Field weakening control system for axial flux hybrid excitation synchronous motor based on multi-source feedback

Abstract

This invention relates to the field of motor control technology and discloses a field weakening control system for an axial flux hybrid excitation synchronous motor based on multi-source feedback. The system includes constructing a digital twin model of the motor; collecting real-time operating data from multiple sources and synchronizing it to the model; establishing a three-dimensional electromagnetic field state space based on the model; extracting features using manifold neural operators to generate a multi-dimensional excitation state feature set; determining an excitation anomaly feature index set based on the feature set distribution density; collecting data in real-time and extracting current excitation feature parameters; mapping them to the three-dimensional electromagnetic field state space; calculating the spatial correlation with the anomaly index set; and outputting a field weakening demand correlation factor. Finally, iteratively updating the operator parameters based on the correlation factor, feature set, and real-time data, generating control commands to dynamically adjust the coupling ratio of the electrical excitation and permanent magnet excitation components. This system achieves comprehensive perception and precise control of the motor's operating state, improving the adaptability and reliability of field weakening control.

Description

Technical Field

[0001] This invention relates to the field of motor control technology, specifically to a field weakening control system for an axial flux hybrid excitation synchronous motor based on multi-source feedback. Background Technology

[0002] In fields such as new energy power generation, industrial drives, and rail transportation, axial flux hybrid excitation synchronous motors have been widely used due to their combination of the high efficiency of permanent magnet motors and the adjustable magnetic field of electrically excited motors. The operating performance of these motors largely depends on field weakening control technology, which adjusts the excitation component to achieve stable operation at high speeds. However, existing field weakening control systems still face many challenges in practical applications. Traditional field weakening control methods are mostly based on simplified mathematical models of the motor, relying on preset control parameters and empirical formulas. However, the electromagnetic field distribution of axial flux hybrid excitation synchronous motors exhibits strong nonlinearity and strong coupling characteristics. Furthermore, under different operating conditions, dynamic changes in parameters such as motor temperature, load, and speed lead to continuous changes in the electromagnetic field state. Simplified models cannot accurately reflect the actual operating state, easily resulting in insufficient control accuracy.

[0003] Existing systems have limitations in utilizing multi-source operational data. During motor operation, data such as current, voltage, speed, and temperature are collected in a scattered manner, lacking an effective fusion mechanism, resulting in incomplete extraction of characteristic parameters. Some systems rely solely on a single physical quantity for field weakening judgment, making it difficult to capture subtle changes in the electromagnetic field state. Under complex operating conditions, this can easily lead to misjudgments or lags, affecting the timeliness of field weakening control.

[0004] The coupling adjustment method of the excitation component has defects. The proportional adjustment of the electrical excitation component and the permanent magnet excitation component needs to respond to changes in the electromagnetic field in real time, but the parameter update cycle of traditional control algorithms is long and it is difficult to adapt to rapidly changing operating conditions. When the motor faces sudden load fluctuations or a sudden temperature rise, the excitation coupling proportional adjustment lags, which may lead to insufficient motor output, reduced efficiency, or even overheating and other safety hazards.

[0005] Existing technologies lack spatial analysis methods for electromagnetic field states. The distribution of electromagnetic fields inside a motor has three-dimensional spatial characteristics, and traditional two-dimensional analysis methods cannot fully present the spatial correlation of key features such as magnetic field strength and magnetic flux density distribution. This leads to insufficient accuracy in identifying abnormal excitation states, making it difficult to predict field weakening requirements in advance, and limiting further improvements in field weakening control performance. These problems collectively restrict the application expansion of axial flux hybrid excitation synchronous motors in high-performance drive scenarios. Summary of the Invention

[0006] The purpose of this invention is to provide a field weakening control system for an axial flux hybrid excitation synchronous motor based on multi-source feedback, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a field weakening control system for an axial flux hybrid excitation synchronous motor based on multi-source feedback, the system comprising:

[0008] Central processing module, the central processing module being used to execute:

[0009] A digital twin model of an axial flux hybrid excitation synchronous motor is constructed. Real-time operating data is collected by multi-source sensors deployed on the motor body, and the real-time operating data is synchronized to the digital twin model.

[0010] The digital twin model topology includes an electromagnetic field calculation module, a heat conduction analysis module, and a mechanical dynamics module;

[0011] A three-dimensional electromagnetic field state space is established based on the digital twin model, and motor operation characteristic parameters are extracted through manifold neural operators to generate a multi-dimensional excitation state feature set.

[0012] Based on the distribution density of the multi-dimensional excitation state feature set in the three-dimensional electromagnetic field state space, the excitation anomaly feature index set is determined;

[0013] Real-time acquisition of motor current, voltage, speed and temperature data; extraction of current excitation characteristic parameters and mapping to the three-dimensional electromagnetic field state space; calculation of the spatial correlation between the current excitation characteristic parameters and the excitation anomaly characteristic index set; output of field weakening demand correlation factor.

[0014] Based on the field weakening demand correlation factor, the multi-dimensional excitation state feature set, and real-time operating data, the manifold neural operator parameters are iteratively updated to generate field weakening control commands and dynamically adjust the coupling ratio between the electric excitation component and the permanent magnet excitation component.

[0015] Preferably, the real-time operating data collected by the multi-source sensors includes armature winding current waveform, air gap flux density distribution, rotor position angle, stator core temperature gradient, and cooling medium flow rate; the real-time operating data is subjected to time-domain noise reduction and frequency-domain filtering to generate a motor operating health dataset.

[0016] Preferably, the manifold neural operator includes a Laplace kernel integral module for performing feature decomposition on the motor operating health dataset:

[0017] Discretize the geometry of the motor design domain into a finite element mesh and solve for the Laplace operator characteristic function of the mesh nodes;

[0018] The armature current waveform and air gap flux density are spectrally decomposed using the Laplace operator characteristic function to extract the fundamental component amplitude, harmonic distortion rate and spatial phase offset, forming the multi-dimensional excitation state feature set.

[0019] Preferably, the excitation anomaly characteristic index set is determined through the following steps:

[0020] The coordinate axes of the three-dimensional electromagnetic field state space are defined based on the parameter dimensions of the multi-dimensional excitation state feature set.

[0021] Calculate the cluster density of characteristic points of historical field weakening fault events of the motor in the three-dimensional electromagnetic field state space;

[0022] The excitation characteristic parameters corresponding to spatial regions where the density of feature points exceeds a preset threshold are selected to form the excitation anomaly characteristic index set.

[0023] Preferably, the calculation of the magnetic field weakening demand correlation factor includes:

[0024] The amplitude and phase angle of the current fundamental flux linkage are analyzed based on the real-time current waveform and rotor position angle.

[0025] The current fundamental flux linkage amplitude and phase angle are combined into a dynamic field strength vector and mapped to the three-dimensional electromagnetic field state space;

[0026] Calculate the weighted sum of the Euclidean distances between the dynamic field strength vector and each feature point in the excitation anomaly feature index set, and output the field weakening demand correlation factor after normalization.

