Permanent magnet motor fault imaging diagnosis method and system based on magnetic field inverse inversion

Abstract

This invention discloses a method and system for fault imaging diagnosis of permanent magnet motors based on magnetic field inversion. The method includes: constructing a material permeability distribution map and an excitation source current density distribution map of the motor, inputting these maps into a digital twin of the motor, and predicting the magnetic field vector; wherein the digital twin is constructed by training and freezing parameters of a physically guided differentiable neural network; subtracting the measured values ​​to obtain a residual vector, and initiating fault inversion when the residual is significant; iteratively adding perturbations to the input image, and using the digital twin to calculate the gradient of the loss function relative to the added perturbations to update the added perturbations, until a match between the reconstructed magnetic field vector and the measured values ​​is found, and the added perturbation corresponds to the fault imaging map; and determining the fault location and severity based on the fault imaging map. This invention enables high-precision and interpretable diagnosis of permanent magnet motor faults.

Description

Technical Field

[0001] This invention belongs to the field of industrial equipment fault diagnosis technology, specifically relating to a fault imaging diagnosis method and system for permanent magnet motors based on magnetic field inverse inversion. It is applicable to the non-invasive, online, high-precision positioning and quantitative diagnosis of early electrical and mechanical faults in permanent magnet synchronous motors in harsh industrial environments such as mining and metallurgy. Background Technology

[0002] Permanent magnet synchronous motors, with their high power density and high efficiency, are widely used in critical power systems such as mine hoisting and transportation. However, the complex mine environment, coupled with the motor's long-term exposure to heavy loads, frequent starts and stops, and stress from dust and moisture, makes it prone to early failures such as stator inter-turn short circuits, localized demagnetization of permanent magnets, and rotor dynamic eccentricity. If these failures are not detected and addressed in a timely manner, they will lead to motor performance degradation and reduced energy efficiency, and in severe cases, may cause equipment damage or even safety accidents, resulting in huge economic losses.

[0003] Existing technologies (such as CN202510039973) typically employ deep learning models like the Transformer for fault classification. However, these methods belong to data-driven supervised learning, requiring the pre-collection of massive amounts of labeled fault samples (such as rotor bar breakage and inter-turn short circuit data). In actual mining production, obtaining destructive fault samples of high-value equipment is virtually impossible, rendering these models ineffective in zero-shot scenarios. Another type of technology (such as CN202510362434), while introducing large text models for auxiliary diagnosis, is essentially still based on text retrieval and generation (RAG) from a historical case database, lacking the ability to analyze the physical mechanisms of the electromagnetic fields inside the motor. They can only provide a qualitative judgment of 'potentially causing a fault,' failing to achieve quantitative inversion and visualization of fault physical parameters (such as permeability distribution and short-circuit current density).

[0004] Faults in permanent magnet synchronous motors typically manifest as abnormal changes in electromagnetic field distribution, vibration spectrum, and temperature rise patterns. Fault diagnosis can be performed using signal processing or data-driven methods. Sensors collect signals such as current, voltage, vibration, and magnetic fields, extracting time-domain, frequency-domain, or time-frequency-domain features, and constructing machine learning or deep learning models to achieve fault identification and classification. However, existing diagnostic methods generally suffer from insufficient sensitivity to early, subtle fault characteristics, poor real-time performance under varying loads and strong interference environments, and high false positive rates due to fluctuations in operating conditions. Especially for permanent magnet motors in harsh environments such as mining, traditional methods struggle to simultaneously achieve non-invasiveness, high-precision positioning, and strong anti-interference capabilities. Summary of the Invention

[0005] This invention provides a method and system for imaging diagnosis of permanent magnet motor faults based on magnetic field inversion, enabling high-precision and interpretable diagnosis of permanent magnet motor faults.

[0006] To achieve the above technical objectives, the present invention adopts the following technical solution: A fault imaging diagnosis method for permanent magnet motors based on magnetic field inversion includes: Obtain parameter vectors to comprehensively characterize the motor's operation and health status, and construct a material permeability distribution map of the motor. and excitation source current density distribution diagram ; Distribution map and Input the digital twin of the motor to predict the output magnetic field vector The digital twin is constructed by training and freezing the parameters of a physically guided differentiable neural network. Predicted value of magnetic field vector The residual vector is obtained by subtracting the measured magnetic field vector from the residual vector. ; Determine whether the residual vector is significant based on a given threshold, and initiate fault inversion if the residual is significant; Iteratively towards the distribution map and Add a perturbation and use the digital twin to calculate the gradient of the loss function with respect to the added perturbation. Then use the gradient to update the added perturbation until the reconstructed predicted magnetic field vector matches the measured magnetic field vector. The added perturbation is the fault imaging map. The fault location and severity are determined based on the fault imaging diagram.

