set of instantaneous state vectors for all N servo axes is constructed. .

[0062] Where i represents the index of the servo axis, which is a positive integer ranging from 1 to N; N represents the total number of servo axes; This represents the d-axis current component of axis i at time k, i.e., the excitation current in Field Oriented Control (FOC), measured in amperes (A), and is used in the thermal dynamics unit of step S1300. Calculate the instantaneous power loss of axis i ; This represents the q-axis current component of axis i at time k, i.e., the torque current in the field-oriented control (FOC), measured in amperes (A), and is used in the thermal dynamics unit of step S1300. Calculate the instantaneous power loss of axis i ; This represents the motor speed of axis i at time k, in radians per second (rad / s), used in step S1200 to assemble the N-dimensional real-time velocity vector. and the Coriolis force matrix ; This represents the motor position of axis i at time k, in radians (rad), and is used in step S1200 to assemble the N-dimensional real-time position vector. To calculate dynamic weights and inertia matrix ; The physical temperature of the intelligent power unit IPM at time k, representing axis i, is a temperature reading collected by a physical sensor embedded in the intelligent power unit IPM, expressed in Kelvin (K) or Celsius (°C), and is used for the thermal dynamics unit in step S1300. The measured values ​​are used for calibration, and also for delay drift units. Calculate the total current loop delay of axis i. ; The temperature of the current sensor at time k, representing axis i, is the temperature reading of the current sensor acquired by the physical sensor, in Kelvin (K) or Celsius (°C), and is used in the delay drift unit of step S1300. Calculate the total current loop delay of axis i. ; This indicates the transpose operation.

[0063] Step S1200, based on kinematic configuration parameters Analyze the physical connections between servo axes and construct a static topological adjacency matrix. And using the instantaneous state vector set Real-time position vector Driving state-dependent edge weight function To calculate dynamic weights Construct a dynamic weighted adjacency matrix .

[0064] Specifically, this step aims to utilize the kinematic configuration parameters of a multi-axis CNC machine tool system. Construct a dynamically weighted adjacency matrix that can dynamically quantify the dynamic coupling relationships between multiple axes. This construction process includes not only defining the static topological adjacency matrix of the kinematic coupling graph, but also... More importantly, it lies in establishing a state-dependent edge weight function. The edge weight function Based on step S1100, construct the instantaneous state vector set for all N servo axes. In the process, extract the motor position of each servo axis. This is assembled into a complete N-dimensional real-time position vector. It is used to quantify the intensity of external coupling disturbances that vary nonlinearly in real time.

[0065] In the specific implementation process, this step includes the following stages:

[0066] The first stage is the node definition stage. N servo axes are defined as N graph nodes. Each node It corresponds to one servo axis and is used to carry the instantaneous state vector acquired in step S1100. .in, This indicates that it contains N servo axes. The set of graph nodes; N represents the total number of servo axes; Represents the set of graph nodes The i-th servo axis in the graph; i represents the index of the graph node, which is a positive integer ranging from 1 to N, used to uniquely identify and traverse the set of graph nodes. Each node in .

[0067] The second stage is the static topology definition stage. This involves defining the kinematic configuration parameters based on the multi-axis CNC machine tool system. This analysis examines the physical connections and kinematic dependencies between axes, such as the C-axis being mounted on the A-axis. Through this analysis, a binary, static topological adjacency matrix is ​​constructed. .

[0068] The topological adjacency matrix medium elements The rule for determining the value is: if and only if the system kinematic configuration parameter The topological adjacency matrix determines the motion of the i-th axis, such as velocity or acceleration, and can have a dynamic impact on the j-th axis, such as inertia, Coriolis force, or gravitational load disturbances. element in row i and column j The value is assigned to 1; conversely, if the two axes are dynamically decoupled, such as the two X-axis of a gantry structure having no direct kinematic dependence, then the value is assigned to 0. Therefore, the topological adjacency matrix... The static path for message passing in a graph neural network (GNN) is defined a priori, indicating which axes may have physical coupling.

[0069] in, The kinematic configuration parameters of a multi-axis CNC machine tool system are a set of prior information describing the physical stacking structure, linkage parameters and kinematic dependencies of each servo axis of the system. The topological adjacency matrix is ​​a... A binary square matrix whose elements take values ​​of 0 or 1, used to statically describe the possibility of physical coupling between N servo axes; i represents the topological adjacency matrix. The row index is a positive integer ranging from 1 to N, used to identify the source node in the coupling relationship, i.e., the i-th servo axis that exerts the dynamic influence; j represents the topological adjacency matrix. The column index is a positive integer ranging from 1 to N, used to identify the target node in the coupling relationship, i.e., the j-th servo axis that is subject to dynamic influence.

[0070] The third stage is the construction of the dynamic edge weight function. To overcome the problem that external coupling perturbations cannot be quantified in real time in existing technologies, this step constructs a state-dependent edge weight function. This function Used at time k, based on the N-dimensional real-time position vector Calculate the dynamic weighted adjacency matrix used for message passing. Its elements The specific calculation formula is as follows:

[0071] ;

[0072] in, The dynamically weighted adjacency matrix at time k is a A real matrix whose elements Real-time quantization measures the instantaneous intensity of the external coupling perturbation of axis i to axis j at time k; k represents the current discrete time. The dynamic weights, representing state dependencies, are N-dimensional real-time position vectors. A changing scalar value.

[0073] The state-dependent dynamic weights By edge weight function The calculation yields the value used for real-time quantization of the current N-dimensional real-time position vector. Below, the instantaneous coupling strength of axis i to axis j. This function is preferably based on offline identification or online calculation of the system's multi-body dynamics (MBD) model to quantify the instantaneous coupling strength with the N-dimensional real-time position vector. Related inertia matrix and Coriolis force matrix Coupling, the specific calculation formula is as follows:

[0074] ;

[0075] in, The edge weight function represents the state dependency, used to map the real-time dynamic characteristics of a multi-axis CNC machine tool system to specific edges in a graph neural network (GNN). instantaneous weight ; The inertia matrix of the system is a A symmetric positive definite matrix whose elements vary with the real-time position vector of the system. It exhibits nonlinear variation and is the main source of describing system inertial coupling; The Coriolis force matrix of the system is a A matrix whose elements vary with the system's real-time position vector and real-time velocity vector Common changes are the source of velocity-dependent coupling in a system; The N-axis real-time velocity vector at time k is a A column vector whose elements The motor speed of axis i at time k is collected in step S1100. .

[0076] Step S1300, based on the instantaneous state vector set Calculate instantaneous power loss from electrical and thermal state data. Combined with the driver hardware thermal parameters Constructing thermal dynamic units To estimate the hardware temperature, based on the hardware temperature via a delay drift unit. Calculate the total delay of the current loop. Construct a set of physical prior constraints .

[0077] Specifically, this step aims to determine the thermal parameters of the driver hardware. Construct a set of physical prior constraints This set of physical prior constraints Establish a path for each servo axis based on electrical instantaneous power loss. To hardware temperature Then there's the total delay of the current loop, which determines the high-frequency stability of the system. The physical causal chain of electrical-thermal-delay provides a key constraint for graph neural networks (GNNs) to identify phase lag caused by internal parameter drift.

[0078] In the specific implementation process, the physical prior constraint set The construction consists of two parts, one being the thermal dynamic unit. Secondly, the delayed drift unit .