[0027] Preferably, the generation of the field weakening control command includes:

[0028] The weakening magnetic demand correlation factor and the multi-dimensional excitation state feature set are input into the updated manifold neural operator;

[0029] The encoder-approximator-decoder structure of the Laplace kernel integral module is used to predict the set value of the electric excitation winding current and the equivalent magnetic flux compensation of the permanent magnet.

[0030] The field weakening control command is generated based on the ratio between the set value of the electric excitation winding current and the equivalent magnetic flux compensation of the permanent magnet.

[0031] Preferably, the iterative update of the manifold neural operator employs the Fourier descriptor algorithm:

[0032] Convert the motor operating health dataset into a frequency domain feature tensor;

[0033] The low-rank core matrix of the frequency domain feature tensor is extracted by tensor network decomposition.

[0034] The weight coefficients of the Laplace kernel integral module of the manifold neural operator are corrected using the low-rank core matrix.

[0035] Preferably, the dynamic adjustment of the digital twin model includes:

[0036] After executing the field weakening control command, the actual air gap magnetic flux density change rate and temperature gradient distribution were collected;

[0037] Compare the residuals between the actual data and the predicted values ​​of the digital twin model to generate model correction coefficients;

[0038] The coordinate scale of the three-dimensional electromagnetic field state space is scaled according to the model correction coefficients, and the spatial boundary of the excitation anomaly characteristic index set is updated.

[0039] Preferably, the system further includes conflict detection for the field weakening control strategy:

[0040] Establish the coupling relationship matrix between the electrical excitation response delay time and the permanent magnet temperature demagnetization coefficient;

[0041] The control conflict risk value is calculated based on the product of the dynamic field strength vector and the coupling relationship matrix.

[0042] When the control conflict risk value exceeds the safety threshold, a redistribution strategy for the excitation component ratio is triggered.

[0043] Preferably, the closed-loop execution flow of the field weakening control system is as follows:

[0044] Dynamically load the field weakening control command based on the correlation factors related to the real-time field weakening demand.

[0045] Feedback data is collected once every preset period to update the digital twin model parameters and excitation anomaly characteristic index set;

[0046] The current control loop terminates when the rate of change of the weak magnetic field demand correlation factor is less than the convergence threshold in three consecutive iterations.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] This field weakening control system constructs a digital twin model of an axial flux hybrid excitation synchronous motor, synchronizing real-time operating data collected by multiple sensors into the model, thus achieving dynamic mapping between the physical motor and the virtual model. This mapping mechanism can reproduce the motor's operating state in real time, including complex physical processes such as electromagnetic field distribution and temperature field changes. It breaks through the limitations of traditional control systems that rely on simplified models, enabling the system to more comprehensively perceive the actual operating conditions of the motor.

[0049] The deployment of multi-source sensors provides the system with abundant raw data, covering multiple dimensions such as current, voltage, speed, and temperature. This data, after synchronous processing, is input into the digital twin model, laying the foundation for subsequent feature extraction. Compared to a single data source, multi-dimensional data input can capture various characteristics of motor operation, reducing feature loss due to data limitations and making the extracted excitation state features closer to the actual state of the motor.

[0050] The three-dimensional electromagnetic field state space established based on the digital twin model provides a spatial perspective for the analysis of excitation characteristics. The three-dimensional space can accurately represent the distribution of the electromagnetic field within the motor, including spatial characteristics such as magnetic field strength and magnetic flux density gradient at different locations, overcoming the shortcomings of traditional two-dimensional analysis in spatial description. The manifold neural operator extracts operational characteristic parameters in the three-dimensional space, effectively capturing the nonlinear and strongly coupled characteristics of the electromagnetic field. The generated multi-dimensional excitation state feature set is more representative and can comprehensively reflect the operating state of the excitation system.

[0051] The determination of the excitation anomaly characteristic index set is based on the distribution density of the multi-dimensional excitation state characteristic set in three-dimensional space. Anomaly patterns are identified by analyzing the spatial distribution patterns of the characteristics. This method does not rely on preset thresholds but instead mines anomaly features from the natural distribution of the characteristics. It can adapt to the characteristic change patterns under different operating conditions, has a stronger ability to identify potential anomalies in the excitation system, and can promptly detect abnormal factors that may affect weak magnetic properties.

[0052] By calculating the spatial correlation between the current excitation characteristic parameters and the excitation anomaly characteristic index set, the field weakening demand correlation factor can be obtained, which quantifies the degree of correlation between the current state and the abnormal state, providing a precise demand signal for field weakening control. This quantification method avoids the experience-based qualitative judgment in traditional control, making the assessment of field weakening demand more objective, and allowing for dynamic adjustment of the control strategy based on the actual correlation degree.

[0053] The system iteratively updates the parameters of the manifold neural operator, enabling continuous optimization of feature extraction capabilities as operational data accumulates, thus adapting to performance changes in the motor over long-term operation. Simultaneously, it generates field-weakening control commands based on the updated parameters, dynamically adjusting the coupling ratio between the electrical excitation component and the permanent magnet excitation component. This allows for real-time response to changes in the electromagnetic field, ensuring that the excitation component adjustment better matches the current operating conditions, avoiding performance fluctuations caused by adjustment lag, and guaranteeing stable motor operation under different speeds and loads. Attached Figure Description

[0054] Figure 1 This is a timing diagram of the field weakening control system for the axial flux hybrid excitation synchronous motor based on multi-source feedback described in this invention.

[0055] Figure 2 A flowchart of the eigenvalue decomposition of manifold neural operators;

[0056] Figure 3 This is a diagram of multi-source sensor data acquisition and preprocessing.

[0057] Figure 4 A flowchart for calculating the correlation factor for weak magnetic field demand;

[0058] Figure 5 A three-dimensional feature space and a magnetic weakening control map;

[0059] Figure 6 A flowchart for dynamically adjusting a digital twin model. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Please see Figure 1 This invention provides a field weakening control system for an axial flux hybrid excitation synchronous motor based on multi-source feedback, the specific implementation of which is as follows:

[0062] A digital twin model of an axial flux hybrid excitation synchronous motor is constructed. Real-time operating data is collected by multi-source sensors deployed on the motor body and synchronized to the digital twin model.

[0063] A three-dimensional electromagnetic field state space is established based on the digital twin model, and motor operation characteristic parameters are extracted through manifold neural operators to generate a multi-dimensional excitation state feature set.

[0064] Based on the distribution density of the multi-dimensional excitation state feature set in the three-dimensional electromagnetic field state space, the excitation anomaly feature index set is determined.

[0065] Real-time acquisition of motor current, voltage, speed and temperature data, extraction of current excitation characteristic parameters and mapping to the three-dimensional electromagnetic field state space, calculation of the spatial correlation between the current excitation characteristic parameters and the excitation anomaly characteristic index set, and output of field weakening demand correlation factor.

[0066] Based on the field weakening demand correlation factor, the multi-dimensional excitation state feature set, and real-time operating data, the manifold neural operator parameters are iteratively updated to generate field weakening control commands and dynamically adjust the coupling ratio between the electric excitation component and the permanent magnet excitation component.

