A high-precision robust control system and method for a megawatt large-torque motor

Abstract

The application discloses a megawatt large-torque motor high-precision robust control system and method, relates to the technical field of motor driving and algorithm control cooperation, and comprises the following steps: collecting multi-source operation data, performing electromechanical parameter cross-domain reconstruction and state decoupling, and outputting current motor state quantity; performing cross-domain decoupling closed loop output decoupling error and dynamic parameter set; performing space-time state reconstruction and disturbance analysis; combining algorithm power margin ratio, performing dimension reduction reconstruction and feature compensation to generate equivalent disturbance parameters and reconstruction instruction set; modulating disturbance to generate front-end weight, and outputting voltage reference component through amplitude compensation and manifold evolution; and outputting switch signal to power device through cross-domain mapping and over-limit edge closed loop reconstruction deduction carrier frequency. The application overcomes the limitations of traditional methods, such as ignoring cross-domain coupling phase shift, fixed gain observation lag and boundary over-limit distortion, and provides accurate decision basis for megawatt large-torque motor high-precision control.

Description

Technical Field

[0001] This invention relates to the field of high-power power electronics and motor drive technology, and in particular to a high-precision robust control system and method for a megawatt-level high-torque motor. Background Technology

[0002] With the increasing complexity and dynamism of megawatt-level high-torque motor applications, traditional motor control technologies, relying solely on single time-domain errors and static parameter models for state assessment, struggle to accurately quantify the complex coupling losses and multi-source time-varying disturbances between electromechanical parameters. Existing methods utilize fixed-gain observers or conventional proportional-integral algorithms for error adjustment, completely ignoring physical characteristics such as standing wave reflections at high- and low-frequency disturbance boundaries, cross-spatial phase shifts caused by dynamic inductance parameter perturbations, and implicit characteristic attenuation under critical computing power constraints. This control approach, solely based on time-domain statistical indicators, leads to spatiotemporal decoupling disturbance analytical distortion, making it impossible for the motor control system to accurately identify boundary exceedance characteristics and spatiotemporal resonance distortion when facing shock-induced sudden loads and multi-scale frequency band energy drift. This, in turn, reduces the accuracy of disturbance feedforward compensation and the robustness of boundary protection reconstruction.

[0003] Therefore, how to provide a high-precision and robust control system and method for megawatt-level high-torque motors is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention proposes a high-precision robust control system and method for megawatt-level high-torque motors. It generates generalized disturbances through spatiotemporal state reconstruction based on decoupling errors and dynamic inductance parameter sets, and performs spatiotemporal decoupling disturbance analysis based on dynamic frequency band masking and impedance crossover reconstruction, outputting low-frequency and high-frequency disturbance estimates. Combining the monitored chip computing power margin ratio, it performs dimensionality reduction reconstruction and long-range reconnection compensation for characteristic attenuation. After boundary protection correction and frequency band fusion, it generates equivalent disturbance parameters and a reconstruction instruction set. The equivalent disturbance parameters are modulated to generate feedforward weights and superimposed with spatiotemporal resonance limiting compensation to output the desired current command. After manifold pre-evolution and orthogonal decoupling limiting compensation, the output voltage reference component is determined. Finally, after boundary-crossing edge closed-loop reconstruction, the carrier frequency is deduced to output the switching signal to the power device. This invention overcomes the limitations of traditional methods, such as ignoring cross-domain coupling phase shift, fixed-gain observation lag, and boundary-crossing distortion, providing accurate decision-making basis for high-precision control of megawatt-level high-torque motors.

[0005] A high-precision robust control method for a megawatt-class high-torque motor according to an embodiment of the present invention includes the following steps: Collect multi-source operating data to perform cross-domain reconstruction of electromechanical parameters to generate reference resistance and rotating coordinate system current, and output the current motor state quantity through disturbance decoupling state reconstruction. Based on the current motor state variables, a backstepping decoupling matrix is ​​constructed, and a cross-domain decoupling closed loop of electromechanical parameters is executed to output the decoupling error and dynamic inductance parameter set. The spatiotemporal state reconstruction of the decoupling error and dynamic inductance parameter set is performed to generate a generalized perturbation. Based on dynamic frequency band masking and impedance crossover reconstruction, spatiotemporal decoupling perturbation analysis is performed to output low-frequency perturbation estimates and high-frequency perturbation estimates. By monitoring the chip's computing power margin ratio and combining low-frequency disturbance estimates with high-frequency disturbance estimates, dimensionality reduction reconstruction and long-range reconnection compensation feature attenuation are performed. After boundary protection correction and fusion of frequency bands, equivalent disturbance parameters and reconstruction instruction sets are generated. Extract equivalent disturbance parameters, reconstruct instruction set, current motor state variables and decoupling error, calculate speed error and modulate disturbance to generate feedforward weight, combine reconstruct instruction set to output quadrature axis reference current, perform spatiotemporal resonance amplitude limiting compensation after smooth reconstruction and superimpose it on quadrature axis reference current, and output desired current command. Based on the desired current command, speed error and decoupling error, the system performs spatiotemporal coupling constraint scheduling, manifold pre-evolution and orthogonal decoupling limiting compensation, and outputs the voltage reference component. Extract the voltage reference component and reconstruct the instruction set, perform cross-domain coupled topology inverse mapping to generate the mapping component, dynamically optimize the output duty cycle parameter through the domain constraint, and deduce the carrier frequency based on the over-limit edge closed-loop reconstruction, and output the switching signal to the power device.

[0006] Optionally, the disturbance decoupling state reconstruction specifically includes: The stator current and winding temperature of a megawatt-level high-torque motor are collected. The armature heat loss power is calculated based on the stator current. The thermal coupling temperature rise coefficient is calculated based on the armature heat loss power and winding temperature. The thermal load following compensation is performed on the winding temperature according to the thermal coupling temperature rise coefficient to generate an equivalent thermal characteristic temperature. The nonlinear mapping calculation is performed based on the equivalent thermal characteristic temperature, and the reference resistance is output. Based on the stator current, the fundamental current component and harmonic current component are generated by signal analysis, and the corresponding first instantaneous phase and second instantaneous phase are calculated to construct the first frequency domain phase-sensitive transformation matrix and the second frequency domain phase-sensitive transformation matrix. The fundamental current component is generated by performing frequency-domain phase-sensitive spatial mapping on the fundamental current component using the first frequency-domain phase-sensitive transformation matrix. The harmonic current component is generated by performing frequency-domain phase-sensitive spatial mapping on the harmonic current component using the second frequency-domain phase-sensitive transformation matrix. The fundamental current and the harmonic current are superimposed to output the rotating coordinate system current. The speed of a megawatt-level high-torque motor is collected, and the dynamic amplitude boundary corresponding to each is calculated by combining the reference resistance and the current in the rotating coordinate system. Based on the dynamic amplitude boundary, the adaptive scaling factor is calculated, and the reference resistance, the current in the rotating coordinate system, and the speed are normalized using the adaptive scaling factor to generate normalized resistance components, normalized current components, and normalized speed components. Calculate the dynamic volatility of the normalized current component and deduce the adaptive over-limit threshold based on the dynamic volatility; if the normalized current component exceeds the adaptive over-limit threshold, calculate the over-limit warning compensation term and superimpose it on the normalized current component to generate the boundary correction current component. By splicing together the normalized resistance component, the boundary-corrected current component, and the normalized speed component, the current motor state of the megawatt-level high-torque motor is output.

[0007] Optionally, the cross-domain decoupling closed loop of electromechanical parameters specifically includes: Based on the current motor state variables, the rotating coordinate system current is extracted, the spatiotemporal distribution characteristics of the current vector are calculated, a cross-domain spatial mapping bridge is constructed based on the spatiotemporal distribution characteristics of the current vector to calculate the magnetic circuit nonlinear response distribution, and local flux linkage incremental differential calculation is performed based on the magnetic circuit nonlinear response distribution to output transient adaptive dynamic inductance. Based on the current motor state variables, the reference resistance is extracted. Combined with the transient adaptive dynamic inductance as the electromechanical parameter composite constraint, the time-varying cross-coupling degree distribution of electrical parameters is calculated. Based on the time-varying cross-coupling degree distribution of electrical parameters, an initial state feedback matrix is ​​constructed. Dynamic scheduling and diagonal dominant configuration of matrix elements are performed on the initial state feedback matrix, and the backstepping adaptive decoupling matrix is ​​output. The current motor state is input into the backstep adaptive decoupling matrix to perform channel independent orthogonal analysis, stripping global coupling features to separate disturbance residuals, extracting boundary extreme value features based on the disturbance residuals, performing state correction closed-loop calculation based on the boundary extreme value features to output decoupling error, and combining the boundary extreme value features with the transient adaptive dynamic inductor to perform parameter mapping and encapsulation, outputting a dynamic inductor parameter set.