[0079] First, thermal dynamic unit The establishment of this unit. Based on the laws of physical thermodynamics, this thermal dynamics unit is used to describe the dynamic temperature changes of key heating elements on a single axis i, such as intelligent power units (IPMs) and current sensors. In the subsequent step S1400, the constraint graph neural network (GNN) is used to identify the internal temperature of the system, ensuring that the identification result conforms to physical convergence.

[0080] Taking the intelligent power unit IPM of axis i as an example, in order to overcome the measurement lag of traditional physical sensors and the fact that they can only measure the case temperature rather than the critical junction temperature, this step preferably uses a lumped parameter RC thermal network model to construct a virtual temperature sensor, namely a thermal dynamic unit. Used to estimate the core junction temperature of axis i in real time. The discrete-time thermal model is constructed as follows:

[0081] ;

[0082] in, Indicates the time of axis i. The junction temperature of the intelligent power unit (IPM) is a thermal dynamic unit. The predicted output value of the junction temperature at time k for the next time step, in Kelvin (K) or Celsius (°C), is used to iteratively update the thermal dynamics unit. The state; This represents the junction temperature of the intelligent power unit (IPM) at time k, which is a thermal dynamic unit. The current state input, whose value is determined by the previous time step. of Iteratively derived; The control period is represented by the time interval between time k and time t. The time interval between them is in seconds (s). The equivalent heat capacity from the intelligent power unit (IPM) of axis i to the heat sink is a priori physical parameter characterizing the thermal properties of the hardware, and its value is derived from the thermal parameters of the driver hardware. This is used to describe the ability of an intelligent power unit (IPM) to store thermal energy, and the unit is joules per Kelvin (J / K). This represents the instantaneous power loss of axis i at time k. It is a scalar value that changes with the electrical state at time k and is used to quantify the total heat power generated by the intelligent power unit (IPM) at time k. The unit is watts (W). Indicates ambient temperature, used to characterize the reference temperature around the driver or the base of the heat sink; This represents the equivalent thermal resistance of axis i at the IPM junction of the intelligent power unit to the environment. It is a priori physical parameter characterizing the thermal properties of the hardware, and its value is derived from the thermal parameters of the driver hardware. The unit is Kelvin per watt (K / W); The heat dissipation power of axis i at time k is calculated based on Ohm's law of thermal conduction to determine the junction temperature of the intelligent power unit (IPM) at time k. With ambient temperature The heat flow transferred to the environment driven by the temperature difference between them; This represents the net increase in power of axis i at time k.

[0083] The instantaneous power loss This is the cause of the junction temperature of the IPM in the i-axis intelligent power unit. The rising direct electrical input, i.e., the heat source. This instantaneous power loss. The calculation of this is crucial for achieving electro-thermal conversion, and the specific calculation formula is as follows:

[0084] ;

[0085] in, The on-resistance of axis i is represented by the junction temperature of the intelligent power unit (IPM) along axis i at time k. The changing function output value, in ohms. This is used to characterize the resistance of the power switch in the intelligent power unit (IPM) when it is in the on state, and to calculate the instantaneous power loss accordingly. The conduction loss component in the middle; The motor d-axis current component at time k represents axis i, which is the excitation current component in Field Oriented Control (FOC). Its value is acquired in step S1100 and the unit is ampere (A). The motor q-axis current component at time k represents axis i, which is the torque current component in field-oriented control (FOC). Its value is collected in step S1100 and the unit is ampere (A). The switching loss function of the intelligent power unit (IPM) for axis i is a complex nonlinear function used to calculate the power loss generated by the power switch in the IPM during pulse width modulation (PWM) switching. Its value depends on the instantaneous current amplitude of axis i. DC bus voltage of servo driver and the switching frequency of pulse width modulation (PWM) Multiple input variables; This represents the instantaneous current amplitude of axis i at time k, and its value is usually expressed as... Calculated; This indicates the DC bus voltage of the servo driver, measured in volts (V), and is used to determine the voltage stress that the switching transistors withstand. This indicates the switching frequency of pulse width modulation (PWM), measured in Hertz (Hz), and is used to determine the number of switching actions per unit time. The conduction loss is represented by a calculation based on Joule's law, i.e. .

[0086] Second, delayed drift unit The establishment of the delay drift unit. It is a mapping from thermal to high-frequency characteristics, designed to integrate thermal dynamics units. The output, such as the junction temperature of the intelligent power unit IPM at time k-axis i acquired in step S1100, is shown. The current sensor temperature at time k-axis i This is related to the high-frequency stability of the servo drive system. The root cause of this high-frequency stability issue lies in the total delay of the current loop of axis i. This is the direct physical reason for the system's phase margin drift with temperature, and also the root cause of the high-frequency self-oscillation that easily occurs in fixed PI controllers under high gain. The specific calculation process is as follows:

[0087]

[0088] ;

[0089] in, This represents the total current loop delay of axis i at time k, and is a scalar value that varies with temperature, in seconds (s). This represents the delay drift unit, used to measure the hardware temperature at time k, i.e. and As input, and mapped to the total current loop delay at time k-axis i. ; The nominal fundamental delay of the system is a priori calibrated positive real constant in seconds (s). Its value represents the fundamental loop delay of the servo drive system calibrated in a cold state, which does not change with temperature, and serves as the total current loop delay of axis i. The baseline value; This represents the delay drift function of the intelligent power unit (IPM), used to determine the junction temperature of the IPM on axis i at time k. As input, it maps out the amount of delay change caused by the temperature drift of the intelligent power unit (IPM); This represents the current sensor delay drift function, used to determine the current sensor temperature of axis i at time k. As input, it maps the amount of delay change caused by the temperature drift of the current sensor.

[0090] Step S1400: Set the instantaneous state vectors As a node feature, the dynamically weighted adjacency matrix As an edge, combined with the set of physical prior constraints Build input graph snapshot Based on the set of physical prior constraints Define the physical prior constraint loss term , and data-driven loss terms The fusion process results in fusion loss. With the fusion loss To optimize the target, the input graph snapshot As input, train a graph neural network; as output, a thermally coupled graph neural network model. .

[0091] Specifically, this step aims to integrate the instantaneous state vector set acquired in step S1100. The dynamic weighted adjacency matrix constructed in step S1200 and the set of physical prior constraints constructed in step S1300 By integrating these components, a unified and trainable thermally coupled graph neural network model can be constructed. This enables dynamic characterization of external force coupling and internal thermal drift.

[0092] In the specific implementation process, the construction logic of this step is as follows:

[0093] First, construct a snapshot of the input graph at time k. The instantaneous state vector set collected in step S1100 Assign to the graph nodes defined in step S1200 As a node feature at time k Simultaneously, the dynamically weighted adjacency matrix constructed in step S1200 is... Snapshot of the input graph at time k edge , used for message passing in graph neural networks (GNNs) to characterize the coupling of forces.

[0094] Secondly, define the physical prior constraint loss term for the graph neural network (GNN). The physical prior constraint set constructed in step S1300 The physical prior constraint loss term is a key constraint term in the loss function during the training of a graph neural network (GNN). It is used to characterize thermal drift.

[0095] Subsequently, design fusion loss This fusion loss The data-driven loss term used to ensure prediction accuracy will be used. and physical prior constraint loss term used to ensure physical consistency The combination is a thermally coupled graph neural network model. The optimization objective of the training is calculated using the following formula:

[0096] ;

[0097] in, The fusion loss is represented by a data-driven loss term. and physical prior constraint loss term It is obtained by weighted summation; This represents the data-driven loss term, used to penalize the prediction error of the graph neural network (GNN) on the observable data. This represents the weighting coefficient of the physical prior constraint loss, and its value is greater than 0. It is used in the fusion loss. Balanced data-driven loss term and physical prior constraint loss term The relative importance of the two items.