[0067] Example 1: See Figure 2Multi-source sensors deployed on the motor body continuously collect real-time operating data. The sensor types include current sensors, flux density probes, rotor position encoders, distributed temperature sensors, and flow meters. The collected data includes instantaneous waveforms of the armature winding currents in each phase, flux density distribution at multiple detection points along the air gap circumference, real-time changes in the rotor rotation angle, axial and radial temperature gradient distribution of the stator core, and instantaneous flow rate of the cooling medium in the pipes. The armature winding current waveform is acquired using a Rogowski coil structure, with common-mode interference eliminated by a differential amplifier circuit; the air gap flux density distribution is measured using a Hall effect sensor array, with the detection points arranged at equal intervals according to the number of motor pole pairs; the rotor position angle is fed back by an absolute encoder mounted on the shaft end; the stator core temperature gradient is obtained using platinum resistance sensors embedded in the slot bottom and tooth top; and the cooling medium flow rate is recorded in real-time by a turbine flow meter. All sensor data undergoes signal conditioning via an isolated sampling circuit, is digitized by a high-speed synchronous acquisition card, and ultimately forms a multi-channel data stream with strictly aligned timestamps.

[0068] Temporal denoising is performed on the data stream. For the armature winding current waveform, a wavelet threshold denoising algorithm employs a multi-scale decomposition strategy: the signal is decomposed into multiple levels of approximation coefficients and detail coefficients using the Mallat algorithm, and the optimal number of decomposition levels is automatically selected based on the motor's operating frequency characteristics. The detail coefficients of each level are processed by an adaptive threshold function, with the threshold set as a nonlinear function of the standard deviation estimate of the noise level and the number of decomposition levels. The processed detail coefficients and the retained approximation coefficients are then reconstructed to form the denoised current waveform. For the air gap flux density distribution data, frequency domain filtering is performed using a three-dimensional Fourier transform: the sequence of detection points along the circumference is used as the spatial dimension, and the time sampling points are used as the time dimension to construct a spatiotemporal cube data structure. The fundamental component and higher harmonic components are separated using a fast Fourier transform, and a bandpass filter is designed to extract the effective signal within a specific frequency band. The filtered fundamental component is then reconstructed into the spatial domain flux density distribution using an inverse Fourier transform.

[0069] The denoised data is input into the digital twin model. The model's topology includes an electromagnetic field calculation module, a heat conduction analysis module, and a mechanical dynamics module. The electromagnetic field calculation module is constructed based on the finite element method, discretizing the motor's stator and rotor geometry into tetrahedral element meshes, with each element assigned corresponding material permeability and conductivity. The heat conduction analysis module uses a thermal network model, with nodes corresponding to the stator core temperature sensor positions, and network branch parameters calculated based on cooling medium flow rate and material thermal conductivity. The mechanical dynamics module uses multibody dynamics equations, combined with rotor position angle data, to solve for bearing vibration characteristics. The three modules are coupled via a data bus. The output magnetic flux distribution of the electromagnetic field module serves as the heat source input for the heat conduction module, the temperature field distribution of the heat conduction module is fed back to the electromagnetic field module to correct material properties, and the vibration results from the mechanical dynamics module trigger the boundary condition reconfiguration of the electromagnetic field module. The model periodically receives denoised real-time operating data, synchronously updates internal state variables, and outputs a complete dataset containing transient magnetic field intensity distribution, core temperature field cloud map, and vibration spectrum, constituting the motor's operating health dataset.

[0070] Data preparation and preprocessing before construction include: 1. Geometric and material data acquisition: 3D laser scanning technology is used to acquire the geometric parameters of the motor stator and rotor, including stator inner / outer diameter, core stack thickness, permanent magnet dimensions, and air gap length, ensuring the geometric model accuracy error is ≤ ±0.02mm; a material property database is established, including the permeability curves (relative permeability values ​​under different magnetic field strengths) and conductivity (…) of the stator core silicon steel sheets. The remanence of the permanent magnet (≥1.2T at 20℃), coercivity (≥900kA / m), and resistivity of the winding copper wire ( ); 2. Sensor data preprocessing: preprocess the real-time data collected by multi-source sensors: armature current is de-noised by wavelet threshold (decomposed into 8 levels of approximation coefficients and detail coefficients and then reconstructed), and air gap magnetic flux density is extracted by three-dimensional Fourier transform to extract the fundamental component; the preprocessed data must meet the following requirements: sampling frequency ≥ 1 kHz, data integrity ≥ 99.5%.

[0071] The construction method of the multiphysics module includes: 1. Electromagnetic field calculation module: Based on finite element software, a model is constructed, and the stator and rotor geometry is discretized into tetrahedral element meshes. The element size of the air gap and permanent magnet edge region is controlled within 0.5-2mm (the proportion of deformed elements is <5%). The stator outer circle is set as the parallel boundary of magnetic field lines, the rotor surface as the rotation boundary, and the air gap as the coupling boundary. The armature winding current waveform and the remanent magnetization excitation of the permanent magnet are applied to solve the air gap magnetic flux density distribution and electromagnetic torque. The error between the calculation results and experimental values ​​should be ≤±5%; 2. Thermal conduction analysis module: A thermal network model is used to divide ≥100 thermal nodes, where the stator core slot bottom and tooth top nodes correspond one-to-one with the temperature sensor positions. The convective heat transfer coefficient is calculated according to the cooling method (10-20 W / (m²·K) for air cooling and 500-1000 W / (m²·K) for liquid cooling), and the contact thermal resistance of the components (0.1-0.3 W / (m²·K) between the core and the casing) is determined by the thermal conduction formula. m²·K / W); Using the iron and copper losses output from the electromagnetic field module as heat source inputs, solve for the temperature field distribution, with an error ≤ ±3℃; 3. Mechanical Dynamics Module: Establish a rotor-bearing system model based on the multibody dynamics equations, and define the rotor moment of inertia (calculated based on material density and geometric dimensions) and bearing stiffness coefficient (m²·K / W); The damping coefficient (100-500 N·s / m) is calculated; the rotor unbalance excitation (unbalance ≤5g·mm) and electromagnetic torque fluctuation excitation are applied, and the radial vibration displacement (≤0.05mm) and vibration velocity (≤2.8 mm / s) are calculated with an error of ≤±10%.

[0072] The module coupling and real-time interaction mechanism includes: 1. Data bus architecture: Module collaboration is achieved through an EtherCAT protocol data bus (communication rate ≥100Mbps): the electromagnetic field module transmits magnetic flux distribution data to the heat conduction module every 10ms; the heat conduction module feeds back temperature field data every 20ms to correct the material's magnetic permeability (the magnetic permeability of silicon steel sheets decreases by 2%-3% for every 10℃ increase in temperature); the mechanical dynamics module transmits vibration data every 50ms, triggering the reconfiguration of the air gap boundary conditions in the electromagnetic field module (vibration causes air gap changes ≤±0.01mm); 2. Standardized interface design: The model output adopts an FMU interface file conforming to the FMI standard, which includes transient magnetic field intensity distribution, core temperature field cloud map, and vibration spectrum dataset, supporting real-time data interaction with the upper-level health management system.