[0007] Furthermore, the parameter vector includes the motor's basic parameters, real-time operating parameters, and fault identification parameters. The basic parameters include rated power, rated speed, number of pole pairs, number of stator slots, winding connection method, remanence of permanent magnet material, stator inner diameter, air gap length, and core length. The real-time operating parameters include speed and q-axis current.

[0008] Furthermore, when constructing the distribution map, a two-dimensional axisymmetric computational domain including the stator, rotor, permanent magnet and air gap is constructed with the motor axis as the axis of symmetry, and it is discretized into a regular grid with a fixed resolution. Then, a physical property map with two channels is generated for each grid. The first channel is a material permeability distribution map: based on the material corresponding to each grid position, a corresponding relative permeability scalar value is assigned, forming a static material permeability distribution map. ;in, Represents the radial coordinates in a two-dimensional axisymmetric cylindrical coordinate system. This represents the axial coordinate in a two-dimensional axisymmetric cylindrical coordinate system.

[0009] The second channel represents the excitation source current density: based on the rotational speed, The shaft current and winding distribution are encoded with normalized current density values ​​at the corresponding grid positions, forming a dynamic excitation source current density distribution map. .

[0010] Furthermore, the digital twin of the motor employs a convolutional neural network with an encoder-decoder structure, and the input is a dual-channel physical field image. and The output is a magnetic vector potential distribution map; the end of the convolutional neural network is connected to a post-processing layer, which converts the magnetic vector potential distribution map into a magnetic field vector at the sensor location through curl calculation and coordinate mapping.

[0011] Furthermore, when training a physically guided differentiable neural network, the loss function adopts the weak form of the Maxwell equations for static magnetic fields in axisymmetric cylindrical coordinates, i.e., minimizing the co-energy of the coupled fields; the specific loss function is as follows:

[0012] ; in, This is the physical loss function used for network training; To calculate the number of sampling points within the domain; the first term The second term represents the magnetic field energy storage density. This represents the work done by the current source; For the network in the first sampling points The azimuth component of the predicted magnetic vector potential. The radial coordinates are in an axisymmetric coordinate system. , The input images are respectively and Mapped to the The material permeability and current density azimuth components at each sampling point; The boundary constraint term loss is used to force the calculation of the magnetic vector potential at the outer boundary of the domain. It is zero; The weight coefficients for the boundary penalty term.

[0013] Furthermore, the magnetic vector potential of the network output is calculated using a finite-difference convolution operator based on the Yee mesh. Spatial coordinates The partial derivatives of the loss function are calculated and substituted into the loss function formula to calculate the loss value. Then, the loss value is minimized using the gradient descent algorithm, and all weight parameters of the network are updated using the backpropagation algorithm.

[0014] Furthermore, the gradient of the inversion loss function with respect to the added perturbation is calculated using a digital twin, and is expressed as: ; ; In the formula, The fault inversion loss function is used to characterize the difference between the magnetic field vector predicted by the digital twin and the measured magnetic field vector; the fault inversion loss function is calculated... Relative to disturbance gradient This yields the optimized direction for updating the perturbation; Represents a digital twin; This represents the actual dual-channel physical property diagram of the motor, derived from the actual material permeability distribution diagram of the motor. and excitation source current density distribution diagram constitute; This indicates the added perturbation; express The measured magnetic field vector at time t.

[0015] Furthermore, the added perturbation is updated using gradients and an adaptive moment estimation optimizer, and in each iteration... The update rule for the added perturbation is as follows: ; in, For learning rate, gradient In the The first-order moment estimate after bias correction in the next iteration is used to characterize the exponentially weighted average of the gradient of the fault inversion loss function; gradient In the The second-order moment estimate after bias correction in the next iteration is used to characterize the exponentially weighted average of the squared gradient; The table contains numerically stable terms to avoid the denominator being zero.

[0016] Furthermore, based on the fault imaging map, the fault location and severity are determined, specifically: Based on the physical disturbance corresponding to the material's magnetic permeability channel The spatial distribution forms a material permeability fault imaging map. When the absolute value of the disturbance at a certain grid position in the material permeability fault imaging map is greater than a preset material permeability disturbance threshold... When this occurs, the grid location is determined as a region of abnormal magnetic properties of the material, and the degree of abnormal magnetic properties is determined based on the absolute value of the disturbance at the grid location. Based on the physical disturbance corresponding to the excitation source current density channel The spatial distribution forms a current density fault imaging map. When the absolute value of the disturbance at a certain grid position in the current density fault imaging map is greater than a preset current density disturbance threshold... When this occurs, the grid location is identified as a region of abnormal current density, and the degree of current anomaly is determined based on the absolute value of the disturbance at the grid location. The spatial coordinates of the abnormal area in the fault imaging map are mapped to the physical model of the motor to determine the specific location of the fault.