[0008] Optionally, the spatiotemporal decoupling perturbation analysis specifically includes: The transient convergence gradient and statistical dispersion of the error are calculated based on the decoupling error. The dynamic feedback gain weight is calculated based on the transient convergence gradient and the statistical dispersion. The adaptive state smoothing correction calculation is performed based on the decoupling error and the dynamic feedback gain weight, and the smoothing error is output. Based on the smoothing error and the dynamic inductor parameter set, calculate the cross-space phase shift compensation coefficient, and perform phase amplitude alignment and nonlinear state differential tracking calculation based on the smoothing error, the dynamic inductor parameter set and the cross-space phase shift compensation coefficient to output the generalized disturbance; The multi-scale frequency band cross-correlation energy distribution is calculated based on the generalized perturbation. The frequency band boundary nonlinear masking suppression threshold is calculated based on the multi-scale frequency band cross-correlation energy distribution. Transient adaptive gain compensation and residual suppression reconstruction are performed based on the frequency band boundary nonlinear masking suppression threshold, and the time-frequency decoupled generalized perturbation is output. Based on the time-frequency decoupling generalized perturbation, a multi-scale receptive field spatial distribution is constructed, the center-periphery antagonistic response gradient at the high-low frequency crossing boundary is calculated, and edge-side suppression compensation and dynamic sharpening reconstruction are performed based on the center-periphery antagonistic response gradient to output the spatiotemporal orthogonal perturbation. Based on the spatiotemporal orthogonal perturbation, calculate the local time-varying spectral density distribution and characteristic dynamic impedance parameters. Based on the local time-varying spectral density distribution and characteristic dynamic impedance parameters, calculate the standing wave ratio and impedance mismatch at the high-low frequency crossing boundary. Based on the standing wave ratio and impedance mismatch, perform traveling wave reflection compensation and characteristic impedance matching reconstruction, and output continuous orthogonal perturbation at the boundary. Based on the continuous orthogonal perturbation of the boundary, calculate the frequency band energy drift characteristics and extreme boundary thresholds. Based on the frequency band energy drift characteristics and extreme boundary thresholds, perform adaptive frequency band crossing boundary delineation and region-independent analytical calculation, and output low-frequency perturbation estimates and high-frequency perturbation estimates.

[0009] Optionally, the boundary protection correction fusion frequency band specifically includes: Cross-band spatiotemporal feature extraction is performed based on low-frequency and high-frequency disturbance estimates. The frequency band energy distribution weight and spatiotemporal statistical significance grid are calculated. Adaptive mapping and dimensionality reduction reconstruction are performed based on the frequency band energy distribution weight and spatiotemporal statistical significance grid. High-order matrix operations are converted into scalar operators, and an adaptive dimensionality reduction scalar operator is output. Based on the adaptive dimensionality reduction scalar operator and the chip computing power margin ratio obtained by monitoring, the total intensity distribution of low-frequency disturbance estimates and high-frequency disturbance estimates is calculated. Based on the computing power margin ratio and the total intensity distribution, computing power composite constraint scheduling is performed, and the adaptive dimensionality reduction reconstruction matrix is ​​output. The node betweenness centrality and clustering coefficient distribution are calculated based on the adaptive dimensionality reduction and reconstruction matrix. The characteristic topological distance between local feature islands is calculated based on the node betweenness centrality and clustering coefficient distribution. Long-range edge reconnection is performed based on the characteristic topological distance, and the long-range reconnected connected topology is output. Based on the characteristic topological distance, the long-range reconnection topology is mapped to a node-level topology. Based on the node-level topology, the signal gain accumulation value of each branch node is calculated. Based on the signal gain accumulation value, the characteristic transmission damping coefficient is calculated. Based on the characteristic transmission damping coefficient, the characteristic attenuation caused by the drop in computing power is dynamically compensated, and the characteristic compensation weight matrix is ​​output. Based on the feature compensation weight matrix and the long-range reconnection topology calculation boundary connectivity fusion distribution, a topology protection zone is constructed at the computing power composite constraint boundary. The key mutation features of high-frequency disturbance estimation are calculated based on the boundary connectivity fusion distribution and the weights are transmitted without loss along the edge of the topology protection zone. Based on the edge lossless transmission weights, feature correction is performed on the adaptive dimensionality reduction reconstruction matrix, and the topology protection correction matrix is ​​output. Based on the topology protection correction matrix, frequency band fusion calculation is performed on the low-frequency disturbance estimate and the high-frequency disturbance estimate to output the equivalent disturbance parameters. Then, a reconstruction instruction set is compiled based on the equivalent disturbance parameters and the topology protection correction matrix.

[0010] In its implementation, this invention inherits the basic logic of perturbation feature extraction based on multi-scale frequency band decomposition in traditional signal processing, but it has been deeply modified to address the problems of dimensionality reduction drop and boundary distortion caused by limited computing power under complex dynamic conditions.

[0011] Traditional techniques typically only superimpose high- and low-frequency disturbance estimates as feedforward compensation quantities, failing to feed back the cognitive uncertainty caused by computing power fluctuations to the feature topology space for active correction. This results in key mutation features being severely attenuated when computing power drops and continuing to generate interference in subsequent modulation.

[0012] This invention creatively constructs a closed-loop correction system from "computational power composite constraint scheduling" to "topology protection feature reconstruction". Based on high and low frequency disturbance estimation, it calculates an adaptive dimensionality reduction scalar operator and executes computational power composite constraint scheduling to generate an adaptive dimensionality reduction reconstruction matrix. Then, based on the feature topology distance, it performs long-range edge reconnection to construct a long-range reconnected connected topology, realizing accurate mapping of global disturbance features to local topology levels. Subsequently, it calculates the feature transmission damping coefficient to dynamically compensate for feature attenuation and generate a feature compensation weight matrix. Based on this, it constructs a topology protection zone at the boundary of computational power composite constraint to calculate edge lossless transmission weights. Finally, it performs feature correction on the dimensionality reduction reconstruction matrix and outputs a topology protection correction matrix and reconstruction instruction set.

[0013] This modified system completely solves the problems of key disturbance feature attenuation and boundary distortion caused by the separation of computing power fluctuation and feature representation in traditional methods. It realizes the transformation of abstract computing power uncertainty into concrete topological lossless transmission protection force, effectively weakens feature attenuation and shock wave disturbance response surge in computing power drop areas, significantly improves the boundary continuity of disturbance estimation and the robustness of the system in computing power-constrained environments, and provides a highly reliable parameter decision basis for subsequent accurate amplitude limiting compensation and modulation reconstruction.

[0014] Optionally, the spatiotemporal resonance limiting compensation specifically includes: Extract the rotational speed and calculate the speed error from the current motor state variables, extract the rotating coordinate system current from the current motor state variables, calculate the spatiotemporal coupling constraint distribution based on the speed error and the rotating coordinate system current, and output the composite state constraint vector; Based on the composite state constraint vector, calculate the nonlinear dynamic activation threshold, construct a time-varying compensation gating matrix, perform amplitude nonlinear modulation on the equivalent disturbance parameter based on the time-varying compensation gating matrix, and output a time-varying feedforward compensation weight vector. Based on the reconstructed instruction set and the time-varying feedforward compensation weight vector, the adaptive scheduling coefficient of the control law is calculated. Based on the adaptive scheduling coefficient of the control law, a dynamic surface control law is constructed for the speed error. Based on the dynamic surface control law, the speed error is converted into a quadrature axis reference current. Based on the quadrature reference current and the time-varying feedforward compensation weight vector, the dynamic energy barrier distribution and compensation release rate coefficient are calculated. Based on the dynamic energy barrier distribution and compensation release rate coefficient, the state adaptive gain coefficient is calculated. Based on the state adaptive gain coefficient, the amplitude rate limiting and dynamic smoothing of the time-varying feedforward compensation weight vector are performed, and the dynamic smoothing compensation weight vector is output. Based on the dynamic smooth compensation weight vector and the cross-axis reference current, the boundary distortion resonant frequency band is extracted, the frequency band selective amplification factor and the phase continuous shaping operator are calculated, and the frequency band coupling filtering and phase reconstruction are performed on the dynamic smooth compensation weight vector based on the frequency band selective amplification factor and the phase continuous shaping operator to output the harmonic resonance continuous compensation parameters. Based on the decoupling error and the rotational speed error, a disturbance boundary trigger criterion is calculated. Based on the disturbance boundary trigger criterion, boundary adaptive limiting is performed on the harmonic resonance continuous compensation parameters. The limited harmonic resonance continuous compensation parameters are superimposed on the quadrature axis reference current to perform feedforward compensation and output the desired current command.