[0073] The model calibration and self-correction process includes: 1. Operating condition verification: Select 10 typical operating conditions within the range of speed 500-3000 r / min and load rate 0-120%, collect the current, temperature, and magnetic flux density data of the physical motor and compare them with the model simulation results, and calculate the error threshold (current ±5%, temperature ±3℃, magnetic flux density ±5%); 2. Self-correction mechanism: Optimize the correction coefficients using the response surface methodology: Select shaft power and rotational frequency as operating condition variables, obtain the actual values ​​of stator current and excitation current through testing, construct the correction coefficient vector (xi1-xi4), minimize the deviation between the model calculated value and the actual value, and ensure long-term operating accuracy.

[0074] The Laplace kernel integral module of the manifold neural operator is initiated. This module receives the motor operating health dataset output by the digital twin model, with magnetic field strength distribution data as the primary input source. The Laplace kernel integral module first analyzes the geometry of the motor design domain, which includes the stator inner surface, the permanent magnet array profile, and the air gap boundary profile. The geometry is discretized into a non-uniform finite element mesh: regions with large curvature variations use locally refined meshes, while flat regions use sparse meshes. The spatial coordinates of the mesh nodes are input to the Laplace characteristic equation solver, which uses an inverse power iteration method to calculate the characteristic function space. The characteristic function values ​​generate an orthogonal normalized basis function system at the mesh nodes. The analyzed motor design domain geometry is discretized using a non-uniform finite element mesh. Regions with large curvature changes (such as the edges of permanent magnets and narrow air gaps) are locally meshed with finer meshes, while flat regions (such as the stator core plane) are meshed with sparse meshes. The mesh element type is tetrahedral, and each element is assigned the permeability (such as the permeability of silicon steel sheets and the remanent permeability of permanent magnets) and conductivity parameters corresponding to the material. The Laplace operator eigenfunctions of the mesh nodes are solved using an inverse power iteration method, with the magnetic field distribution under no-load conditions of the motor as the initial boundary condition. The iteration terminates when the error between two consecutive eigenvalue calculations is less than [a certain value]. The obtained characteristic functions must satisfy the orthogonal normalization condition (the inner product of any two different characteristic functions is 0, and the inner product of the same characteristic function is 1), and finally generate an orthogonal normalized basis function system covering all grid nodes.

[0075] Spectral decomposition of the armature current waveform is performed based on this function system. Current waveform data is extracted from the motor operating health dataset and a two-dimensional time-frequency diagram is generated through time-frequency conversion. The basis vectors of the characteristic function system are orthogonally projected onto the current time-frequency diagram: the inner product of each characteristic function and the time-frequency diagram is calculated to generate a projection coefficient sequence. The maximum amplitude of the coefficient sequence corresponds to the fundamental current component, and the amplitudes of the remaining coefficients constitute the harmonic component distribution. The fundamental component amplitude is normalized to a percentage scalar value, and the distortion rate index of the harmonic component distribution is obtained by calculating the squared amplitude and root mean square of the harmonic component distribution. Spatial phase analysis is performed on the air gap flux density data: the flux density distribution map is unfolded into a spatial waveform signal according to angular coordinates and convolved with the characteristic function system. The phase angle sequence of the output response is smoothed and filtered, and the standard deviation of its phase offset angle from the ideal sine waveform is statistically analyzed. Simultaneously, magnetic flux density distribution data of air gap circumferential detection points (the number of detection points matches the number of motor pole pairs, such as 16 detection points for an 8-pole motor) are extracted from the motor operation health dataset to construct a spatiotemporal cube data structure of "spatial dimension (detection point sequence) - time dimension (sampling point sequence)". A three-dimensional fast Fourier transform is performed on this structure to separate the fundamental component and higher harmonic components. A bandpass filter (passband range is the motor fundamental frequency ± 5Hz) is designed to extract the effective signal. The Laplace operator characteristic function is convolved with the filtered fundamental component to obtain the spatial phase response sequence. After 5-point moving average smoothing filtering, the standard deviation of its phase offset angle from the ideal sine waveform is calculated.

[0076] The final generated excitation state feature set contains three core parameters: the normalized amplitude of the fundamental current component, the scalar value of the current harmonic distortion rate, and the standard deviation of the spatial phase offset angle. Each parameter is strictly bound to the timestamp of the digital twin model, forming a vector dataset arranged in a time series. The dimensionality of this dataset is determined by the sampling period and the parameter type, where the fundamental amplitude and harmonic distortion rate are single-valued sequences, while the spatial phase offset is a multi-dimensional vector sequence with a vector length equal to the number of spatial detection points. All data are uniformly stored in a three-dimensional tensor structure, with the first dimension being the time axis, the second dimension being the parameter category axis, and the third dimension being the spatial location axis, forming a complete multi-dimensional excitation state feature set. The calculation rules for the three core parameters are as follows: 1. Fundamental component amplitude: The amplitude of the fundamental component of the armature current is normalized. The normalization reference is the amplitude of the fundamental current under rated operating conditions of the motor, i.e., "Normalized fundamental amplitude = Actual fundamental amplitude / Rated fundamental amplitude". The result is retained to 3 decimal places, with a value range of 0.5-1.2; 2. Harmonic distortion rate: Calculated according to "Harmonic distortion rate = (Root mean square of the sum of squares of the amplitudes of each harmonic component) / Amplitude of the fundamental component × 100%". The result is retained to 1 decimal place, with a value range of 0%-15%. The 3rd, 5th, and 7th harmonics are monitored in particular; 3. Spatial phase offset: Taking the first detection point on the air gap circumference as the reference, the difference between the phase of the remaining detection points and the phase of the reference point is calculated. The standard deviation of all differences is taken, in degrees (°), and the integer part is retained, with a value range of 0°-30°. The resulting multi-dimensional excitation state feature set is stored using a three-dimensional tensor structure. The first dimension is the time axis (timestamps are synchronized with the sensor acquisition cycle, such as 10ms / time), the second dimension is the parameter category axis (including normalized fundamental amplitude, harmonic distortion rate, and spatial phase offset), and the third dimension is the spatial position axis (corresponding to the number of air gap detection points). The spatial phase offset has a specific value on the spatial position axis, while the normalized fundamental amplitude and harmonic distortion rate are global parameters. The feature set is stored using a circular buffer (the earliest record is discarded when the time axis data reaches 1000 timestamps) and is transferred via zero-copy sharing of memory. The transfer is triggered when the digital twin model completes the calculation of the current time step.

[0077] The feature set is updated in real time and input into the downstream processing flow. The update is triggered when the digital twin model completes the calculation cycle of the current time step. The feature set data storage adopts a circular buffer structure. When the time axis data volume reaches the preset length, the earliest time point record is automatically discarded and the latest calculation result is added. The parameter passing process achieves zero-copy transmission through shared memory, avoiding the time delay introduced by data copying. The feature set's lifecycle covers the entire process from neural operator startup to control loop termination, continuously recording the dynamic evolution trajectory of the motor's operating state.

[0078] See Figure 3This showcases the analysis of multi-source sensor data. It includes three core observation indicators: Current waveform denoising: a comparison of the original current waveform (dark gray) and the denoised waveform (light gray), demonstrating how wavelet threshold denoising effectively eliminates high-frequency interference while preserving fundamental characteristics; Air gap magnetic flux density distribution: displaying the magnetic flux density measurements at 36 detection points along the circumference, exhibiting a typical sinusoidal distribution; Stator core temperature gradient: axial (dark gray) and radial (light gray) temperature distributions reveal cooling non-uniformity.