[0017] A fault imaging diagnostic system for permanent magnet motors based on magnetic field inverse inversion includes: The signal acquisition module is used to: acquire signals deployed on the casing of the permanent magnet motor. Using signals from a triaxial magnetic field sensor, a measured magnetic field vector is constructed; The physics field image construction module is used to: construct a material permeability distribution map of the motor by acquiring parameter vectors that comprehensively characterize the motor's operation and health status. and excitation source current density distribution diagram ; The motor digital twin module, constructed by training and freezing parameters of a physically guided differentiable neural network, is used to: construct a distribution map and As input, predict the output magnetic field vector. ; The residual calculation and decision module is used to: calculate the predicted value of the magnetic field vector. The residual vector is obtained by subtracting the measured magnetic field vector from the residual vector. And determine whether the residual vector is significant based on a given threshold, and initiate fault inversion when the residual is significant; The inverse imaging module is used to: iteratively map the distribution map and Add a perturbation and use the digital twin to calculate the gradient of the loss function with respect to the added perturbation. Then use the gradient to update the added perturbation until the reconstructed predicted magnetic field vector matches the measured magnetic field vector. The added perturbation is the fault imaging map. The fault determination module is used to locate and determine the severity of faults based on fault imaging images.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention improves the physical interpretability of fault diagnosis results by constructing a dual-channel physical field input of material permeability distribution and excitation source current density distribution, and by using a physically guided differentiable digital twin to establish the correspondence between the internal physical state of the motor and the external magnetic field response.

[0019] Furthermore, by comparing the difference between the predicted magnetic field and the measured magnetic field of the digital twin, and by performing gradient inverse optimization on the input physical field disturbance, the predicted magnetic field gradually approaches the measured magnetic field, thereby realizing the inverse inversion from external magnetic field anomaly to changes in the internal physical state of the motor.

[0020] Finally, based on the spatial distribution of material permeability perturbation and excitation source current density perturbation obtained by reverse optimization, fault imaging results are formed, thereby realizing fault type judgment, fault location determination and anomaly degree characterization. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method described in the embodiments of this application; Figure 2 This is a schematic diagram of the architecture and training principle of the physically guided neural network used in the digital twin of the motor described in this application embodiment. Detailed Implementation

[0022] The embodiments of the present invention will be described in detail below. These embodiments are based on the technical solutions of the present invention and provide detailed implementation methods and specific operation processes to further explain the technical solutions of the present invention.

[0023] Example 1

[0024] This embodiment provides a method for imaging diagnosis of permanent magnet motor faults based on magnetic field inverse inversion, such as... Figure 1 , Figure 2 As shown, it includes the following steps: Step 1: Obtain the parameter vector used to comprehensively characterize the motor's operation and health status, and construct the material permeability distribution map of the motor. and excitation source current density distribution diagram .

[0025] In this embodiment, a parameter vector is defined to comprehensively characterize the motor's operating and health status. Including the basic parameters of the motor Real-time operating parameters and fault identification parameters The basic parameters include rated power. Rated speed Extreme logarithms Number of stator slots Winding connection method, remanence of permanent magnet material and the main geometric dimensions of the motor (such as the stator inner diameter) Air gap length Core length Real-time operating parameters include speed. and q-axis current The q-axis current Through torque command The result is obtained through conversion.

[0026] When constructing the distribution map, a two-dimensional axisymmetric computational domain including the stator, rotor, permanent magnet, and air gap is constructed with the motor axis as the axis of symmetry. This domain is then discretized into a regular grid with a fixed resolution, and a physical property map with two channels is generated for each grid.

[0027] The first channel is a material permeability distribution map: based on the material (silicon steel sheet, permanent magnet, copper winding, air) at each grid location, a corresponding relative permeability scalar value is assigned, forming a static material permeability distribution map. ;in, Represents the radial coordinates in a two-dimensional axisymmetric cylindrical coordinate system. This represents the axial coordinate in a two-dimensional axisymmetric cylindrical coordinate system.