[0015] Optionally, the manifold pre-evolution and orthogonal decoupling limiting compensation specifically include: Based on the desired current command, the speed error, and the decoupling error, calculate the time-domain statistical characteristics of the speed error and the spatial boundary characteristics of the decoupling error. Based on the desired current command, the time-domain statistical characteristics of the speed error, and the spatial boundary characteristics of the decoupling error, perform cross-domain tensor product correlation mapping and calculate the corresponding coefficients of the spatiotemporal coupling constraint quantization. Based on the spatiotemporal coupling constraint quantization corresponding coefficients, multi-dimensional composite constraints and dynamic scheduling are executed, multi-constraint dynamic scheduling convergence rate coefficients are calculated, and adaptive sliding surface parameters and decoupling error space boundary characteristics are output. Based on the adaptive sliding surface parameters and the rotational speed error, a time-varying sliding surface manifold is constructed. Based on the decoupling error spatial boundary features and the adaptive sliding surface parameters, a running mode feature sequence is extracted. Based on the running mode feature sequence, scene-based feature stripping is performed, and local feature-sensitive observation gradients are calculated. Based on the local feature sensitive observation gradient, calculate the boundary limit adaptive early warning factor, and based on the local feature sensitive observation gradient and the boundary limit adaptive early warning factor, perform pre-intervention controlled second-order calculus evolution on the time-varying sliding surface manifold to output the super-twisted second-order sliding mode control law. Based on the super-twisted second-order sliding mode control law, the integral sliding surface switching control component and the equivalent robust control component are calculated. Based on the physical actuator constraints, the execution boundary space mapping basis matrix is ​​calculated. Based on the execution boundary space mapping basis matrix, the integral sliding surface switching control component and the equivalent robust control component are subjected to cross-space orthogonal decoupling mapping, and the orthogonal decoupling component is calculated. Based on the execution boundary space mapping basis matrix, calculate the over-limit distortion evaluation index for the orthogonal decoupling component, perform adaptive amplitude limiting closed-loop compensation based on the over-limit distortion evaluation index, output the inner loop correction term, and superimpose the inner loop correction term with the desired current command to output the voltage reference component.

[0016] Optionally, the over-limit edge closed-loop reconstruction specifically includes: Based on the reconstructed instruction set, spatiotemporal coupling constraint basis features are extracted; based on the voltage reference component, multidimensional weighted boundary features are extracted; based on the voltage reference component and the spatiotemporal coupling constraint basis features, cross-domain topological inverse mapping calculation is performed, and physical space mapping components are output. Based on the physical space mapping components, joint time-frequency analytical calculation is performed to output amplitude statistical distribution characteristics and phase gradient distribution characteristics. Based on the amplitude statistical distribution characteristics, target scope division is performed. Based on the target scope and the phase gradient distribution characteristics, scene configuration time-series weight extraction is performed to output initial scene time-series weights. Based on the multidimensional weight boundary characteristics, dynamic scheduling is performed on the initial scene time-series weights to output modulated wave controlled duty cycle parameters. Based on the controlled duty cycle parameter of the modulation wave, the narrow pulse over-limit depth value is calculated, the narrow pulse over-limit depth level factor is extracted based on the narrow pulse over-limit depth value, the carrier graded frequency deduction and pulse edge reconstruction quantization are performed based on the narrow pulse over-limit depth level factor and the controlled duty cycle parameter of the modulation wave, and a switching signal is output. The switching signal is output to the control terminal voltage of the power device.

[0017] A high-precision robust control system for a megawatt-level high-torque motor according to an embodiment of the present invention includes the following modules: The cross-domain reconfiguration module is used to collect multi-source operating data, perform cross-domain reconfiguration of electromechanical parameters to generate reference resistance and rotating coordinate system current, and output the current motor state quantity through disturbance decoupling state reconfiguration. The cross-domain decoupling module is used to construct a backstepping decoupling matrix based on the current motor state variables, execute a cross-domain decoupling closed loop for electromechanical parameters, and output the decoupling error and dynamic inductance parameter set. The perturbation analysis module is used to perform spatiotemporal state reconstruction on the decoupling error and dynamic inductance parameter set to generate generalized perturbations, and to perform spatiotemporal decoupling perturbation analysis based on dynamic frequency band masking and impedance crossover reconstruction, outputting low-frequency perturbation estimates and high-frequency perturbation estimates. The boundary protection module is used to monitor the chip's computing power margin ratio and combine low-frequency disturbance estimates with high-frequency disturbance estimates to perform dimensionality reduction reconstruction and long-range reconnection to compensate for feature attenuation. After boundary protection correction and fusion of frequency bands, equivalent disturbance parameters and reconstruction instruction sets are generated. The amplitude limiting compensation module is used to extract equivalent disturbance parameters, reconstruct the instruction set, current motor state variables and decoupling error, calculate speed error and modulate disturbance to generate feedforward weight, combine the reconstructed instruction set to output quadrature axis reference current, perform spatiotemporal resonance amplitude limiting compensation after smooth reconstruction and superimpose it on quadrature axis reference current, and output the desired current command. The manifold evolution module is used to perform spatiotemporal coupling constraint scheduling, manifold pre-evolution and orthogonal decoupling limiting compensation based on the desired current command, speed error and decoupling error, and output voltage reference component; The edge reconstruction module is used to extract the voltage reference component and reconstruction instruction set, perform cross-domain coupled topology inverse mapping to generate the mapping component, dynamically optimize the output duty cycle parameter through the domain constraint, and deduce the carrier frequency based on the over-limit edge closed-loop reconstruction, and output the switching signal to the power device.

[0018] This invention provides a high-precision robust control method for megawatt-level high-torque motors, involving motor drive and computer control coordination technology, which has the following beneficial effects: (1) This invention generates generalized disturbances by performing cross-domain decoupling of electromechanical parameters and spatiotemporal state reconstruction, and outputs high and low frequency disturbance estimates based on dynamic frequency band masking and impedance crossover reconstruction. It can accurately separate multi-scale crossover boundary features and cross-space phase shift components in shock wave sudden load, effectively eliminate disturbance estimation distortion caused by single time domain observation, solve the problem of disturbance analysis lag caused by neglecting cross-domain coupling phase shift in traditional control methods, and significantly improve the accuracy of disturbance feedforward compensation under complex working conditions.

[0019] (2) This invention combines the chip computing power margin ratio to perform dimensionality reduction reconstruction and long-distance reconnection to compensate for feature attenuation, and generates equivalent disturbance parameters and reconstruction instruction sets by boundary protection correction and fusion of frequency bands. It can dynamically schedule computing load according to computing power fluctuations and protect the lossless transmission of core control features, solving the problem of control performance degradation caused by feature attenuation when computing power is limited by traditional fixed algorithms, and greatly improving the robustness and stability of the control system under the condition of tight computing resources.

[0020] (3) By performing spatiotemporal coupling constraint scheduling, manifold pre-evolution and orthogonal decoupling amplitude limiting compensation to output voltage reference components, this invention can fully consider the physical boundary constraints and spatiotemporal resonance distortion of the actuator, solve the limitations of the traditional integral anti-integral saturation method that is prone to waveform distortion and control oscillation under critical over-limit conditions, and improve the smoothness and safety of control command output under boundary limit conditions.

[0021] (4) This invention generates mapping components by performing cross-domain coupled topology inverse mapping, and outputs the carrier frequency to the power device based on the over-limit edge closed-loop reconstruction. It can actively suppress narrow pulse over-limit and optimize the switching timing in the modulation output stage, overcome the defect of classical carrier modulation method that is difficult to adapt to extreme boundary changes, and provide accurate decision basis for the efficient and reliable operation of megawatt-level high torque motor system. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a high-precision robust control method for a megawatt-level high-torque motor proposed in this invention; Figure 2 This is a flowchart illustrating the working principle of the boundary protection correction fusion frequency band of the high-precision robust control method for megawatt-level high-torque motor proposed in this invention. Figure 3 This is a schematic diagram of the structure of a high-precision robust control system for a megawatt-level high-torque motor proposed in this invention. Detailed Implementation

[0023] The invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0024] refer to Figure 1 and Figure 2 A high-precision robust control method for megawatt-class high-torque motors includes the following steps: Collect multi-source operating data to perform cross-domain reconstruction of electromechanical parameters to generate reference resistance and rotating coordinate system current, and output the current motor state quantity through disturbance decoupling state reconstruction. Based on the current motor state variables, a backstepping decoupling matrix is ​​constructed, and a cross-domain decoupling closed loop of electromechanical parameters is executed to output the decoupling error and dynamic inductance parameter set. The spatiotemporal state reconstruction of the decoupling error and dynamic inductance parameter set is performed to generate a generalized perturbation. Based on dynamic frequency band masking and impedance crossover reconstruction, spatiotemporal decoupling perturbation analysis is performed to output low-frequency perturbation estimates and high-frequency perturbation estimates. By monitoring the chip's computing power margin ratio and combining low-frequency disturbance estimates with high-frequency disturbance estimates, dimensionality reduction reconstruction and long-range reconnection compensation feature attenuation are performed. After boundary protection correction and fusion of frequency bands, equivalent disturbance parameters and reconstruction instruction sets are generated. Extract equivalent disturbance parameters, reconstruct instruction set, current motor state variables and decoupling error, calculate speed error and modulate disturbance to generate feedforward weight, combine reconstruct instruction set to output quadrature axis reference current, perform spatiotemporal resonance amplitude limiting compensation after smooth reconstruction and superimpose it on quadrature axis reference current, and output desired current command. Based on the desired current command, speed error and decoupling error, the system performs spatiotemporal coupling constraint scheduling, manifold pre-evolution and orthogonal decoupling limiting compensation, and outputs the voltage reference component. Extract the voltage reference component and reconstruct the instruction set, perform cross-domain coupled topology inverse mapping to generate the mapping component, dynamically optimize the output duty cycle parameter through the domain constraint, and deduce the carrier frequency based on the over-limit edge closed-loop reconstruction, and output the switching signal to the power device.