[0079] Example 2: See Figure 4 The construction of the three-dimensional electromagnetic field state space is based on a multi-dimensional excitation state feature set. This feature set contains three core parameters: the amplitude of the fundamental current component, the current harmonic distortion rate, and the standard deviation of the spatial phase offset. These three parameters are defined as three orthogonal coordinate axes of the spatial coordinate system: the fundamental amplitude parameter corresponds to the X-axis, the distortion rate parameter corresponds to the Y-axis, and the standard deviation of the phase offset corresponds to the Z-axis. The coordinate axis dimensions are normalized, using the maximum and minimum values ​​of the historical operating data of each parameter as the scaling reference, and converting the actual values ​​into relative values ​​from 0 to 1 through linear mapping. The origin of the coordinate system is defined as the reference value of each parameter under ideal no-load conditions, ensuring that the feature points of different operating states are radially distributed in space.

[0080] Feature points from historical field weakening fault events are imported into this spatial model. The fault event data originates from the motor's historical operation database, containing feature set data collected within a specific time window before and after the fault occurrence. Each fault event is represented in space as a discrete feature point, whose position coordinates are determined by the fundamental amplitude, distortion rate, and phase shift recorded at that moment. The feature point cluster density calculation employs an adaptive kernel density estimation algorithm: a three-dimensional Gaussian kernel function is generated centered on each feature point, with the kernel function bandwidth adaptively adjusted according to the point density in the spatial region. Density value calculation traverses the entire spatial grid; the density value of each grid cell is the sum of the probability densities of all kernel functions at that cell location. The grid cell size is set to a fixed fraction of the parameter variation range to ensure uniformity of spatial resolution.

[0081] When the density value of a spatial region exceeds a preset threshold, that region is marked as an abnormally high-incidence area. The preset threshold is determined based on the statistical values ​​of density distribution under historical normal operating conditions: the upper limit of the density distribution of characteristic points under normal operating conditions is calculated, multiplied by a safety factor, and used as the judgment criterion. The spatial boundary of the abnormally high-incidence area is determined using a convex hull algorithm: the coordinates of all characteristic points within the region are extracted, and the minimum convex polyhedron boundary is calculated. The mathematical expression of each boundary surface is stored as a spatial plane equation. Finally, the set of boundary equations for all abnormally high-incidence areas constitutes the excitation anomaly characteristic index set. This index set is stored in a tree structure, with the root node recording the overall spatial density distribution, and child nodes storing the boundary parameters of each region in a hierarchical spatial octree structure.

[0082] The real-time data acquisition system continuously acquires motor current waveforms and rotor position angle data. The current waveform is acquired at a high frequency through a three-phase synchronous sampling circuit, with the sampling rate set to an integer multiple of the power system's fundamental frequency. The fundamental current component analysis employs digital phase-locked loop (PLL) technology: the PLL reference signal is generated from the rotor position angle signal through coordinate transformation; the phase difference between the reference signal and the actual current is compared using a phase detector; and the fundamental phase offset is output after passing through a loop filter. The fundamental amplitude is calculated using a sliding window RMS algorithm: the squares of the current sample values ​​are summed within one fundamental cycle, and the square root is taken to obtain the effective value. The rotor position angle is directly output by an absolute encoder, and the angle resolution meets the requirements for spatial phase calculation.

[0083] The fundamental flux linkage amplitude and phase angle obtained from the analysis are combined to form the dynamic field strength vector. The vector uses complex representation: the real part is the normalized value of the flux linkage amplitude, and the imaginary part is the radian value of the phase angle. When mapping this vector to the three-dimensional electromagnetic field state space, a third dimension parameter needs to be added. The real-time current harmonic distortion rate is calculated through FFT analysis: a Fast Fourier Transform is performed on the current sampling window, and the ratio of the sum of the squares of the amplitudes of all harmonic components other than the fundamental wave to the fundamental wave amplitude is the distortion rate. The standard deviation of the spatial phase offset is calculated using the current air gap flux density distribution data: the deviation of the instantaneous phase angle from the average value at multiple detection points is statistically analyzed to obtain the standard deviation. Thus, the complete coordinates of the dynamic field strength vector in three-dimensional space are determined.

[0084] The spatial correlation calculation process is initiated. The data structure of the excitation anomaly characteristic index set is read, and all high-incidence anomaly sub-nodes are traversed. For each anomaly region, the following operations are performed: Calculate the minimum distance from the dynamic field strength vector to the region's boundary. If the vector is located inside the region, the distance is recorded as a negative value. Multiply this distance by the region's density value to obtain the initial correlation degree for that region. The correlation degrees of all regions are weighted and summed to obtain the total correlation value, with the weighting factor being the historical fault occurrence frequency of each region. The total correlation value is input into the Sigmoid function for normalization, and the output range is mapped to the interval between 0 and 1, generating a field weakening demand correlation factor. The larger the value of this factor, the higher the similarity between the current state and historical fault characteristics, and the more urgent the need for field weakening control.

[0085] The data update mechanism adopts an event-triggered approach. Once the field weakening demand correlation factor is generated, the system automatically records the spatial coordinates of the current dynamic field strength vector, storing it as a new feature point in the historical database. This point participates in subsequent density distribution recalculation: the update process of the excitation anomaly feature index set is only triggered when the change in spatial density distribution caused by the new point exceeds a set threshold. The update process uses an incremental algorithm, recalculating only the density values ​​of the local space surrounding the new point, thus optimizing computational efficiency. The spatial coordinate scaling benchmark is periodically calibrated: after accumulating a certain number of new feature points, the maximum and minimum values ​​of each parameter are recalculated, and the coordinate mapping relationship is adjusted to maintain the effectiveness of the spatial distribution.

[0086] The lifecycle management of the abnormal feature indicator set includes a version control mechanism. A new indicator set version is generated after each major update, and historical version data is archived and stored. When a change in the motor's structural parameters is detected, the system automatically clears the historical feature point database and initializes a completely new three-dimensional spatial structure. The indicator set data uses distributed storage; the master node stores complete data, while each execution node caches local spatial partition data, ensuring data synchronization through a consistency protocol. The data access interface uses a binary protocol to reduce serialization overhead and meet the time constraints of real-time control.

[0087] See Figure 5 This document showcases the analysis of a field weakening control system. It includes three key processes: a three-dimensional electromagnetic field state space (X-axis fundamental current amplitude, Y-axis harmonic distortion rate, Z-axis phase offset standard deviation – light gray dots: normal operating conditions, dark gray triangles: abnormal operating conditions); evolution of field weakening demand correlation factors (reflecting the spatial correlation between real-time status and historical faults – 0-1 range, >0.7 triggers control); and dynamic adjustment of excitation components (electric excitation (dark gray) and permanent magnet excitation (light gray) components changing with field weakening demand, with the dashed line representing the proportion of electric excitation (40%-80% dynamic range).