[0098] The physical prior constraint loss term This is the total physical loss of an N-axis system, and its specific calculation method is as follows:

[0099] ;

[0100] in, This represents the heat loss weighting coefficient, which has a value greater than 0 and is used in the physical prior constraint loss term. Consistent loss of equilibrium heat The relative importance; This represents the delay loss weighting coefficient, which has a value greater than 0 and is used in the physical prior constraint loss term. Medium-balanced delay consistency loss The relative importance; The summation operator over axis index i is used to accumulate the physical losses of all N servo axes, i.e., the thermal consistency loss. and delay consistency loss The total physical loss of the N-axis system is obtained, which is the physical prior constraint loss term. .

[0101] Thermal uniformity loss of axis i The aim is to utilize the thermal dynamic unit defined in step S1300. The constraint used to constrain the graph neural network (GNN) in identifying the junction temperature inside axis i is the thermodynamic basis for the GNN's ability to predict internal parameter drift and solve high-gain oscillation problems. The specific calculation formula is as follows:

[0102] ;

[0103] in, The summation operator over time index k is used to accumulate the thermal consistency loss of a single axis i at each time k over the time series of the training data; k represents the time index of the discrete-time system. The square of the L2 norm, i.e., the square of the Euclidean distance, is used to calculate the prediction value of a graph neural network (GNN). and physical model target value The distance between them; Representation of graph neural network model The junction temperature of the intelligent power unit IPM on axis i at time k+1 The predicted value.

[0104] Delay consistency loss of axis i The aim is to use the delay drift unit defined in step S1300 The total delay of the current loop calculated in step S1300 by the constraint graph neural network (GNN) The specific calculation formula for its identification is as follows:

[0105] ;

[0106] in, Representation of graph neural network model Total current loop delay at time k-axis i The predicted value.

[0107] Finally, by taking a snapshot of the input graph As input data, and with fusion loss As the optimization objective, a graph neural network model is trained to output a thermally coupled graph neural network model. That is, the optimized graph neural network model .

[0108] S2000, input image snapshot Input thermally coupled graph neural network model Output the set of hidden states After parameter identification decoder and dynamic identification decoder From the set of hidden states respectively Identify the total delay of the system-level current loop. and coupling torque model Combined with future motion instructions Predict the total system-level current loop delay for the next cycle. Coupled torque with the next cycle Through mapping function Fusion, output the next cycle composite perturbation voltage set .

[0109] Specifically, this step aims to utilize the thermally coupled graph neural network model output from step S1400. Input graph snapshots at time k, constructed in real time As input, the online identification system's current internal high-frequency characteristics and external coupling model are combined with future motion commands from the servo PLC motion controller. Predict the composite disturbance voltage of the next cycle .

[0110] Further, step S2000 includes:

[0111] Step S2100: Input image snapshot Input to thermally coupled graph neural network model In the thermally coupled graph neural network model Internally, L-layer graph neural network computation is performed, utilizing the input graph snapshot. Dynamic weighted adjacency matrix The neighborhood information is weighted and aggregated to output a set of hidden states. .

[0112] Specifically, this step aims to execute the thermally coupled graph neural network model output from step S1400. The forward propagation calculation takes a snapshot of the input graph at time k. Input into the thermally coupled graph neural network model The thermo-coupled graph neural network model is described in the text. Internally, through its learned message propagation mechanisms, such as multi-layer graph convolution, it aggregates neighbor information and updates node states layer by layer, ultimately outputting a set of time-k hidden states that can simultaneously represent force coupling and thermal drift information. .

[0113] Due to the thermal coupling graph neural network model In step S1400, a loss term including physical prior constraints is included. Fusion loss The optimized set of hidden states is the output. The thermo-mechanical coupling state information contained within the system, which conforms to physical laws, forms a reliable basis for the subsequent steps S2200 and S2300 to identify physical quantities.

[0114] In the specific implementation process, the input graph snapshot at time k is... Input into the pre-trained thermally coupled graph neural network model In the process, calculate and output the set of hidden states at time k. .

[0115] The thermal coupling graph neural network model Internal calculations typically involve A stack of graph neural network layers. Taking the first layer as an example. Taking layer calculation as an example, first, the nodes... In the Hidden state of a layer Through the learned weight matrix Perform a linear transformation. Secondly, for the nodes... Each neighbor node , and put it in the first Hidden state of a layer Through the learned weight matrix A linear transformation is performed, and the neighbor information after the linear transformation is weighted using normalized dynamic edge weights. All weighted neighbor information is then summed to complete information aggregation. Finally, the nodes... The linear transformation result and the neighbor node aggregation result are added together and then processed by a nonlinear activation function to finally obtain a single node. In the Hidden state of a layer The specific process formula is as follows:

[0116] ;

[0117] in, Represents a node At time k, the first The hidden state vector of the layer is used to carry the nodes. and its neighborhood in the Feature representation after layer fusion; Represents a thermally coupled graph neural network model The network layer index is a value ranging from 0 to... A non-negative integer, used to represent a thermally coupled graph neural network model. The hierarchy of internal calculations; Represents a thermally coupled graph neural network model The total number of floors; This represents a nonlinear activation function, used to introduce nonlinearity to enhance the expressive power of the model; Indicates the first Layers are used to transform nodes The weight matrix learned from its own features has parameters that are obtained by optimizing the fusion loss in step S1400. What I learned; Represents a node At time k, the first The hidden state vector of the layer; For nodes The set of neighboring nodes The summation operator is used to aggregate data from nodes. Information about the neighborhood; Represents a node The set of neighboring nodes, which is defined by the topological adjacency matrix in step S1200. Sure; The k-th dynamic weighted adjacency matrix represents the first... Line number Column elements; Representing the nodes respectively and nodes The degree; This represents the symmetric normalization factor, used to adjust the weights of neighbor information aggregation to prevent excessive differences in node degree from causing computational instability. Indicates the first Layers are used to transform neighboring nodes. The learned weight matrix of features has parameters that are also optimized in step S1400 by adjusting the fusion loss. What I learned; Represents a node At time k, the first The hidden state vector of the layer.

[0118] Through the above iterative calculations at level L, i.e. from arrive This yields the final output set of hidden states. .

[0119] Step S2200: Utilize the thermally coupled graph neural network model Parameter identification decoder From the set of hidden states The total delay of the second current loop on each axis was identified in the middle. And determine the total delay of the system-level current loop. Based on the total delay of the system-level current loop and the target closed-loop bandwidth of the servo current loop Calculate the stability margin of the system .

[0120] Specifically, this step aims to utilize the thermally coupled graph neural network model output from step S1400. Parameter identification decoder The set of hidden states output from step S2100 In real time, it identifies the system axes affected by internal parameter drift. The high-frequency characteristic change, i.e., the total delay of the second current loop. Based on this identification result, the stability margin of the current system is calculated. This provides a direct and quantitative basis for stability in subsequent steps.

[0121] The parameter identification decoder It is a thermally coupled graph neural network model. The components are optimized by adjusting the fusion loss during the training process in step S1400. The optimal parameters are learned. Specifically, during this training process, latency consistency loss is used. Force this parameter to identify the decoder Delay drift unit in output approximation step S1300 The calculation results enable the parameter identification decoder. It has the ability to accurately decode delayed information that conforms to physical laws from a hidden state.