[0025] In this embodiment, the disturbance decoupling state reconstruction specifically includes: Read the motor stator current and winding temperature, calculate the armature heat loss power by multiplying the square of the stator current by the preset equivalent resistance of 0.05 ohms; divide the armature heat loss power by the preset heat dissipation coefficient of 0.2 watts per degree Celsius and add the winding temperature to obtain the thermal coupling temperature rise coefficient; multiply the thermal coupling temperature rise coefficient by the preset following compensation coefficient of 0.05 and add the winding temperature to obtain the equivalent thermal characteristic temperature, input the Sigmoid activation function to output the reference resistance at the reference temperature of 25 degrees Celsius.

[0026] Perform a Fourier transform on the stator current to extract the fundamental and harmonic current components and calculate the instantaneous phase; construct a two-row, two-column phase-sensitive transformation matrix with elements of 1 in the upper left and lower left corners, 1 in the upper right corner, and -1 in the lower right corner; substitute the phase into the matrix to perform a mapping calculation, generate the fundamental and harmonic rotating coordinate system currents and add them together, and output the rotating coordinate system current.

[0027] Read the motor speed, calculate the range of the reference resistance, rotating coordinate system current and speed within the sliding window as the dynamic amplitude boundary; divide the upper limit of the preset normalization interval by 1 to obtain the adaptive scaling factor, subtract the minimum value of each from the original value and multiply by the factor, and output the normalized resistance, current and speed components.

[0028] The normalized current component's sliding variance within a preset 1 second is calculated as the dynamic volatility. This variance is multiplied by a preset extrapolation coefficient of 3 and a preset margin of 0.1 to obtain the adaptive over-limit threshold. If the current component exceeds this threshold, the difference is calculated and multiplied by a preset warning gain of 0.5 as a compensation term, which is then added to the current component to output the boundary-corrected current component.

[0029] The normalized resistance, boundary correction current, and normalized speed components are concatenated end to end to construct a three-dimensional state vector, and the current motor state is output.

[0030] In this embodiment, the cross-domain decoupling closed loop of electromechanical parameters specifically includes: Extract the rotating coordinate system current from the current motor state variables, calculate its sliding variance within a preset space window of 0.1 seconds as the spatiotemporal distribution characteristics of the current vector; substitute it into a two-row, two-column mapping bridge matrix with elements of 0.5 at the top left and bottom left corners, 0.5 at the top right corner, and -0.5 at the bottom right corner to perform multiplication calculation, and input the Tanh activation function to generate the magnetic circuit nonlinear response distribution; calculate the adjacent differences of this distribution within a preset step size of 0.01 seconds, multiply it by a preset magnetic flux reference coefficient of 0.02 Weber, and output the transient adaptive dynamic inductance.

[0031] Extract the reference resistance from the current motor state variables and concatenate it with the transient adaptive dynamic inductance to construct a composite constraint quantity of electromechanical parameters; calculate the product of the reference resistance and the transient adaptive dynamic inductance and divide it by the sum of their squares to obtain the time-varying cross-coupling degree distribution of the electrical parameters; construct a two-row, two-column initial state feedback matrix, setting the upper left corner as the reference resistance, the lower right corner as the transient adaptive dynamic inductance, the upper right corner as the coupling degree distribution, and the lower left corner as the negative number of the coupling degree distribution; multiply the lower left corner element by a preset diagonal dominance coefficient of 0.1, so that the difference in the absolute values ​​of the main diagonals is greater than a preset threshold of 0.1, and output the backstepping adaptive decoupling matrix.

[0032] The current motor state input is input to the backstepping adaptive decoupling matrix and multiplied to obtain the decoupling channel output component to remove global coupling features and separate disturbance residuals. The maximum absolute value of the disturbance residual sequence is extracted as the boundary extreme value feature. The difference between the value and the preset residual reference value of 0.01 is calculated and multiplied by the preset correction gain coefficient of 5.0, and then added to the disturbance residual sequence to output the decoupling error. The boundary extreme value feature and the transient adaptive dynamic inductor are concatenated to construct a two-dimensional parameter vector and output the dynamic inductor parameter set.

[0033] In this embodiment, the spatiotemporal decoupling perturbation analysis specifically includes: Read the decoupling error sequence, calculate the absolute value of the difference between adjacent values ​​within a 0.1-second window as the transient convergence gradient of the error, and calculate the standard deviation of the values ​​within the window as the statistical dispersion; multiply the gradient by a preset gain of 0.3 and add the dispersion multiplied by a preset gain of 0.7 to obtain the dynamic feedback gain weight; multiply the decoupling error by this weight, add the smoothing error from the previous time step multiplied by a preset smoothing factor of 0.8, and accumulate to output the smoothing error.

[0034] Extract the smoothing error and the dynamic inductance parameter set, calculate the smoothing error divided by the arctangent of the dynamic inductance parameter set as the cross-space phase shift compensation coefficient; multiply the smoothing error by this coefficient to perform phase alignment, and superimpose the dynamic inductance parameter set to perform amplitude alignment; input the alignment state value into the Tanh activation function, calculate the rate of change of the value within a 0.01 second step and multiply it by a preset coefficient of 0.5, and output the generalized perturbation.

[0035] The cross-correlation coefficients of the generalized perturbation in the 1-2 kHz and 0-1 kHz frequency bands are calculated as the multi-scale frequency band cross-correlation energy distribution. The cross-correlation coefficients are input into the Sigmoid activation function, and the output value is multiplied by a preset benchmark of 0.5 to obtain the frequency band boundary nonlinear masking suppression threshold. The threshold is subtracted from the generalized perturbation, and the difference is multiplied by a preset gain of 1.2 and superimposed on the generalized perturbation to output the time-frequency decoupled generalized perturbation.

[0036] The time-frequency decoupled generalized perturbation is divided into 1 central data point and 8 peripheral data points to construct a multi-scale receptive field spatial distribution; the mean value of the center minus the mean value of the periphery is calculated to obtain the center-periphery antagonistic response gradient at the high-low frequency crossing boundary; it is multiplied by a preset coefficient of 0.6 to perform edge-side suppression compensation, and the time-frequency decoupled generalized perturbation is multiplied by a preset coefficient of 1.5 to perform dynamic sharpening reconstruction, outputting the spatiotemporal orthogonal perturbation.

[0037] A fast Fourier transform is performed on the spatiotemporal orthogonal disturbance to calculate the local time-varying spectral density distribution. The mean of the distribution is divided by the density to obtain the characteristic dynamic impedance parameters. The values ​​of 1 to 2 kHz and 0 to 1 kHz in the impedance parameters are extracted. The absolute value of the high frequency divided by the low frequency is used as the standing wave ratio (SWR), and the absolute value of the difference between the high and low frequencies is divided by the sum to obtain the impedance mismatch. The SWR is multiplied by a preset coefficient of 0.4 to perform traveling wave reflection compensation, and the mismatch is multiplied by a preset coefficient of 0.2 to perform impedance matching reconstruction. The two results are superimposed on the spatiotemporal orthogonal disturbance to output a continuous orthogonal disturbance at the boundary.

[0038] The mean change rate of continuous orthogonal disturbances within a 1-second window is calculated as the frequency band energy drift feature. The extreme value boundary threshold is obtained by extracting the sequence range and multiplying it by a preset coefficient of 0.9. The sum of the two is used as the boundary demarcation point of the frequency band crossing. It is determined whether the frequency of each point is less than the demarcation point. If it is less, it is classified into the low-frequency channel; otherwise, it is classified into the high-frequency channel. The mean values ​​of the data points of the two channels are extracted respectively, and the low-frequency and high-frequency disturbance estimates are output.

[0039] In its implementation, this invention inherits the basic logic of extracting external disturbances based on state error in traditional disturbance observation technology, but it has been deeply modified to address the problem of disturbance analysis distortion caused by multi-band signal crossover distortion and traveling wave reflection interference in complex dynamic conditions of megawatt-level high-torque motors.