[0088] Example 3: In the field weakening control command generation stage, the updated manifold neural operator receives the field weakening demand correlation factor and historical multi-dimensional excitation state feature set from the real-time system. The manifold neural operator employs a hierarchical processing architecture, with its core component, the Laplace kernel integral module, consisting of an encoder, an approximator, and a decoder connected sequentially. The encoder first performs feature fusion on the input data: the field weakening demand correlation factor is used as a scalar weight coefficient and weighted and concatenated with the three-dimensional tensor of the historical excitation state feature set. The concatenated tensor undergoes spatial dimension compression through a multi-channel convolutional layer, outputting a dimensionality-reduced feature vector. This vector is input to the approximator component, which uses a radial basis function network structure with Gaussian kernel functions as activation functions for its hidden layer nodes. The network output contains two independent branches: the first branch generates a scalar output of the electric excitation winding current setpoint, and the second branch outputs a vector sequence of the equivalent magnetic flux compensation amount of the permanent magnet.

[0089] The current setpoint scalar is limited to a range within the safe operating range of the electrically excited winding. The length of the flux compensation vector sequence matches the number of permanent magnet pole pairs, with each vector element corresponding to the compensation amount for a specific pole pair. The decoder component performs proportional fusion on the two branch outputs: calculating the ratio of the current setpoint to the average flux compensation amount, and converting this ratio into a percentage coupling ratio coefficient. The final generated field weakening control command contains three data fields: the target value of the electrically excited current (in amperes), the permanent magnet compensation distribution map (spatial vector), and the coupling ratio coefficient (percentage) of the electrically excited and permanent magnet excitation components. The command is transmitted to the power drive unit via a high-speed communication bus. The drive unit adjusts the excitation power supply output according to the target current value and simultaneously adjusts the permanent magnet operating point according to the compensation distribution map.

[0090] The parameter updates of the manifold neural operator employ the Fourier descriptor algorithm. Upon initiation, the latest motor operating health dataset is obtained from the digital twin model. The dataset undergoes a frequency domain transformation: three-dimensional Fourier transforms are performed on the time, spatial, and parameter category axes. The transformed complex spectral data constitutes a frequency domain feature tensor, with the tensor dimension consistent with the original dataset. Tensor network decomposition uses the Tucker model: the original tensor is decomposed into a product of a kernel tensor and multiple factor matrices. The kernel tensor extraction process is implemented through alternating least squares iterations, with a target decomposition order set to a preset value. After decomposition, the low-rank approximation of the kernel tensor is retained, while higher-order components with singular values ​​below a threshold are discarded.

[0091] The resulting low-rank kernel tensor is used to correct the parameters of the Laplacian kernel integral module. Specifically, the kernel tensor is expanded into a two-dimensional matrix, and its Hadamard product with the current weight matrix of the neural operator is calculated. The product result is normalized and then replaces a specific subset of the original weight coefficients. Weight updates focus on the encoder's convolution kernel parameters and the decoder's fully connected layer parameters, while the coordinates of the radial basis function center points of the approximator components remain unchanged. The updated manifold neural operator is immediately incorporated into the real-time control process, and its output is checked for consistency with the predictions of historical versions of the operator. If a significant deviation is detected, a weight rollback mechanism is triggered to revert to the previous stable version.

[0092] The command loading and execution process in closed-loop control is as follows: The real-time monitoring module for field weakening demand correlation factors continuously outputs the latest factor values. The system maintains a command priority queue; when the factor value change exceeds a set threshold, the current control command is marked as pending update. Before loading an updated command, a safety check must be performed: verifying whether the target value of the electric excitation current exceeds the driver's operating range and whether the permanent magnet compensation exceeds the material's magnetic flux saturation. After passing the check, the new command is loaded in two stages: the first stage smoothly transitions the electric excitation current value in the form of a ramp function; the second stage injects the magnetic flux compensation after the current stabilizes. The time constant of the transition process is dynamically adjusted according to the rotor speed; the higher the speed, the smaller the time constant.

[0093] The control loop termination logic is based on the changing trend of the magnetic weakening demand-related factor. The system maintains a circular buffer to store the factor values ​​of the most recent three iterations, and performs the following calculations each time a new factor is generated:

[0094]

[0095] in: Indicates the factor value at the current time; and These are the values ​​from the first two iterations, respectively. Indicates the average magnitude of change. When When the parameters remain below the preset convergence threshold, the system freezes the parameter update process of the manifold neural operator, maintaining the current output of the magnetic weakening control command. Continuous monitoring is performed during the freeze period. If the value exceeds the threshold set ratio again, the control loop is immediately restarted. After the loop terminates, the power drive unit switches to command sustain mode, in which only current closed-loop tracking is performed, and no longer responds to new field weakening control command update requests.

[0096] The iteration cycle management employs an adaptive clock mechanism. The basic iteration cycle is set to a fixed value, but is dynamically fine-tuned based on the motor's operating status: when the armature current harmonic distortion rate increases, the cycle is automatically shortened to increase the control frequency; when the temperature gradient distribution tends to stabilize, the cycle is extended to reduce the computational load. At the beginning of each cycle, the central processing module verifies the readiness status of each submodule, and only executes a new iteration after confirming the integrity of data synchronization. Historical iteration data records include timestamps, input feature sets, output commands, and execution status codes, forming a complete control process traceability chain.

[0097] Example 4: See Figure 6After the field weakening control command is executed, the multi-source sensors immediately initiate the feedback monitoring process. In a specific example, a certain type of axial flux motor receives a field weakening control command during a speed increase. The command requires adjusting the setpoint of the electrical excitation current to a specific value and simultaneously increasing the flux compensation in the N-pole region of the permanent magnet. After the command is loaded, sixteen flux density probes deployed around the air gap circumference capture transient data at a high sampling rate, with each probe recording the rate of change of flux density over time. Eight temperature sensors installed at the bottom of the stator core slots synchronously collect the temperature gradient distribution, with the sensors arranged equidistantly along the axis to form a temperature matrix.

[0098] Actual monitoring data is compared with the predicted values ​​of the digital twin model. The residual calculation for the rate of change of magnetic flux density uses a point-by-point differencing method: the difference between the measured and predicted values ​​at the same spatial location and time stamp is calculated to generate a residual sequence. The temperature gradient residual calculation considers spatial distribution characteristics: the measured and predicted temperatures at eight sensor locations are constructed as vectors, and the absolute difference between corresponding elements of the vectors is calculated. The residual data is compiled into a monitoring report, which includes three elements: spatial location number, time series index, and residual values.

[0099] The model correction coefficients are generated based on residual statistical analysis. For the magnetic flux density residual, the root mean square value of the residuals of all probes throughout the entire sampling period is calculated. For the temperature gradient residual, the Manhattan norm of the vector difference is calculated. The two residual indices are weighted and fused, and then converted into correction coefficients using a logarithmic scaling function. The specific rule is as follows: when the residual index is below the normal fluctuation range, the correction coefficient approaches 1; when the residual exceeds the warning threshold, the correction coefficient decreases exponentially.

[0100] Referring to Table 1, based on the data in the table, the system calculates the root mean square of the magnetic flux residual to be 0.035 and the norm of the temperature gradient residual to be 14.3. The weighted residual index is then calculated, and a coefficient of 0.89 is output through a correction function.

[0101] Table 1: Residual data at key locations at a certain moment.