[0122] In the specific implementation process, the first step is to perform axis-by-axis current loop total loop delay identification. During the real-time identification phase, each node... Hidden state vector Input to the pre-trained parameter identification decoder Decoded as the axis At any moment Total delay of the second current loop The identification results can reflect the impact of hardware thermal drift on latency characteristics in real time and in a physically consistent manner.

[0123] Based on the control requirements of multi-axis CNC machine tool systems, such as system stability assurance, the total delay set of the current loops for each axis is identified. In this context, a total system-level current loop delay that can represent the overall or critical stability margin of the current system is determined. The preferred conservative strategy is to select the largest total delay of the second current loop among all axes as the estimate of the system delay, i.e. .in, Indicates time The system-level total current loop delay is determined based on the total second current loop delay identified by each of the N servo axes. Determined to assess the overall dynamic stability of the system; This represents the maximum value operator, used to determine the most unfavorable delay state of the system under multi-axis cooperative operation conditions.

[0124] Based on the system-level current loop total delay By applying servo control theory, a quantitative relationship between delay and stability margin is established, thereby calculating the stability margin of the current servo current loop. The aforementioned servo control theory states that the total delay of the system-level current loop... This introduces a phase lag proportional to the frequency, which reduces the system's stability margin. The specific process formula is as follows:

[0125] ;

[0126] in, Indicates the calculated time The system's stability margin is used to quantify how far the system is from an unstable state, such as high-frequency oscillations. This represents the arctangent function, used to calculate the total delay caused by the system-level current loop. At the target closed-loop bandwidth The phase lag introduced at that point; The target closed-loop bandwidth of the servo current loop is a system design parameter, measured in radians per second (rad / s), and is a key frequency point used to evaluate stability. This represents pi, and in this case, it is used to express the arctangent function. The output radian value is converted into an angle value; Indicates 180 degrees, used with... This is used to perform unit conversion from radians to degrees.

[0127] Step S2300: Utilize the thermally coupled graph neural network model Dynamics identification decoder in From the set of hidden states Decode the coupling torque model Real-time calculation of coupling torque estimates .

[0128] Specifically, this step aims to utilize the thermally coupled graph neural network model output from step S1400. Dynamics identification decoder in The set of hidden states output from step S2100 In the process, the dynamic model describing the external coupled disturbance, namely the coupled torque model, is identified in real time. This provides a quantifiable basis for external dynamic coupling compensation in subsequent steps.

[0129] The dynamic identification decoder For thermally coupled graph neural network models An important component, whose parameters are optimized in step S1400 by adjusting the fusion loss. Learned. Data-driven loss term. This data plays a crucial role in driving the loss term. Constrained Dynamics Identification Decoder The system's state at the next moment must be predicted, thus forcing the dynamic identification decoder to... During training, the system dynamics, including inter-axis coupling, are implicitly learned. Therefore, this identification decoder... Its purpose is to extract implicit, complex coupled dynamics from the hidden state set. The model is decoded and expressed in an explicitly usable form as a usable coupled torque model. .

[0130] In the specific implementation process, during the real-time identification phase, the set of hidden states is... Input to the pre-trained dynamics recognition decoder In the middle, it is decoded into a value that can represent the current moment. Nonlinear model of system coupling characteristics, i.e., coupling torque model .

[0131] Identified coupling torque model It has the following function: given the target axis index and its neighboring nodes When the state information is available, the model can calculate in real time the force exerted by k on the axis at the current moment. Estimated coupling torque on Preferably, the hidden state calculated in step S2100 is used as input, i.e., the node... The hidden state vector at time k Because it incorporates a thermally coupled graph neural network model. Extracted high-level physical characteristic information.

[0132] The coupling torque model Capable of capturing neighbor axes The motion of the axis, such as position and velocity, is transmitted to the axis through physical mechanisms such as inertia and Coriolis force. The complex nonlinear coupling effect on the surface. Due to the coupling torque model... The input depends on the set of hidden states. It can adaptively reflect the edge weight function that was not constructed in step S1200. It fully captures time-varying factors, such as load changes and friction changes, thereby enabling dynamic modeling and real-time impact of external high-frequency disturbances.

[0133] Step S2400: Utilize future motion commands from the servo PLC motion controller. and system-level current loop total delay Predict the total delay of the system-level current loop in the next cycle. Based on the coupling torque model Predicting the coupling torque in the next cycle The total delay of the system-level current loop in the next cycle will be... and the coupling torque of the next cycle Through mapping function The fusion is performed to output the composite perturbation voltage set for the next cycle. .

[0134] Specifically, this step aims to address the system-level total current loop delay identified in step S2200. Coupled torque model output from step S2300 And combined with future motion commands from the servo PLC motion controller Predict the next control cycle The composite disturbance acting on each axis is considered, taking into account both the thermal drift effect caused by temperature rise and the force coupling effect caused by multi-axis interactions. This composite disturbance is then mapped to the composite disturbance voltage set for the next cycle. This provides an input basis for adaptive compensation of the current loop.

[0135] In the specific implementation process, firstly, the next control cycle is predicted. System-level current loop total delay The goal of this prediction is to estimate the total system-level current loop delay caused by thermal effects. The predicted value for the next period can be obtained using the following two methods:

[0136] Method 1: Prediction based on the current state. When the total current loop delay changes relatively slowly within adjacent control cycles, the system-level total current loop delay identified in step S2200 can be used directly. As the predicted value for the next cycle, i.e. .

[0137] Method 2: Prediction based on a physical model. This method follows the S1300 to establish a physical causal chain of electricity-thermal-delay, utilizing the motion controller from the servo PLC in the current cycle. Calculated and used for the next cycle Future motion instructions Especially reference shaft current This enables more accurate forward-looking forecasts. Specifically, it includes the following sub-steps:

[0138] First, predict the junction temperature of the intelligent power unit (IPM) for each axis in the next cycle. Based on the current moment Status, such as the junction temperature of the intelligent power unit IPM output in step S1300. and the data collected in step S1100 Axis current components and future motion instructions References in shaft current Estimate the instantaneous power loss in the next cycle. The predicted instantaneous power loss Input to thermal dynamics unit In the middle, the next cycle is calculated. Predicted junction temperature of the intelligent power unit IPM .

[0139] Second, predict the total delay of each axis current loop in the next cycle. The predicted junction temperature of the intelligent power unit (IPM) is obtained. And the predicted temperature value of the current sensor using a similar method described above. Input to the delay drift unit established in step S1300 In the middle, the calculation is obtained Total predicted delay of each axis current loop in the cycle .

[0140] Third, determine Periodic system-level latency The predicted total delay of each axis current loop. In this process, the system-level latency is determined based on system stability requirements. For example, a conservative strategy can be adopted to select the maximum value.

[0141] Secondly, predict the effect of the next control cycle on each servo axis. Coupling torque on The coupling torque model identified in step S2300 is used. and future motion commands from the servo PLC motion controller This will be achieved through the following sub-steps:

[0142] First, estimate the instantaneous state vector of the neighbor in the next cycle. Based on the axis data acquired in step S1100 The instantaneous state vector at time k and future motion instructions for the next cycle Estimate the next moment neighbor axis The instantaneous state vector.

[0143] Second, apply the coupled torque model. Make a prediction: estimate the instantaneous state vector of the neighbors in the next cycle. Input to the identified coupled torque model In the middle, the next cycle is calculated. Acting on the shaft Coupling torque on .