[0040] Traditional techniques typically use a single-band filter to simply extract state errors, which cannot handle impedance mismatch and residual coupling at the high-low frequency crossing boundary. This causes the disturbance estimate to be distorted at the boundary and continuously induces current oscillations in subsequent control.

[0041] This invention creatively constructs a cross-band perturbation analysis system from "spatiotemporal state reconstruction" to "impedance matching reconstruction." Based on decoupling error and dynamic inductance parameter set, it calculates cross-spatial phase shift compensation coefficients and performs nonlinear differential tracking to generate a generalized perturbation containing full-band characteristics. Subsequently, based on the generalized perturbation, it calculates multi-scale frequency band cross-correlation energy distribution, generates a nonlinear masking suppression threshold, performs residual suppression reconstruction, and outputs a time-frequency decoupled generalized perturbation. Next, it constructs a multi-scale receptive field to calculate the center-periphery antagonistic response gradient at the high-low frequency crossover boundary, performs edge-side suppression compensation, and outputs a spatiotemporal orthogonal perturbation. Then, it calculates the local time-varying spectral density and characteristic dynamic impedance parameters to obtain the standing wave ratio and impedance mismatch at the high-low frequency crossover boundary, performs traveling wave reflection compensation and characteristic impedance matching reconstruction, and outputs a boundary continuous orthogonal perturbation. Finally, based on the boundary continuous orthogonal perturbation, it calculates the frequency band energy drift characteristics and extreme value boundary threshold, performs adaptive frequency band crossover boundary delineation and regional independent analysis, and outputs low-frequency perturbation estimates and high-frequency perturbation estimates.

[0042] This modified system completely solves the problems of disturbance observation lag and phase distortion caused by crossover boundary distortion and traveling wave reflection in traditional methods. It realizes the transformation of multi-scale frequency band residuals into orthogonal disturbance resolution with continuous and lossless boundary, effectively weakens the response surge and reflection distortion at high and low frequency crossover, and significantly improves the accuracy and spatiotemporal continuity of disturbance estimation under complex working conditions. It provides a high-fidelity disturbance feedforward basis for subsequent accurate amplitude limiting compensation and robust control.

[0043] In this embodiment, the boundary protection correction fusion frequency band specifically includes: Read the low-frequency and high-frequency disturbance estimates, divide the square of the low-frequency estimate by the sum of the squares of the two to obtain the frequency band energy distribution weight; divide the absolute value of the difference between high and low frequencies by the sum of the standard deviation of the high-frequency estimate and 0.1 to obtain the spatiotemporal statistical significance grid; construct a two-row, two-column adaptive mapping matrix, set the top left and bottom right as weights, the top right as the grid, and the bottom left as the grid inverse, calculate the absolute value of its determinant and input it into the Sigmoid activation function, multiply by 0.5 to output the adaptive dimensionality reduction scalar operator.

[0044] Read the ratio of the operator to the chip's computing power margin, multiply the sum of the absolute values ​​of the high and low frequency estimates by the operator to obtain the total intensity distribution; divide the computing power margin ratio by the sum of the total intensity distribution and 0.01 to obtain the computing power composite constraint scheduling factor; construct the basic matrix, set the upper left corner to the factor, the lower right corner to 1.0, and the rest to 0, multiply each element by the total intensity distribution, and output the adaptive dimensionality reduction reconstruction matrix.

[0045] The matrix elements are used as nodes. The sum of the product of the nodes' connections is used as the betweenness centrality. The ratio of the actual number of connections to the number of possible connections is used as the clustering coefficient. The betweenness centrality is divided by the clustering coefficient to obtain the feature topological distance. The threshold is set to 0.5. The weight of the connection with a topological distance greater than 0.5 is set to 1.0, and otherwise it is set to 0.1. The long-range reconnection connected topology is output.

[0046] Extract the feature topological distance and map nodes with connection weights greater than 0.5 to a hierarchical topology; calculate the connection weights of branch nodes and their child nodes as the cumulative signal gain value; subtract the product of this cumulative value and 0.3 from 1.0 to obtain the feature transmission damping coefficient; multiply the damping coefficient by 2.0 and add it to the original values ​​of the corresponding nodes in the topology to output the feature compensation weight matrix.

[0047] Multiply the feature compensation weight matrix with the corresponding elements of the long-range reconnected connected topology to obtain the boundary connectivity fusion distribution; extract the boundary regions with computing power factors less than 0.3 and set the topology protection band width to 0.2; divide the mean value within the protection band by the mean value outside the band to obtain the edge lossless transmission weight; multiply the weight by each element of the adaptive dimensionality reduction reconstruction matrix to output the topology protection correction matrix.

[0048] The low-frequency and high-frequency disturbance estimates are added together and multiplied by the determinant of the topology protection correction matrix to output the equivalent disturbance parameters. The equivalent disturbance parameters and each element of the topology protection correction matrix are concatenated byte by byte to compile and generate a reconstruction instruction set.

[0049] In its implementation, this invention inherits the basic logic of perturbation feature extraction based on multi-scale frequency band decomposition in traditional signal processing, but it has been deeply modified to address the problems of dimensionality reduction drop and boundary distortion caused by limited computing power under complex dynamic conditions.

[0050] Traditional techniques typically only superimpose high- and low-frequency disturbance estimates as feedforward compensation quantities, failing to feed back the cognitive uncertainty caused by computing power fluctuations to the feature topology space for active correction. This results in key mutation features being severely attenuated when computing power drops and continuing to generate interference in subsequent modulation.

[0051] This invention creatively constructs a closed-loop correction system from "computational power composite constraint scheduling" to "topology protection feature reconstruction". Based on high and low frequency disturbance estimation, it calculates an adaptive dimensionality reduction scalar operator and executes computational power composite constraint scheduling to generate an adaptive dimensionality reduction reconstruction matrix. Then, based on the feature topology distance, it performs long-range edge reconnection to construct a long-range reconnected connected topology, realizing accurate mapping of global disturbance features to local topology levels. Subsequently, it calculates the feature transmission damping coefficient to dynamically compensate for feature attenuation and generate a feature compensation weight matrix. Based on this, it constructs a topology protection zone at the boundary of computational power composite constraint to calculate edge lossless transmission weights. Finally, it performs feature correction on the dimensionality reduction reconstruction matrix and outputs a topology protection correction matrix and reconstruction instruction set.

[0052] This modified system completely solves the problems of key disturbance feature attenuation and boundary distortion caused by the separation of computing power fluctuation and feature representation in traditional methods. It realizes the transformation of abstract computing power uncertainty into concrete topological lossless transmission protection force, effectively weakens feature attenuation and shock wave disturbance response surge in computing power drop areas, significantly improves the boundary continuity of disturbance estimation and the robustness of the system in computing power-constrained environments, and provides a highly reliable parameter decision basis for subsequent accurate amplitude limiting compensation and modulation reconstruction.

[0053] In this embodiment, spatiotemporal resonance limiting compensation specifically includes: Extract the current motor state variables by subtracting the preset reference speed of 1000 rpm from the current speed to obtain the speed error, and extract the rotating coordinate system current; multiply the absolute value of the speed error by the preset coupling coefficient of 0.5, divide by the sum of the absolute value of the rotating coordinate system current and 0.1 amperes to obtain the spatiotemporal coupling constraint distribution; multiply the distribution by the preset constraint gain of 2.0 and input it into the Tanh activation function, concatenate the output value with the rotating coordinate system current, and output the composite state constraint vector.

[0054] Extract the composite state constraint vector, calculate the absolute value of the product of its first and second dimensions, and add 0.2 to obtain the nonlinear dynamic activation threshold; construct a two-row, two-column time-varying compensation gating matrix, setting the top left corner to the threshold, the bottom right corner to its opposite, and the top right and bottom left corners to 0.5; extract the equivalent perturbation parameters, multiply them with the corresponding elements of the matrix, and sum them, inputting the result into the Sigmoid activation function, and outputting the time-varying feedforward compensation weight vector.

[0055] Divide the reconstructed instruction set byte length by 1000 and multiply by 1.5 to obtain the scheduling base value. Add the absolute value of the time-varying feedforward compensation weight vector multiplied by 0.8 to obtain the control law adaptive scheduling coefficient. Multiply the speed error by this coefficient and subtract 5.0, input it into the Tanh activation function, multiply the output value by 10 Amperes and add the reference current of 50 Amperes to output the quadrature axis reference current.

[0056] The dynamic energy barrier distribution is obtained by multiplying the square of the quadrature-axis reference current by 0.05 and adding the absolute value of the time-varying feedforward compensation weight vector by 2.0. Dividing by 0.02 seconds yields the compensation release rate coefficient, which is then multiplied by 0.5 and input into the Sigmoid activation function to obtain the state adaptive gain coefficient. The time-varying feedforward compensation weight vector is multiplied by this gain coefficient, and if the absolute value exceeds 20 amperes, it is truncated to the 20 ampere sign. The value from the previous moment is then added and multiplied by 0.9 to output the dynamic smooth compensation weight vector.