[0102]

[0103] The coordinate scale of the three-dimensional electromagnetic field state space is adjusted according to this coefficient. The original coordinate axis ranges are scaled proportionally: the fundamental amplitude axis range is adjusted from [0.5, 1.2] to [0.445, 1.068], the harmonic distortion rate axis range is adjusted from [0%, 15%] to [0%, 13.35%], and the phase offset axis range is adjusted from [0°, 30°] to [0°, 26.7°]. The coordinates of all historical feature points in the space are updated synchronously, and the density distribution is recalculated. After the update, a shrinkage of the boundary of a certain high-incidence area is detected: the original boundary vertex coordinates (0.78, 9.2%, 18°) are adjusted to (0.694, 8.2%, 16°). The system automatically deletes two abnormal regions whose density values ​​are lower than the new threshold due to coordinate scaling, and finally generates an updated excitation anomaly feature index set.

[0104] The conflict detection module starts after the control command takes effect. In the example, the adjustment of the electric excitation current setpoint causes a delay in the excitation winding response. The delay time is measured by the rise time of the current closed-loop control, recording the time difference from when the command is issued to when the actual current reaches 90% of the target value. Simultaneously, permanent magnet temperature monitoring shows that the temperature in the N-pole region rises due to the increased magnetic flux compensation, and the temperature change rate recorded by temperature sensor PM_N04 exceeds that of adjacent regions. The system retrieves a pre-stored coupling relationship matrix. The rows of this matrix represent the electric excitation response delay time levels (e.g., 0-5ms, 5-10ms, 10-15ms), and the columns represent the permanent magnet temperature change rate levels (e.g., 0-3°C / s, 3-6°C / s, >6°C / s). The matrix element values ​​are the probability values ​​of conflicts triggered by similar historical events.

[0105] When the delay time is in the range of 5-10ms and the temperature change rate is in the range of 3-6°C / s, the collision probability is 0.35.

[0106] When the delay time is greater than 10ms and the temperature change rate is greater than 6°C / s, the collision probability jumps to 0.81.

[0107] The current dynamic field strength vector and matrix operation process: The vector contains the fundamental amplitude of 0.76 and phase angle of 12° obtained from real-time analysis, mapped to three-dimensional spatial coordinates (0.76, 4.8%, 12°). The correlation operation between this coordinate and the matrix adopts a spatial distance weighting method: the distance from the coordinate point to the center of each level of the matrix is ​​calculated, and the closer the distance, the higher the weight. The final product operation outputs a conflict risk value of 0.63.

[0108] When the risk value exceeds the safety threshold of 0.6, the excitation component redistribution strategy is triggered. The system reduces the setpoint of the electric excitation current according to preset rules: the original target value is reduced by a fixed percentage, while the permanent magnet flux compensation is increased by the same percentage. The redistribution process uses a ramp function transition, and the transition time is dynamically set according to the rotational speed. In the example, the current setpoint is reduced from the original value, and the compensation in the S-pole region of the permanent magnet is increased accordingly. After the new command after redistribution is verified for safety, it is issued. The system simultaneously marks this event as a new type of conflict case and updates the probability value of the corresponding segment in the coupling relationship matrix.

[0109] The model's dynamic adjustment and conflict detection form a closed loop. After each coordinate scaling, the system automatically initiates baseline calibration of the conflict matrix: recalculating the historical data distribution of each tier interval and updating the probability weight coefficients. After a conflict event is resolved, the actual running data is used as new samples to input into the digital twin model, initiating a new round of residual calculation and model calibration.

[0110] Example 5: The closed-loop execution process of the field weakening control system uses the real-time field weakening demand correlation factor as the control core. This factor is continuously calculated and output by the monitoring module, and its value changes continuously between zero and one. The system maintains a dynamic loading queue to manage field weakening control commands. The queue adopts a priority sorting mechanism: the change between the current value of the correlation factor and the value at the time of the previous command generation is used as the basis for priority determination; the larger the change, the higher the priority. High-priority commands can interrupt the loading process of low-priority commands that are currently being executed. Command loading adopts a two-stage mode: in the first stage, the set value of the electric excitation current is transmitted to the power drive unit through a high-speed data bus; in the second stage, the permanent magnet flux compensation vector is distributed to each area controller in a time-division multiplexing manner. The loading interval is adaptively adjusted according to the motor speed; the higher the speed, the shorter the interval.

[0111] The data acquisition system operates according to a fixed cycle, which is set to a constant value. At the beginning of each cycle, multi-source sensor data readings are triggered synchronously: the armature winding three-phase current acquires instantaneous values ​​through an isolated sampling circuit; the rotor position encoder outputs absolute angles; the stator core temperature sensor returns eight-point array data; and the air gap flux density probe uploads sixteen-channel waveforms. All data is verified, packaged into data frames, timestamped, and stored in a circular buffer. The buffer uses a paged storage structure, and a digital twin model update event is triggered every certain number of data frames.

[0112] The model update employs an incremental learning strategy. Each update extracts the latest data frames from the buffer and fuses them with the current model state. The electromagnetic field calculation module focuses on updating the permeability parameter: inferring the local saturation characteristics of the core material based on measured magnetic flux density. The heat conduction analysis module corrects the cooling medium flow rate parameter: adjusting the heat network branch conduction coefficients according to changes in temperature gradient distribution. The mechanical dynamics module calibrates the bearing friction model based on rotor position angle data. After parameter updates, the three-dimensional electromagnetic field state space is recalculated: newly added feature points are imported into the spatial structure, and density distribution scanning is re-executed. The spatial scanning uses a local update algorithm, processing only the sub-cubic region centered on the newly added point. When a newly added point causes the density of a region to exceed a threshold, a new anomalous region boundary is created; if it causes the density of an existing region to fall below the threshold, the region is deleted. The updated excitation anomaly feature index set is synchronized to all processing nodes via shared memory.

[0113] The convergence determination mechanism is based on the changing trend of the weak magnetic field demand-related factors. The system maintains a three-unit ring register to store the factor values ​​of the latest three iterations. Each time a new factor is generated, a triple check is performed: checking whether the factor value is within the valid range; checking whether the time interval between two adjacent data collections meets the periodic requirement; and checking whether the data source has been digitally signed and authenticated. After passing the checks, the rate of change sequence is calculated: first, the absolute difference between adjacent factor values ​​is calculated, and then the ratio of this difference to the time interval is calculated. When all three consecutive ratios are lower than the preset convergence threshold, the system activates the control loop to terminate the program.

[0114] The loop termination process is executed in steps: a freeze command is sent to the power drive unit to stop receiving new control commands; the manifold neural operator enters read-only mode, pausing parameter updates; the multi-source sensors switch to low-power sampling mode, reducing the data acquisition frequency. The system maintains the output state of the last valid field weakening control command: the power drive unit continuously tracks the setpoint of the electric excitation current, and the area controller maintains the current permanent magnet flux compensation distribution. During the freeze period, the field weakening demand correlation factor is continuously monitored; if its rate of change exceeds the threshold set ratio again, the control loop is immediately restarted. The restart process performs initialization: clearing historical data in the ring register; resetting the incremental learning parameters of the digital twin model; and the power drive unit switches to dynamic response mode.