[0144] Finally, the next control cycle is applied to each axis. Coupling torque on and the total delay of the system-level current loop in the next cycle Through mapping function This is collectively converted into an electrical domain compensation quantity that the servo PLC motion controller can directly operate, i.e., the composite disturbance voltage set for the next cycle. The specific process formula is as follows:

[0145]

[0146] ;

[0147] in, Indicates the predicted next period The composite perturbation voltage set contains all Predicted equivalent voltage disturbance values ​​for each axis of the dq axis; Indicates the predicted next period axis of The shaft equivalent disturbance voltage is a scalar value, measured in volts (V). Indicates the predicted next period axis of The shaft equivalent disturbance voltage is a scalar value, measured in volts (V). This represents a mapping function used to convert predicted physical domain perturbations into equivalent dq-axis voltage perturbations in the electrical domain. This represents the set of motor parameters, including but not limited to torque constant, d / q-axis inductance, stator resistance, etc., and is a mapping function. The prior physical parameters necessary to perform a physical-to-electrical conversion; Indicates time The controller's internal state, such as the accumulated value of the integrator, is a mapping function. Estimating the total delay of the system-level current loop in the next cycle The resulting voltage deviation may require additional input to improve estimation accuracy.

[0148] The mapping function Two aspects need to be considered comprehensively, namely, the torque disturbance to The effect of axis voltage conversion and delay on d / q axis voltage.

[0149] The torque disturbance to Shaft voltage conversion, predicted coupling torque The main manifestation is to An external disturbance to the shaft current loop. To counteract this disturbance and maintain the desired... The shaft current requires the application of an additional... Axis voltage. This mapping function. Using motor parameter set For example, torque constant, Axial inductance, resistance, and current operating point, such as time. motor speed Calculate the coupling torque to offset the prediction The expected equivalent The axis voltage disturbance component is included in the prediction. Shaft equivalent disturbance voltage .

[0150] The effect of the delay on the d / q axis voltage, and the predicted total delay of the system-level current loop. This can cause controllers, such as PI controllers, to... The control voltage calculated periodically will exhibit phase lag and potential amplitude deviation compared to the zero-delay case. This mapping function... Based on the predicted total system-level current loop delay The dynamic characteristics of the controller, such as those determined by the target bandwidth or PI gain parameter, and the time... Controller internal state To calculate the predicted total delay of the system-level current loop. The equivalent d / q-axis voltage deviation dynamically generated by the controller. This equivalent d / q-axis voltage deviation also needs to be calculated and added to the prediction. Shaft equivalent disturbance voltage and prediction Shaft equivalent disturbance voltage middle.

[0151] S3000, based on future motion instructions Next cycle reference shaft current and reference shaft current Motor speed and motor parameter set The standard feedforward voltage is calculated using the standard field-oriented control model, combined with the composite disturbance voltage set. Component-by-component superposition is performed to generate and output a set of feedforward compensation voltages. .

[0152] Specifically, this step aims to implement future motion commands from the servo PLC motion controller. each axis Reference shaft current and reference shaft current The motor speed collected in step S1100 and the composite disturbance voltage set predicted in step S2400 Calculate and output the next cycle. Feedforward compensation voltage set This feedforward compensation voltage set By actively superimposing the standard FOC feedforward voltage and the predicted composite disturbance voltage set This enables proactive cancellation of known and predicted disturbances in the next cycle, thereby maintaining the high responsiveness and high-precision tracking performance of the current loop in high-frequency dynamic environments.

[0153] Further, step S3000 includes:

[0154] Step S3100, utilizing future motion commands Next cycle reference shaft current and reference shaft current Collected motor speed and motor parameter set Applying the standard magnetic field orientation control model, calculations were performed to include standard... Shaft feedforward voltage component and standards Shaft feedforward voltage component The standard feedforward voltage.

[0155] Specifically, this step aims to implement future motion commands from the servo PLC motion controller. The next cycle Each axis Reference shaft current and reference shaft current The motor speed collected in step S1100 and motor parameter set Calculate and output each axis based on the standard magnetic field orientation control model. In the next cycle Standard feedforward voltage , including standards Shaft feedforward voltage component and standards Shaft feedforward voltage component Two components. The standard feedforward voltage is designed to compensate for known dynamic effects within the motor in advance, based on an ideal motor model, namely the cross-coupling voltage between the d and q axes and the back electromotive force voltage generated by the permanent magnet, thereby counteracting the inherent dynamic effects within the motor in advance.

[0156] In the specific implementation process, the standard field-oriented control model decoupling and back EMF feedforward formula are applied for each servo axis. Calculate the next cycle Standard feedforward voltage The aforementioned standard Shaft feedforward voltage component The calculation is based on the standard field-oriented control model: when the motor rotates at a speed of... Rotate, and Axial flow through reference shaft current At that time, The shaft induces a cross-coupled electromotive force, the value of which is To counteract this internal induced voltage, it is necessary to... A feedforward voltage of equal magnitude but opposite sign is applied to the shaft, i.e., the standard... Shaft feedforward voltage component .in, Indicates axis Motor nominal Shaft inductance, the value of which is derived from the motor parameter set. The unit is Henry (H), used for calculations. In the shaft feedforward voltage Axis cross-coupling term.

[0157] The standard Shaft feedforward voltage component It includes two compensatory aspects: one is compensation Cross-axis coupled electromotive force. When Axial flow through reference shaft current At that time, due to the motor speed and nominal Shaft inductor , will A cross-coupled electromotive force is induced in the shaft. Second, it compensates for the back electromotive force. Permanent magnet flux linkage The motor operates at the motor speed. When rotating, a back electromotive force proportional to the velocity is generated. Its main function is Axis. Among them, Indicates axis Motor nominal Shaft inductance, the value of which is derived from the motor parameter set. The unit is Henry (H), used for calculations. In the shaft feedforward voltage Axis cross-coupling terms; Indicates axis The nominal flux linkage of the permanent magnet in the motor is determined by the set of motor parameters. The unit is Weber (Wb), used for calculation. The back electromotive force term in the shaft feedforward voltage.

[0158] Step S3200: Combine the standard feedforward voltage and the composite disturbance voltage. Each component is added separately, and the result is combined to obtain the product containing... Shaft feedforward compensation voltage and Shaft feedforward compensation voltage Feedforward compensation voltage set .

[0159] Specifically, this step aims to convert the standard feedforward voltage output from step S3100 into a standard feedforward voltage. The composite disturbance voltage set predicted in step S2400 The layers are stacked to generate and output the final product for the next cycle. Feedforward compensation voltage set of control loop This is the core component for achieving predictive composite disturbance compensation.

[0160] In the specific implementation process, for each axis Standard feedforward voltage Chinese Standard Shaft feedforward voltage component and standards Shaft feedforward voltage component , respectively with the corresponding composite disturbance voltage set The corresponding Shaft equivalent disturbance voltage and Shaft equivalent disturbance voltage Adding the components one by one, we get the following results. Shaft feedforward compensation voltage and Shaft feedforward compensation voltage The specific process formula is as follows:

[0161] ;

[0162] ;

[0163] Finally, the feedforward compensation voltages calculated for all N axes are combined to obtain the set of feedforward compensation voltages for the multi-axis CNC machine tool system. This feedforward compensation voltage set In the next cycle The voltage output injected into the current loop, for example superimposed on the output of the PI controller, is used as part of the final voltage command applied to the inverter to actively counteract ideal model effects and predicted compound disturbances.