[0057] Divide the quadrature-axis reference current by the dynamic smoothing compensation weight vector, and extract the frequency bands with a ratio greater than 2.0 as the boundary distortion resonant frequency bands; divide the center frequency by 50 Hz to obtain the band selective amplification factor, multiply the band phase difference by 0.01 radians and accumulate it, and input it into the Tanh activation function to obtain the phase continuous shaping operator; multiply the dynamic smoothing compensation weight vector by the amplification factor and add the shaping operator, perform band filtering and phase reconstruction, and output the harmonic resonance continuous compensation parameters.

[0058] The absolute value of the decoupling error is calculated and multiplied by 0.6, and the absolute value of the rotational speed error is multiplied by 0.4 to obtain the disturbance boundary trigger criterion. The amplitude limiting threshold is set to 1.5 Amperes. If the criterion is greater than the threshold, the harmonic resonance continuous compensation parameter is truncated to 1.5 Amperes and multiplied by 0.8. Otherwise, the original value is retained. The processed parameter is superimposed on the quadrature axis reference current to perform feedforward compensation and output the desired current command.

[0059] In this embodiment, the manifold pre-evolution and orthogonal decoupling limiting compensation specifically include: The desired current command, speed error, and decoupling error are extracted. The standard deviation of the speed error within 0.1 seconds is calculated as the time-domain statistical feature, and the range of the decoupling error within 0.1 seconds is calculated as the spatial boundary feature. The product of the absolute values ​​of the three is input into the Sigmoid activation function, and the output value is multiplied by 0.5 to obtain the corresponding coefficient of the spatiotemporal coupling constraint quantization.

[0060] Add 0.2 to this coefficient to obtain the convergence rate coefficient of the multi-constraint dynamic scheduling, multiply it by 1.5 to obtain the adaptive sliding surface parameters; multiply this coefficient by 2.0 and superimpose it onto the spatial boundary features to output the updated decoupling error spatial boundary features.

[0061] Multiply the rotational speed error by the adaptive sliding surface parameters to obtain the time-varying sliding surface manifold; take the absolute value of the difference between the updated decoupling error spatial boundary features and the time-varying sliding surface manifold values, input it into the Tanh activation function and multiply by 0.8 to obtain the local feature sensitive observation gradient.

[0062] Divide the gradient by 1.1 to remove the zero constant to obtain the boundary over-limit adaptive warning factor, multiply it by 0.01 seconds as the controlled step size; calculate the first and second derivatives of the time-varying sliding surface manifold within the step size, superimpose the gradient multiplied by 0.5 as a pre-intervention term, and output the super-twisted second-order sliding mode control law.

[0063] The summation within the step size is used as the integral sliding surface switching control component, and the remainder is used as the equivalent robust control component. An identity matrix is ​​constructed based on the maximum voltage of 400 volts divided by the reference voltage of 10 volts as the execution boundary space mapping basis matrix. The two control components are mapped and subtracted through this matrix to output the orthogonal decoupling component.

[0064] Divide its absolute value by matrix trace 2.0 to obtain the over-limit distortion evaluation index; determine whether the index is greater than 0.5. If it is, multiply the decoupling component by 0.6; otherwise, multiply by 1.0 as the inner loop correction term; add it to the desired current command to obtain the output voltage reference component.

[0065] In this embodiment, the over-limit edge closed-loop reconstruction specifically includes: Extract the byte length of the reconstructed instruction set, divide it by 1000, and multiply it by 0.5 to obtain the spatiotemporal coupling constraint basis features; subtract 10 volts from the absolute value of the voltage reference component to obtain the multidimensional weighted boundary features; construct a matrix with the basis features at the top left, the boundary features at the bottom right, 0.2 at the top right, and 0 at the bottom left, and perform matrix multiplication mapping with the voltage reference component and the 0.1 volt bias to output the physical space mapping component.

[0066] The range of this component within 0.1 seconds is calculated as the amplitude statistical distribution feature. The adjacent differences within 0.01 seconds are input into the Tanh activation function to obtain the phase gradient distribution feature. The amplitude feature is divided by 5 volts and rounded up to divide the target domain. The domain number is multiplied by 0.2 seconds and the absolute value of the phase gradient is added. This is input into the Sigmoid activation function to output the initial scene temporal weight. This weight is multiplied by 1.5 by the multidimensional weight boundary feature to output the controlled duty cycle parameter of the modulated wave.

[0067] The absolute value of the difference between this parameter and 0.5 is multiplied by 10.0 to obtain the narrow pulse over-limit depth value; it is mapped to a level factor of 0.3, 0.6 or 1.0 based on the thresholds of 0.2, 0.5 and 0.8; this factor is added to 2.0 and multiplied by 1000 Hz to deduce the carrier frequency, and its period is multiplied by the level factor to obtain the pulse edge quantization value, which is then superimposed on the carrier edge time to reconstruct and generate a switching signal output to the control terminal voltage of the power device.

[0068] refer to Figure 3 A high-precision, robust control system for megawatt-level high-torque motors, specifically comprising the following modules: The cross-domain reconfiguration module is used to collect multi-source operating data, perform cross-domain reconfiguration of electromechanical parameters to generate reference resistance and rotating coordinate system current, and output the current motor state quantity through disturbance decoupling state reconfiguration. The cross-domain decoupling module is used to construct a backstepping decoupling matrix based on the current motor state variables, execute a cross-domain decoupling closed loop for electromechanical parameters, and output the decoupling error and dynamic inductance parameter set. The perturbation analysis module is used to perform spatiotemporal state reconstruction on the decoupling error and dynamic inductance parameter set to generate generalized perturbations, and to perform spatiotemporal decoupling perturbation analysis based on dynamic frequency band masking and impedance crossover reconstruction, outputting low-frequency perturbation estimates and high-frequency perturbation estimates. The boundary protection module is used to monitor the chip's computing power margin ratio and combine low-frequency disturbance estimates with high-frequency disturbance estimates to perform dimensionality reduction reconstruction and long-range reconnection to compensate for feature attenuation. After boundary protection correction and fusion of frequency bands, equivalent disturbance parameters and reconstruction instruction sets are generated. The amplitude limiting compensation module is used to extract equivalent disturbance parameters, reconstruct the instruction set, current motor state variables and decoupling error, calculate speed error and modulate disturbance to generate feedforward weight, combine the reconstructed instruction set to output quadrature axis reference current, perform spatiotemporal resonance amplitude limiting compensation after smooth reconstruction and superimpose it on quadrature axis reference current, and output the desired current command. The manifold evolution module is used to perform spatiotemporal coupling constraint scheduling, manifold pre-evolution and orthogonal decoupling limiting compensation based on the desired current command, speed error and decoupling error, and output voltage reference component; The edge reconstruction module is used to extract the voltage reference component and reconstruction instruction set, perform cross-domain coupled topology inverse mapping to generate the mapping component, dynamically optimize the output duty cycle parameter through the domain constraint, and deduce the carrier frequency based on the over-limit edge closed-loop reconstruction, and output the switching signal to the power device.

[0069] Example 1: To verify the feasibility of this invention in the decoupling and control of high dynamic disturbances in megawatt-level high-torque motors, the method of this invention was applied to a megawatt-level high-torque motor drive system of a large steel group (hereinafter referred to as "System M"). In traditional motor control systems, proportional-integral control or conventional vector control methods are usually used for speed regulation and current tracking. These methods not only have limited ability to suppress nonlinear disturbances and time-varying parameters, but also cannot effectively handle high-frequency resonance and boundary distortion caused by multi-physics coupling, which easily leads to a decrease in control accuracy and damage to power devices. To solve the above problems, System M decided to adopt the high-precision robust control system and method for megawatt-level high-torque motors proposed in this invention.

[0070] During implementation, system M first uses its data acquisition system to acquire real-time operating data such as motor stator current, winding temperature, and motor speed. After signal filtering, coordinate transformation, and feature alignment processing, a high-quality electromechanical state data stream is formed. At the same time, system M accurately expresses the multi-dimensional coupling characteristics of motor thermodynamics and electromagnetism by integrating static physical parameters such as armature heat loss, heat dissipation coefficient, and equivalent resistance.

[0071] System M employs a disturbance decoupling state reconstruction technique to adaptively scale and normalize stator current, speed, and reference resistance. Utilizing a phase-sensitive transformation matrix and a Sigmoid activation function, it efficiently expresses the multivariable spatiotemporal constraint distribution and boundary correction characteristics of the motor. Furthermore, through cross-domain decoupling closed-loop analysis of electromechanical parameters and spatiotemporal decoupling disturbance analysis, it integrates frequency band cross-correlation energy and spatiotemporal orthogonal disturbance characteristics. Using a Tanh activation function and an adaptive masking suppression threshold, it fully captures the evolution of transient disturbances and low-to-high frequency cross-boundaries during motor operation. Subsequently, the learned equivalent disturbance parameters and spatiotemporal orthogonal compensation characteristics are used for time-varying sliding mode surface manifold construction and orthogonal decoupling amplitude limiting compensation, effectively improving the robustness and dynamic response accuracy of motor control.