[0115] An anomaly handling mechanism is implemented throughout the entire closed-loop process. When a data acquisition cycle times out, the system automatically extends the validity period of the current control command until new data is available. If an anomaly is detected in the manifold neural operator output, the system immediately switches to a backup control strategy: generating control commands using historically optimal parameters. When convergence determination fails multiple times consecutively, a fault tracing process is initiated: exporting complete data snapshots from the most recent control cycles, marking the timestamps and execution status of each processing stage. Fault data is uploaded to the diagnostic server via a separate channel, while the system itself executes a safety shutdown procedure, resuming operation only after manual intervention.

[0116] The loop status visualization interface displays key parameters in real time: the numerical curve of the magnetic weakening demand correlation factor, the distribution map of abnormal regions in the three-dimensional electromagnetic field state space, and the currently loaded control command parameters. Operators can manually adjust the convergence threshold parameter and intervene in the termination timing of the control loop. All interactive operations are logged in detail, including operator identity, timestamp, parameter values ​​before modification, and parameter values ​​after modification. The log file uses a cyclic overwrite storage strategy, retaining complete operation records of the most recent control loops. The system runtime statistics function records the percentage of working time for each submodule, providing data support for resource optimization.

[0117] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0118] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A field weakening control system for an axial flux hybrid excitation synchronous motor based on multi-source feedback, characterized in that, Includes a central processing module, which is used to perform: A digital twin model of an axial flux hybrid excitation synchronous motor is constructed. Real-time operating data is collected by multi-source sensors deployed on the motor body, and the real-time operating data is synchronized to the digital twin model. The digital twin model topology includes an electromagnetic field calculation module, a heat conduction analysis module, and a mechanical dynamics module; A three-dimensional electromagnetic field state space is established based on the digital twin model, and motor operation characteristic parameters are extracted through manifold neural operators to generate a multi-dimensional excitation state feature set. Based on the distribution density of the multi-dimensional excitation state feature set in the three-dimensional electromagnetic field state space, the excitation anomaly feature index set is determined; Real-time acquisition of motor current, voltage, speed and temperature data; extraction of current excitation characteristic parameters and mapping to the three-dimensional electromagnetic field state space; calculation of the spatial correlation between the current excitation characteristic parameters and the excitation anomaly characteristic index set; output of field weakening demand correlation factor. Based on the field weakening demand correlation factor, the multi-dimensional excitation state feature set, and real-time operating data, the manifold neural operator parameters are iteratively updated to generate field weakening control commands and dynamically adjust the coupling ratio between the electric excitation component and the permanent magnet excitation component.

2. The field weakening control system for an axial flux hybrid excitation synchronous motor based on multi-source feedback according to claim 1, characterized in that, The real-time operating data collected by the multi-source sensors includes armature winding current waveform, air gap flux density distribution, rotor position angle, stator core temperature gradient, and cooling medium flow rate; the real-time operating data is subjected to time-domain noise reduction and frequency-domain filtering to generate a motor operating health dataset.

3. The field weakening control system for an axial flux hybrid excitation synchronous motor based on multi-source feedback according to claim 2, characterized in that, The manifold neural operator includes a Laplacian kernel integral module for feature decomposition of the motor operating health dataset: Discretize the geometry of the motor design domain into a finite element mesh and solve for the Laplace operator characteristic function of the mesh nodes; The armature current waveform and air gap flux density are spectrally decomposed using the Laplace operator characteristic function to extract the fundamental component amplitude, harmonic distortion rate and spatial phase offset, forming the multi-dimensional excitation state feature set.

4. The field weakening control system for an axial flux hybrid excitation synchronous motor based on multi-source feedback according to claim 3, characterized in that, The set of excitation anomaly characteristic indicators is determined through the following steps: The coordinate axes of the three-dimensional electromagnetic field state space are defined based on the parameter dimensions of the multi-dimensional excitation state feature set. Calculate the cluster density of characteristic points of historical field weakening fault events of the motor in the three-dimensional electromagnetic field state space; The excitation characteristic parameters corresponding to spatial regions where the density of feature points exceeds a preset threshold are selected to form the excitation anomaly characteristic index set.

5. The field weakening control system for an axial flux hybrid excitation synchronous motor based on multi-source feedback according to claim 4, characterized in that, The calculation of the magnetic field weakening demand correlation factor includes: The amplitude and phase angle of the current fundamental flux linkage are analyzed based on the real-time current waveform and rotor position angle. The current fundamental flux linkage amplitude and phase angle are combined into a dynamic field strength vector and mapped to the three-dimensional electromagnetic field state space; Calculate the weighted sum of the Euclidean distances between the dynamic field strength vector and each feature point in the excitation anomaly feature index set, and output the field weakening demand correlation factor after normalization.

6. The field weakening control system for an axial flux hybrid excitation synchronous motor based on multi-source feedback according to claim 5, characterized in that, The generation of the field weakening control command includes: The weakening magnetic demand correlation factor and the multi-dimensional excitation state feature set are input into the updated manifold neural operator; The encoder-approximator-decoder structure of the Laplace kernel integral module is used to predict the set value of the electric excitation winding current and the equivalent magnetic flux compensation of the permanent magnet. The field weakening control command is generated based on the ratio between the set value of the electric excitation winding current and the equivalent magnetic flux compensation of the permanent magnet.

7. The field weakening control system for an axial flux hybrid excitation synchronous motor based on multi-source feedback according to claim 6, characterized in that, The iterative update of the manifold neural operator uses the Fourier descriptor algorithm: Convert the motor operating health dataset into a frequency domain feature tensor; The low-rank core matrix of the frequency domain feature tensor is extracted by tensor network decomposition. The weight coefficients of the Laplace kernel integral module of the manifold neural operator are corrected using the low-rank core matrix.

8. The field weakening control system for an axial flux hybrid excitation synchronous motor based on multi-source feedback according to claim 7, characterized in that, The dynamic adjustment of the digital twin model includes: After executing the field weakening control command, the actual air gap magnetic flux density change rate and temperature gradient distribution were collected; Compare the residuals between the actual data and the predicted values ​​of the digital twin model to generate model correction coefficients; The coordinate scale of the three-dimensional electromagnetic field state space is scaled according to the model correction coefficients, and the spatial boundary of the excitation anomaly characteristic index set is updated.

9. The field weakening control system for an axial flux hybrid excitation synchronous motor based on multi-source feedback according to claim 8, characterized in that, It also includes conflict detection for the field weakening control strategy: Establish the coupling relationship matrix between the electrical excitation response delay time and the permanent magnet temperature demagnetization coefficient; The control conflict risk value is calculated based on the product of the dynamic field strength vector and the coupling relationship matrix. When the control conflict risk value exceeds the safety threshold, a redistribution strategy for the excitation component ratio is triggered.

10. The field weakening control system for an axial flux hybrid excitation synchronous motor based on multi-source feedback according to claim 9, characterized in that, The closed-loop execution process of the field weakening control system is as follows: Dynamically load the field weakening control command based on the correlation factors related to the real-time field weakening demand. Feedback data is collected once every preset period to update the digital twin model parameters and excitation anomaly characteristic index set; The current control loop terminates when the rate of change of the weak magnetic field demand correlation factor is less than the convergence threshold in three consecutive iterations.

Images