[0164] S4000, Calculation Axis current tracking error and Axis current tracking error Based on stability margin The comparison result with the preset safety threshold is processed by the gain scheduler. Dynamic output proportional gain and integral gain Perform PI control calculations and calculate the set of feedback control voltage components. The feedback control voltage component set and feedforward compensation voltage set Superimpose and synthesize the total command voltage The comparison results are used to determine the total command voltage. Apply dynamic safety limiting and output a set of safety voltage commands. .

[0165] Specifically, this step aims to achieve adaptive feedback control of the current loop, based on the future motion commands from the servo PLC motion controller. The next cycle Each axis Reference shaft current and reference shaft current The d-axis current component acquired in step S1100 and q-axis current component The current error between them, combined with the stability margin identified in step S2200. Adaptive scheduling of the PI gain and dynamic limiting of the synthesized voltage are performed, combined with the feedforward compensation voltage set output from step S3200. Calculate and output the set of safe voltage commands applied to the inverter. .

[0166] Further, step S4000 includes:

[0167] Step S4100, calculate the next cycle. refer to shaft current and reference shaft current Respectively with the current time Collected Axis current components and Axis current components The difference between them, generate Axis current tracking error and Axis current tracking error .

[0168] Specifically, this step aims to be based on the next cycle. Each axis Reference shaft current and reference shaft current and the current cycle of Axis current components and Axis current components Calculate the current time of Axis current tracking error and Axis current tracking error The current tracking error mainly reflects the deviation caused by residual disturbances that were not completely eliminated by the feedforward compensation in step S3000, as well as model uncertainties.

[0169] In the specific implementation process, for each axis Calculate the next cycle refer to shaft current and reference shaft current , with the current moment Actual measured value, i.e., collected Axis current components and Axis current components The difference between them gives the current time. The current tracking error, i.e. Axis current tracking error and Axis current tracking error .

[0170] Step S4200: Compare stability margins And a preset safety threshold, if stability margin If the value exceeds the preset safety threshold, the gain scheduler will be activated. The output scaling gain and integral gain Set to the first proportional gain respectively and first integral gain If stability margin If the output is below a preset safety threshold, the proportional gain will be reduced. and integral gain Set to the second proportional gain respectively Second integral gain .

[0171] Specifically, this step aims to address the stability margin identified in real time based on step S2200. Dynamically adjust the gain parameter of the PI controller, i.e., the proportional gain. and integral gain Under the premise of prioritizing system stability, the feedback control loop calculates the residual error in step S4100, i.e. Axis current tracking error and Axis current tracking error This maximizes suppression, thus balancing dynamic performance and robustness.

[0172] In the specific implementation process, at all times Stability margin of identification Input to gain scheduler The gain scheduler. Preferably, it is a neural network model trained offline, but lookup tables or rule-based logistic functions can also be used. Its core function is to establish a stability margin. The mapping relationship to the optimal PI gain parameter. Through this mapping relationship, the system can, within each control cycle, adjust the real-time stability margin. Output a set of dynamically adjusted gain parameters, i.e., proportional gain. and integral gain Among them, the aforementioned Indicates axis At any moment The proportional gain obtained by adaptive scheduling varies with the stability margin. The dynamically changing nonnegative scalar value; Indicates axis At any moment The integral gain obtained by adaptive scheduling varies with the stability margin. A dynamically changing nonnegative scalar value.

[0173] The gain scheduler The core idea of ​​the scheduling logic is: based on real-time stability margin The control intensity is dynamically adjusted as follows:

[0174] If stability margin When the value exceeds the preset safety threshold, it indicates that the system currently has sufficient stability margin and is far from the risk of high-frequency oscillation. At this time, the gain scheduler... The output scaling gain and integral gain The first proportional gain is set to a higher gain value respectively. Second integral gain This is to improve the bandwidth and error decay rate of the feedback control loop, and quickly eliminate the residual error calculated in step S4100, i.e. Axis current tracking error and Axis current tracking error This enables high dynamic response and high-precision tracking.

[0175] If stability margin If the value is below or close to the preset safety threshold, it indicates that the current stability margin of the system is insufficient and there is a risk of high-frequency self-oscillation. This is usually caused by the total delay of the second current loop due to the thermal drift effect in step S1300. Increase. At this time, the gain scheduler Automatically reduce the gain value, adjusting the output proportional gain. and integral gain The second proportional gain is set to the lower gain value respectively. Second integral gain Actively narrowing the loop bandwidth, with a stability-first strategy, suppresses possible high-frequency self-excitation phenomena, ensuring the safe operation of the system under complex thermoelectric dynamic conditions.

[0176] Step S4300, based on Axis current tracking error and Axis current tracking error And adopt proportional gain and integral gain Perform discrete-time PI control calculations to determine the next cycle. Feedback control voltage component set .

[0177] Specifically, this step aims to calculate the current residual error based on step S4100, i.e. Axis current tracking error and Axis current tracking error And the current adaptive gain parameter dynamically scheduled in step S4200, i.e., the proportional gain. and integral gain Perform discrete-time PI control calculations, calculate and output the results for the next control cycle. Feedback control voltage component set This enables real-time correction and dynamic compensation of the current tracking error in the current cycle.

[0178] In the specific implementation process, for each servo axis of shaft and The axes are each applied with a discrete-time PI control law. This control law is based on the current time... Using the current tracking error and adaptive gain parameters as inputs, the next control cycle is calculated. The feedback control voltage output serves as the current moment's... The immediate response to the current tracking error. The feedback control voltage consists of a proportional term and an integral term.

[0179] by Taking the axis as an example, the next cycle of Shaft feedback control voltage The calculation formula is as follows:

[0180] ;

[0181] in, Indicates axis Calculated for the next cycle of Shaft feedback control voltage, measured in volts (V), is used to characterize the feedback control pair. Correction amount for shaft voltage; express Shaft feedback control voltage The proportional term is the term at the current moment. Adaptive proportional gain Multiply by the current time of Axis current tracking error ; express Shaft feedback control voltage The integral term is the axis. At any moment of The integral term of the axis PI controller, usually in volts (V) or volt-seconds (V·s), is used to eliminate potential steady-state deviations in the system.

[0182] The Shaft feedback control voltage integral terms It is at every moment based on Axis current tracking error and the current adaptive integral gain The update is performed using a discrete-time cumulative method, and an integral approximation in the form of backward difference is adopted. The specific formula is as follows:

[0183] ;

[0184] in, Indicates axis In the previous moment of The cumulative integral term of the axis PI controller is used to calculate the current time. Cumulative value of integral terms The foundation; This indicates the control period, measured in seconds (s). Axis current tracking error The duration of the action is used for discrete-time integration calculations.

[0185] Similarly, Shaft feedback control voltage The calculation method and The axes are similar, and the same PI control law is applied for calculation, using the corresponding... Axis current tracking error and adaptive proportional gain and integral gain .

[0186] The next control cycle is formed by performing the above calculations on all N servo axes. Feedback control voltage component set This enables real-time closed-loop regulation of the entire system's current loop.

[0187] Step S4400: Set the feedforward compensation voltage. and feedback control voltage component set Superimpose the values ​​to synthesize the total command voltage. Comparison of stability margins And a preset margin threshold, if stability margin If the value exceeds the preset margin threshold, the dynamic safety limiting unit will be activated. Applying the first limiting policy, if the stability margin If the value falls below a preset margin threshold, the dynamic safety limiting unit will be activated. Apply a second limiting strategy to output a safe voltage command set. .