[0072] During implementation, the technical team of System M discovered that, compared with traditional proportional-integral and vector control methods, the method of this invention significantly improves the motor's disturbance rejection capability and speed control accuracy. Traditional methods cannot accurately characterize the spatiotemporal evolution of multi-band disturbances in the motor and the boundary distortion of the physical actuator, while the method of this invention effectively achieves deep decoupling of broadband disturbances and accurate output of voltage reference components through manifold pre-evolution and limit-crossing edge closed-loop reconstruction.

[0073] To further verify the actual performance of the method of the present invention, system M conducted a detailed comparative test between the method of the present invention and the traditional method. The specific performance data is shown in Table 1: Table 1 Performance Comparison of Control Methods for M-MW High Torque Motors in the System

[0074] As shown in Table 1, the performance of the megawatt-level high-torque motor system was comprehensively improved after applying the method of this invention. Speed ​​control accuracy increased from 1.5% in the traditional method to 0.3%, and disturbance suppression recovery time decreased from 85 milliseconds to 22 milliseconds, significantly improving the system's dynamic performance and directly reducing the risk of oscillations caused by disturbances. The total harmonic distortion (THD) of the current and the transient torque tracking error also achieved significant reductions of over 70%. The calculation time for a single control law decreased from 150 microseconds in the traditional method to 55 microseconds, and the real-time state update capability increased from 5000Hz to 15000Hz, resulting in a significant improvement in control efficiency and response speed. Furthermore, the temperature rise of power devices decreased from 35.5℃ to 18.2℃, significantly improving motor operating efficiency and effectively improving operating conditions and hardware lifespan. The overall system stability margin also significantly improved, increasing from 0.42 to 0.78.

[0075] Through the method of this invention, system M successfully achieved precise anti-disturbance control of megawatt-level high-torque motors, effectively improving motor operating efficiency and dynamic tracking performance, ensuring the high efficiency and safety of industrial production, significantly improving the intelligence and automation level of motor control, significantly reducing energy loss caused by high-frequency harmonics, enhancing the real-time robustness of the system, and providing strong technical support for intelligent drive of high-end equipment.

[0076] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A high-precision robust control method for a megawatt-class high-torque motor, characterized in that, Includes the following steps: Collect multi-source operating data to perform cross-domain reconstruction of electromechanical parameters to generate reference resistance and rotating coordinate system current, and output the current motor state quantity through disturbance decoupling state reconstruction. Based on the current motor state variables, a backstepping decoupling matrix is ​​constructed, and a cross-domain decoupling closed loop of electromechanical parameters is executed to output the decoupling error and dynamic inductance parameter set. The spatiotemporal state reconstruction of the decoupling error and dynamic inductance parameter set is performed to generate a generalized perturbation. Based on dynamic frequency band masking and impedance crossover reconstruction, spatiotemporal decoupling perturbation analysis is performed to output low-frequency perturbation estimates and high-frequency perturbation estimates. By monitoring the chip's computing power margin ratio and combining low-frequency disturbance estimates with high-frequency disturbance estimates, dimensionality reduction reconstruction and long-range reconnection compensation feature attenuation are performed. After boundary protection correction and fusion of frequency bands, equivalent disturbance parameters and reconstruction instruction sets are generated. Extract equivalent disturbance parameters, reconstruct instruction set, current motor state variables and decoupling error, calculate speed error and modulate disturbance to generate feedforward weight, combine reconstruct instruction set to output quadrature axis reference current, perform spatiotemporal resonance amplitude limiting compensation after smooth reconstruction and superimpose it on quadrature axis reference current, and output desired current command. Based on the desired current command, speed error and decoupling error, the system performs spatiotemporal coupling constraint scheduling, manifold pre-evolution and orthogonal decoupling limiting compensation, and outputs the voltage reference component. Extract the voltage reference component and reconstruct the instruction set, perform cross-domain coupled topology inverse mapping to generate the mapping component, dynamically optimize the output duty cycle parameter through the domain constraint, and deduce the carrier frequency based on the over-limit edge closed-loop reconstruction, and output the switching signal to the power device.

2. The high-precision robust control method for a megawatt-level high-torque motor according to claim 1, characterized in that, The boundary protection correction fusion frequency band specifically includes: Cross-band spatiotemporal feature extraction is performed based on low-frequency and high-frequency disturbance estimates. The frequency band energy distribution weight and spatiotemporal statistical significance grid are calculated. Adaptive mapping and dimensionality reduction reconstruction are performed based on the frequency band energy distribution weight and spatiotemporal statistical significance grid. High-order matrix operations are converted into scalar operators, and an adaptive dimensionality reduction scalar operator is output. Based on the adaptive dimensionality reduction scalar operator and the chip computing power margin ratio obtained by monitoring, the total intensity distribution of low-frequency disturbance estimates and high-frequency disturbance estimates is calculated. Based on the computing power margin ratio and the total intensity distribution, computing power composite constraint scheduling is performed, and the adaptive dimensionality reduction reconstruction matrix is ​​output. The node betweenness centrality and clustering coefficient distribution are calculated based on the adaptive dimensionality reduction and reconstruction matrix. The characteristic topological distance between local feature islands is calculated based on the node betweenness centrality and clustering coefficient distribution. Long-range edge reconnection is performed based on the characteristic topological distance, and the long-range reconnected connected topology is output. Based on the characteristic topological distance, the long-range reconnection topology is mapped to a node-level topology. Based on the node-level topology, the signal gain accumulation value of each branch node is calculated. Based on the signal gain accumulation value, the characteristic transmission damping coefficient is calculated. Based on the characteristic transmission damping coefficient, the characteristic attenuation caused by the drop in computing power is dynamically compensated, and the characteristic compensation weight matrix is ​​output. Based on the feature compensation weight matrix and the long-range reconnection topology calculation boundary connectivity fusion distribution, a topology protection zone is constructed at the computing power composite constraint boundary. The key mutation features of high-frequency disturbance estimation are calculated based on the boundary connectivity fusion distribution and the weights are transmitted without loss along the edge of the topology protection zone. Feature correction is performed on the adaptive dimensionality reduction reconstruction matrix based on the edge lossless transmission weight, and a topology protection correction matrix is ​​output. Based on the topology protection correction matrix, frequency band fusion calculation is performed on the low-frequency disturbance estimate and the high-frequency disturbance estimate, and equivalent disturbance parameters are output. A reconstruction instruction set is compiled based on the equivalent disturbance parameters and the topology protection correction matrix.

3. The high-precision robust control method for a megawatt-level high-torque motor according to claim 2, characterized in that, The edge lossless transmission weights specifically include: Graph convolution aggregation is performed on the node weight vector based on the feature compensation weight matrix and the adjacency matrix of the long-range reconnection topology to output a node connectivity strength sequence. The boundary connectivity fusion distribution is then calculated based on the node connectivity strength sequence. Extract the chip computing power margin ratio obtained from monitoring, perform threshold quantization calculation on the chip computing power margin ratio, and output the computing power composite constraint boundary; perform spatial intersection calculation on the boundary connectivity fusion distribution and the computing power composite constraint boundary, and output the topology protection zone; Based on the node coordinate distribution of the topological protection zone, time-domain difference calculation is performed on the high-frequency disturbance estimate to output the disturbance change rate sequence. Statistical significance screening is performed on the disturbance change rate sequence to output key mutation features. The directed transmission path length of the key mutation feature is calculated based on the connected edge weights of the topology protection band. The topology distance attenuation factor is calculated based on the directed transmission path length. Attenuation compensation calculation is performed on the key mutation feature based on the topology distance attenuation factor, and the edge lossless transmission weight is output.