[0188] Specifically, this step aims to collect the feedforward compensation voltage output from step S3200. The set of feedback control voltage components output by step S4300 The synthesis is performed to obtain the preliminary total command voltage. And based on the stability margin identified in real time in step S2200 For the total command voltage Apply dynamic safety limiting and output the final set of safety voltage commands applied to the inverter. .

[0189] In the specific implementation process, for each servo axis , collect feedforward compensation voltages and feedback control voltage component set By adding the components one by one, the next period can be obtained. Total command voltage before limiting .in, Indicates the next period At that time, servo axis Total command voltage of Axis voltage components; Indicates the next period At that time, servo axis Total command voltage of Axis voltage component.

[0190] The initially synthesized total command voltage Input to a dynamic safety limiting unit The dynamic safety limiting unit. Stability margin of synchronous reception and real-time identification This serves as a constraint input signal for limiting scheduling. The dynamic safety limiting unit... By comparing stability margins The unit selectively applies two limiting strategies based on a preset margin threshold. For the total command voltage Apply a non-linear, dynamically adjustable constraint.

[0191] The core idea of ​​the dynamic adjustment is to apply a flexible limit to the output voltage command using the real-time steady state as an adaptive constraint benchmark, so as to ensure that the system can obtain the optimal trade-off between safety and performance under different steady states. The specific logic is as follows:

[0192] If stability margin If the value is higher than the preset margin threshold, it indicates that the system stability margin is sufficient, and the dynamic safety limiting unit... Apply a first limiting strategy. This first limiting strategy is a relatively lenient constraint, limited only to the inverter's physical maximum output voltage and the maximum voltage change rate allowed by the hardware, allowing the voltage command to change rapidly, thereby fully utilizing the system's high-frequency response performance and dynamic tracking capability.

[0193] If stability margin If the value is lower than or equal to a preset margin threshold, it indicates that the system is approaching the edge of instability, for example, due to thermal drift causing a total delay in the current loop. Increase the dynamic safety limiting unit A second limiting strategy is actively applied. This second limiting strategy is a stricter voltage change rate limit or amplitude limit, which actively smooths the final voltage command change, suppresses any high-frequency components or abrupt changes that may exist, and prevents the drastic action of the voltage command itself from triggering potential unstable modes or causing excessive current surges under low stability margins, thereby providing the system with final, dynamic safety protection.

[0194] Through the above adjustments, the system achieves an adaptive balance between stability and response speed, making the limiting process no longer a fixed constraint, but a flexible protection mechanism that dynamically changes with the steady state.

[0195] Finally, based on real-time stability margin After dynamically adjusting the safety limit, output a set of safety voltage commands. And input to the pulse width modulation (PWM) unit to generate the next cycle. The inverter switching signal enables adaptive coordination among control accuracy, dynamic response, and hardware safety.

[0196] Example 2:

[0197] This embodiment, based on Embodiment 1, provides a high-frequency response current loop adaptive control system for a servo drive system, such as... Figure 2 As shown, it includes a model building module, a discrimination and prediction module, a feedforward compensation module, and an adaptive feedback module;

[0198] The model building module is used to collect a set of instantaneous state vectors. Combined with kinematic configuration parameters Construct a dynamic weighted adjacency matrix Utilizing the driver hardware thermal parameters Constructing a set of physical prior constraints The instantaneous state vector set The dynamic weighted adjacency matrix and the set of physical prior constraints Jointly construct input graph snapshot Define the physical prior constraint loss term , and data-driven loss terms Fuse and train graph neural networks to output a thermally coupled graph neural network model. .

[0199] The discrimination and prediction module is used to capture the input image snapshot. Input thermally coupled graph neural network model Output the set of hidden states After parameter identification decoder and dynamic identification decoder From the set of hidden states respectively Identify the total delay of the system-level current loop. and coupling torque model Combined with future motion instructions Predict the total system-level current loop delay for the next cycle. Coupled torque with the next cycle Through mapping function Fusion, output the next cycle composite perturbation voltage set .

[0200] The feedforward compensation module is used to adjust the parameters based on future motion commands. Next cycle reference shaft current and reference shaft current Motor speed and motor parameter set The standard feedforward voltage is calculated using the standard field-oriented control model, combined with the composite disturbance voltage set. Component-by-component superposition is performed to generate and output a set of feedforward compensation voltages. .

[0201] The adaptive feedback module is used to calculate Axis current tracking error and Axis current tracking error Based on stability margin The comparison result with the preset safety threshold is processed by the gain scheduler. Dynamic output proportional gain and integral gain Perform PI control calculations and calculate the set of feedback control voltage components. The feedback control voltage component set and feedforward compensation voltage set Superimpose and synthesize the total command voltage The comparison results are used to determine the total command voltage. Apply dynamic safety limiting and output a set of safety voltage commands. .

[0202] The parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0203] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A high-frequency response current loop adaptive control method for a servo drive system, characterized in that, include: A set of instantaneous state vectors is collected and combined with kinematic configuration parameters to construct a dynamic weighted adjacency matrix. A set of physical prior constraints is constructed using the thermal parameters of the driver hardware. The set of instantaneous state vectors, the dynamic weighted adjacency matrix, and the set of physical prior constraints are used to construct an input graph snapshot. A physical prior constraint loss term is defined, which is fused with the data-driven loss term and used to train a graph neural network. A thermally coupled graph neural network model is then output. The input graph snapshot is input into the thermally coupled graph neural network model, and the output hidden state set is processed by the parameter identification decoder and the dynamic identification decoder to identify the system-level current loop total delay and coupling torque model from the hidden state set. Combined with future motion commands, the system-level current loop total delay and coupling torque of the next cycle are predicted. The next cycle composite disturbance voltage set is output by fusing through the mapping function. Reference for the next cycle based on future motion instructions shaft current and reference The set of shaft current, motor speed and motor parameters is used to calculate the standard feedforward voltage using the standard field-oriented control model. The standard feedforward compensation voltage set is generated and output by combining the composite disturbance voltage set component by component. calculate Axis current tracking error and The shaft current tracking error is calculated by dynamically outputting the proportional gain and integral gain through the gain scheduler based on the comparison results of the stability margin and the preset safety threshold. PI control operation is performed to calculate the set of feedback control voltage components. The set of feedback control voltage components and the set of feedforward compensation voltage components are superimposed to synthesize the total command voltage. Dynamic safety limiting is applied to the total command voltage using the comparison results, and a safe voltage command set is output.

2. The high-frequency response current loop adaptive control method for the servo drive system according to claim 1, characterized in that, The output method of the thermally coupled graph neural network model includes: Synchronously collect all The electrical, mechanical, and thermal state data of each servo axis are used to construct an instantaneous state vector set; The electrical state includes Axis current components and The shaft current component, the mechanical state includes motor speed and motor position, and the thermal state includes the physical temperature of the intelligent power unit and the temperature of the current sensor; Based on the kinematic configuration parameters, the physical connection between servo axes is analyzed, a static topological adjacency matrix is ​​constructed, and the dynamic weight is calculated by using the real-time position vector in the instantaneous state vector set to drive the state-dependent edge weight function. Instantaneous power loss is calculated based on electrical and thermal state data of instantaneous state vector set. A thermal dynamic unit is constructed in combination with driver hardware thermal parameters to estimate hardware temperature. Based on the hardware temperature, the total delay of current loop is calculated through delay drift unit to construct physical prior constraint set. The instantaneous state vector set is used as node features, the dynamic weighted adjacency matrix is ​​used as edges, and the input graph snapshot is constructed by combining the physical prior constraint set. The physical prior constraint loss term is defined based on the physical prior constraint set and fused with the data-driven loss term to form a fusion loss. The fusion loss is used as the optimization objective and the input graph snapshot is used as the input to train the graph neural network and output the thermally coupled graph neural network model.

3. The high-frequency response current loop adaptive control method for the servo drive system according to claim 2, characterized in that, The rule for assigning values ​​to elements in the static topological adjacency matrix is ​​that the element of the corresponding static topological adjacency matrix is ​​assigned a value of 1 if and only if the system kinematic configuration parameters determine the motion of one axis and have a dynamic effect on the other axis; otherwise, if the two axes are decoupled in dynamics and have no direct kinematic dependence, the element is assigned a value of 0.

4. The high-frequency response current loop adaptive control method for a servo drive system according to claim 1, characterized in that, The method for outputting the next cycle composite perturbation voltage set includes: The input graph snapshot is fed into the thermally coupled graph neural network model, and execution is performed within the thermally coupled graph neural network model. The layered graph neural network computes by using the dynamic weighted adjacency matrix in the input graph snapshot to weight and aggregate neighborhood information in order to output a set of hidden states; Using the parameter identification decoder in the thermally coupled graph neural network model, the total delay of the second current loop of each axis is identified from the set of hidden states, and the total delay of the system-level current loop is determined. Based on the total delay of the system-level current loop and the target closed-loop bandwidth of the servo current loop, the system stability margin is calculated. By using the dynamic identification decoder in the thermo-coupled graph neural network model, the coupling torque model is decoded from the set of hidden states, and the estimated value of the coupling torque is calculated in real time. By utilizing the future motion commands of the servo PLC motion controller and the total delay of the system-level current loop, the total delay of the system-level current loop in the next cycle is predicted. Based on the coupling torque model, the coupling torque in the next cycle is predicted. The total delay of the system-level current loop in the next cycle and the coupling torque in the next cycle are fused through a mapping function to output the composite disturbance voltage set for the next cycle.

5. The high-frequency response current loop adaptive control method for a servo drive system according to claim 4, characterized in that, The mapping function comprehensively considers torque disturbances. The effects of axis voltage conversion and delay on d / q axis voltage; The torque disturbance to The conversion of shaft voltage cancels out the predicted coupling torque. External disturbances to the shaft current loop are applied to maintain the desired... shaft current, apply an additional The shaft voltage, the mapping function uses the set of motor parameters to calculate the equivalent of the predicted coupling torque to offset the prediction. Axis voltage disturbance component, and include the predicted Shaft equivalent disturbance voltage; The effect of the delay on the d / q axis voltage is determined by the mapping function, which calculates the equivalent d / q axis voltage deviation dynamically generated by the predicted total system-level current loop delay acting on the controller, based on the predicted total system-level current loop delay, the controller's dynamic characteristics, and the current internal state of the controller. This deviation is then added to the predicted total system-level current loop delay. Shaft equivalent disturbance voltage and prediction In the shaft equivalent disturbance voltage.

6. The high-frequency response current loop adaptive control method for a servo drive system according to claim 1, characterized in that, The method for generating and outputting the feedforward compensation voltage set includes: Utilizing the next cycle reference of future motion commands shaft current and reference The shaft current, the collected motor speed, and the set of motor parameters are used to calculate, using a standard field-oriented control model, including standard... Shaft feedforward voltage components and standard Standard feedforward voltage for the shaft feedforward voltage component; The standard feedforward voltage and the composite disturbance voltage set are added component by component and combined to obtain a result containing... Shaft feedforward compensation voltage and The set of feedforward compensation voltages for shaft feedforward compensation voltages.

7. The high-frequency response current loop adaptive control method for a servo drive system according to claim 6, characterized in that, The standard Shaft feedforward voltage components include compensation Two compensating properties: cross-coupled electromotive force and compensated back electromotive force. The compensation The cross-coupled electromotive force is when the axis is Axial flow through reference When the shaft current is low, due to the motor speed and nominal... Shaft inductance, in A cross-coupled electromotive force is induced in the shaft; The aforementioned compensating back electromotive force is a back electromotive force generated by the permanent magnet flux linkage when the motor rotates at a certain speed, which is proportional to the speed and mainly acts on... axis.

8. The high-frequency response current loop adaptive control method for a servo drive system according to claim 1, characterized in that, The output method of the safe voltage command set includes: Calculate the reference for the next period shaft current and reference The axis currents are respectively compared with the data collected at the current time. Axis current components and The difference between the shaft current components generates Axis current tracking error and Shaft current tracking error; The stability margin is compared with the preset safety threshold. If the stability margin is higher than the preset safety threshold, the proportional gain and integral gain of the output are set to the first proportional gain and the first integral gain respectively by the gain scheduler. If the stability margin is lower than the preset safety threshold, the proportional gain and integral gain of the output are set to the second proportional gain and the second integral gain respectively. in accordance with Axis current tracking error and The shaft current tracking error is investigated, and discrete-time PI control is performed using proportional gain and integral gain to calculate the set of feedback control voltage components for the next cycle. The feedforward compensation voltage set and the feedback control voltage component set are superimposed to synthesize the total command voltage. The stability margin is compared with the preset margin threshold. If the stability margin is higher than the preset margin threshold, a first limiting strategy is applied through the dynamic safety limiting unit. If the stability margin is lower than the preset margin threshold, a second limiting strategy is applied through the dynamic safety limiting unit, and a safe voltage command set is output.

9. A high-frequency response current loop adaptive control system based on a servo drive system, used to implement the high-frequency response current loop adaptive control method of the servo drive system according to any one of claims 1-8, characterized in that, The system includes a model building module, a discrimination and prediction module, a feedforward compensation module, and an adaptive feedback module; The model building module is used to collect a set of instantaneous state vectors, combine them with kinematic configuration parameters, construct a dynamic weighted adjacency matrix, construct a set of physical prior constraints using the driver hardware thermal parameters, construct an input graph snapshot using the set of instantaneous state vectors, the dynamic weighted adjacency matrix, and the set of physical prior constraints, define a physical prior constraint loss term, fuse it with a data-driven loss term and train a graph neural network, and output a thermally coupled graph neural network model. The discrimination and prediction module is used to input the input graph snapshot into the thermally coupled graph neural network model, output a set of hidden states, and identify the system-level current loop total delay and coupling torque model from the set of hidden states through the parameter identification decoder and the dynamic identification decoder, respectively. Combined with future motion commands, it predicts the system-level current loop total delay and coupling torque of the next cycle, and outputs the composite disturbance voltage set of the next cycle through mapping function fusion. The feedforward compensation module is used to reference the next cycle of future motion commands. shaft current and reference The set of shaft current, motor speed and motor parameters is used to calculate the standard feedforward voltage using the standard field-oriented control model. The standard feedforward compensation voltage set is generated and output by combining the composite disturbance voltage set component by component. The adaptive feedback module is used for calculation. Axis current tracking error and The shaft current tracking error is calculated by dynamically outputting the proportional gain and integral gain through the gain scheduler based on the comparison results of the stability margin and the preset safety threshold. PI control operation is performed to calculate the set of feedback control voltage components. The set of feedback control voltage components and the set of feedforward compensation voltage components are superimposed to synthesize the total command voltage. Dynamic safety limiting is applied to the total command voltage using the comparison results, and a safe voltage command set is output.

Images