4. The high-precision robust control method for a megawatt-level high-torque motor according to claim 3, characterized in that, The spatiotemporal decoupling perturbation analysis specifically includes: The transient convergence gradient and statistical dispersion of the error are calculated based on the decoupling error. The dynamic feedback gain weight is calculated based on the transient convergence gradient and the statistical dispersion. The adaptive state smoothing correction calculation is performed based on the decoupling error and the dynamic feedback gain weight, and the smoothing error is output. Based on the smoothing error and the dynamic inductor parameter set, calculate the cross-space phase shift compensation coefficient, and perform phase amplitude alignment and nonlinear state differential tracking calculation based on the smoothing error, the dynamic inductor parameter set and the cross-space phase shift compensation coefficient to output the generalized disturbance; The multi-scale frequency band cross-correlation energy distribution is calculated based on the generalized perturbation. The frequency band boundary nonlinear masking suppression threshold is calculated based on the multi-scale frequency band cross-correlation energy distribution. Transient adaptive gain compensation and residual suppression reconstruction are performed based on the frequency band boundary nonlinear masking suppression threshold, and the time-frequency decoupled generalized perturbation is output. Based on the time-frequency decoupling generalized perturbation, a multi-scale receptive field spatial distribution is constructed, the center-periphery antagonistic response gradient at the high-low frequency crossing boundary is calculated, and edge-side suppression compensation and dynamic sharpening reconstruction are performed based on the center-periphery antagonistic response gradient to output the spatiotemporal orthogonal perturbation. Based on the spatiotemporal orthogonal perturbation, calculate the local time-varying spectral density distribution and characteristic dynamic impedance parameters. Based on the local time-varying spectral density distribution and characteristic dynamic impedance parameters, calculate the standing wave ratio and impedance mismatch at the high-low frequency crossing boundary. Based on the standing wave ratio and impedance mismatch, perform traveling wave reflection compensation and characteristic impedance matching reconstruction, and output continuous orthogonal perturbation at the boundary. Based on the continuous orthogonal perturbation of the boundary, calculate the frequency band energy drift characteristics and extreme boundary thresholds. Based on the frequency band energy drift characteristics and extreme boundary thresholds, perform adaptive frequency band crossing boundary delineation and region-independent analytical calculation, and output low-frequency perturbation estimates and high-frequency perturbation estimates.

5. The high-precision robust control method for a megawatt-level high-torque motor according to claim 4, characterized in that, The generalized disturbance specifically includes: Extract the time-domain waveform extreme point sequence of the smoothing error and the inductance gradient change rate of the dynamic inductance parameter set, and calculate the spatial phase offset angle and amplitude scaling ratio based on the time-domain waveform extreme point sequence and the inductance gradient change rate; An elementary rotation matrix is ​​constructed based on the spatial phase offset angle, and a diagonal scaling matrix is ​​constructed based on the amplitude scaling ratio. The elementary rotation matrix and the diagonal scaling matrix are concatenated to output a composite phase shift operator. Matrix exponentiation is performed on the composite phase shift operator to output cross-space phase shift compensation coefficients. Extract the elementary rotation matrix and diagonal scaling matrix from the cross-space phase shift compensation coefficients, perform matrix multiplication rotation operation on the smoothing error based on the elementary rotation matrix, and output the rotated error sequence. Based on the diagonal scaling matrix, perform diagonal element multiplication scaling on the rotated error sequence to output an amplitude scaling error sequence; perform differential calculation on the amplitude scaling error sequence and the dynamic inductor parameter set to output a phase amplitude alignment error; Extract the time distribution statistical features of the phase amplitude alignment error sequence, and calculate the error variance and the cumulative sign value of the error sequence based on the time distribution statistical features of the error sequence; calculate the adaptive integration step size based on the error variance, and calculate the adaptive differential gain based on the cumulative sign value of the error sequence; Numerical integration is performed on the phase amplitude alignment error based on the adaptive integration step size, and an integral tracking term is output; numerical differentiation is performed on the phase amplitude alignment error based on the adaptive differential gain, and a differential tracking term is output; the integral tracking term and the differential tracking term are superimposed to output a generalized perturbation.

6. The high-precision robust control method for a megawatt-level high-torque motor according to claim 5, characterized in that, The manifold pre-evolution and orthogonal decoupling amplitude limiting compensation specifically include: Based on the desired current command, the speed error, and the decoupling error, calculate the time-domain statistical characteristics of the speed error and the spatial boundary characteristics of the decoupling error. Based on the desired current command, the time-domain statistical characteristics of the speed error, and the spatial boundary characteristics of the decoupling error, perform cross-domain tensor product correlation mapping and calculate the corresponding coefficients of the spatiotemporal coupling constraint quantization. Based on the spatiotemporal coupling constraint quantization corresponding coefficients, multi-dimensional composite constraints and dynamic scheduling are executed, multi-constraint dynamic scheduling convergence rate coefficients are calculated, and adaptive sliding surface parameters and decoupling error space boundary characteristics are output. Based on the adaptive sliding surface parameters and the rotational speed error, a time-varying sliding surface manifold is constructed. Based on the decoupling error spatial boundary features and the adaptive sliding surface parameters, a running mode feature sequence is extracted. Based on the running mode feature sequence, scene-based feature stripping is performed, and local feature-sensitive observation gradients are calculated. Based on the local feature sensitive observation gradient, calculate the boundary limit adaptive early warning factor, and based on the local feature sensitive observation gradient and the boundary limit adaptive early warning factor, perform pre-intervention controlled second-order calculus evolution on the time-varying sliding surface manifold to output the super-twisted second-order sliding mode control law. Based on the super-twisted second-order sliding mode control law, the integral sliding surface switching control component and the equivalent robust control component are calculated. Based on the physical actuator constraints, the execution boundary space mapping basis matrix is ​​calculated. Based on the execution boundary space mapping basis matrix, the integral sliding surface switching control component and the equivalent robust control component are subjected to cross-space orthogonal decoupling mapping, and the orthogonal decoupling component is calculated. Based on the execution boundary space mapping basis matrix, calculate the over-limit distortion evaluation index for the orthogonal decoupling component, perform adaptive amplitude limiting closed-loop compensation based on the over-limit distortion evaluation index, output the inner loop correction term, and superimpose the inner loop correction term with the desired current command to output the voltage reference component.

7. The high-precision robust control method for a megawatt-level high-torque motor according to claim 6, characterized in that, The over-limit edge closed-loop reconstruction specifically includes: Based on the reconstructed instruction set, spatiotemporal coupling constraint basis features are extracted; based on the voltage reference component, multidimensional weighted boundary features are extracted; based on the voltage reference component and the spatiotemporal coupling constraint basis features, cross-domain topological inverse mapping calculation is performed, and physical space mapping components are output. Based on the physical space mapping components, joint time-frequency analytical calculation is performed to output amplitude statistical distribution characteristics and phase gradient distribution characteristics. Based on the amplitude statistical distribution characteristics, target scope division is performed. Based on the target scope and the phase gradient distribution characteristics, scene configuration time-series weight extraction is performed to output initial scene time-series weights. Based on the multidimensional weight boundary characteristics, dynamic scheduling is performed on the initial scene time-series weights to output modulated wave controlled duty cycle parameters. Based on the controlled duty cycle parameter of the modulation wave, the narrow pulse over-limit depth value is calculated, the narrow pulse over-limit depth level factor is extracted based on the narrow pulse over-limit depth value, the carrier graded frequency deduction and pulse edge reconstruction quantization are performed based on the narrow pulse over-limit depth level factor and the controlled duty cycle parameter of the modulation wave, and a switching signal is output. The switching signal is output to the control terminal voltage of the power device.

8. A high-precision robust control system for a megawatt-level high-torque motor, comprising executing the high-precision robust control method for a megawatt-level high-torque motor as described in any one of claims 1 to 7, characterized in that, Includes the following modules: The cross-domain reconfiguration module is used to collect multi-source operating data, perform cross-domain reconfiguration of electromechanical parameters to generate reference resistance and rotating coordinate system current, and output the current motor state quantity through disturbance decoupling state reconfiguration. The cross-domain decoupling module is used to construct a backstepping decoupling matrix based on the current motor state variables, execute a cross-domain decoupling closed loop for electromechanical parameters, and output the decoupling error and dynamic inductance parameter set. The perturbation analysis module is used to perform spatiotemporal state reconstruction on the decoupling error and dynamic inductance parameter set to generate generalized perturbations, and to perform spatiotemporal decoupling perturbation analysis based on dynamic frequency band masking and impedance crossover reconstruction, outputting low-frequency perturbation estimates and high-frequency perturbation estimates. The boundary protection module is used to monitor the chip's computing power margin ratio and combine low-frequency disturbance estimates with high-frequency disturbance estimates to perform dimensionality reduction reconstruction and long-range reconnection to compensate for feature attenuation. After boundary protection correction and fusion of frequency bands, equivalent disturbance parameters and reconstruction instruction sets are generated. The amplitude limiting compensation module is used to extract equivalent disturbance parameters, reconstruct the instruction set, current motor state variables and decoupling error, calculate speed error and modulate disturbance to generate feedforward weight, combine the reconstructed instruction set to output quadrature axis reference current, perform spatiotemporal resonance amplitude limiting compensation after smooth reconstruction and superimpose it on quadrature axis reference current, and output the desired current command. The manifold evolution module is used to perform spatiotemporal coupling constraint scheduling, manifold pre-evolution and orthogonal decoupling limiting compensation based on the desired current command, speed error and decoupling error, and output voltage reference component; The edge reconstruction module is used to extract the voltage reference component and reconstruction instruction set, perform cross-domain coupled topology inverse mapping to generate the mapping component, dynamically optimize the output duty cycle parameter through the domain constraint, and deduce the carrier frequency based on the over-limit edge closed-loop reconstruction, and output the switching signal to the power device.

Images