[0028] The second channel represents the excitation source current density: based on the rotational speed, The shaft current and winding distribution are encoded with normalized current density values ​​at the corresponding grid positions, forming a dynamic excitation source current density distribution map. The equivalent magnetizing current of the permanent magnet is also characterized in this channel.

[0029] Material permeability distribution diagram of the two channels With excitation source current density distribution diagram Combined with health status indicators ( Real-time generation of corresponding dual-channel physical field input images This refers to the reference electromagnetic configuration corresponding to the fault-free state of the motor.

[0030] The healthy state refers to the normal operating state of the motor when no preset fault occurs. The dual-channel physical field input image corresponding to the healthy state serves as the initial reference for fault inverse inversion. When the motor fails, physical disturbances are added to the dual-channel physical field input image corresponding to the healthy state, and the physical disturbances are inversely optimized to determine the changes in the internal physical state of the motor relative to the healthy state.

[0031] Step 2, distribute the map and Input the digital twin of the motor to predict the output magnetic field vector The digital twin is constructed by training and freezing the parameters of a physically guided differentiable neural network.

[0032] In this embodiment, the digital twin of the motor adopts a convolutional neural network with an encoder-decoder structure, and the input is a dual-channel physical field image. and The output is a magnetic vector potential distribution map. This neural network is specifically designed for end-to-end mapping of physical field images: its contraction path (encoder) consists of multiple stages, each containing two 3×3 convolutional layers. An activation function and a 2×2 max-pooling layer are used to progressively extract multi-scale features from the input physical field image. The dilation path (decoder) upsamples the image through transposed convolution and makes skip connections with the feature maps of the corresponding layers in the encoder to fuse low-level fine structural information with high-level physical semantic information. Finally, a 1×1 convolutional layer outputs a magnetic vector potential distribution map with the same resolution as the input. .

[0033] The "digital twin" described in this invention differs fundamentally from traditional data-driven deep learning models (black boxes). Traditional models work by fitting statistical laws of input and output data, and their internal parameters lack physical meaning. In contrast, the Physics-Informed Neural Network (PINN) used in this invention is essentially a differentiable solver that embeds physical laws.

[0034] First, its training does not rely on labeled data, and the loss function... This directly corresponds to the energy minimization principle (weak form) of Maxwell's equations under an axisymmetric static magnetic field. This principle forces the distribution of the magnetic vector potential output by the network. Strictly adhering to the fundamental laws of electromagnetism, the network becomes a physically consistent field solver, rather than a simple data fitter; secondly, because the network structure maintains the consistency with the physical source ( and Distribution) to observation field ( The continuous differentiable mapping of the distribution enables this system to solve electromagnetic inverse problems using automatic differentiation techniques. During diagnosis, by freezing network weights and optimizing the input image, the material properties or physical disturbances of the excitation source can be directly deduced from the magnetic field anomaly, thus realizing the location and quantification of the fault. This inversion capability based on the differentiable physical model is something that traditional black-box models cannot achieve, and it is also the core foundation for this invention to achieve high-precision and interpretable diagnosis.

[0035] To obtain the sensor magnetic field required for diagnosis, a deterministic post-processing layer is connected to the end of the convolutional neural network. This layer uses a pre-defined finite-difference convolution kernel to process the network output. The image is curled to obtain the magnetic flux density. The distribution. Then, based on the pre-calibrated sensor spatial coordinates... From that The triaxial magnetic field components at corresponding locations are extracted from the field distribution to form the final predicted output vector. .

[0036] To embed the physical laws of electromagnetic fields into the training process and fundamentally eliminate the reliance on large amounts of pre-simulation data, this invention employs an unsupervised learning strategy. Based on the previously defined two-dimensional axisymmetric computational domain, the loss function must adopt the weak form of Maxwell's equations for the static magnetic field in axisymmetric cylindrical coordinates, i.e., minimizing the co-energy of the coupled field. This form effectively handles the permeability discontinuity problem at the iron core-air interface, improving solution stability. The specific loss function is as follows: ; in, For physical loss function; To calculate the number of sampling points within the domain; the first term The second term represents the magnetic field energy storage density. This represents the work done by the current source; For the network in the first sampling points The azimuth component of the predicted magnetic vector potential. The radial coordinates are in an axisymmetric coordinate system. , The input images are respectively and Mapped to the The material permeability and current density azimuth components at each sampling point. Specifically, to ensure the well-posedness of the equation solution, the following is introduced: The boundary constraint term loss is used to force the calculation of the magnetic vector potential at the outer boundary of the domain. The value is zero, simulating the physical boundary condition where the magnetic field decays to zero at infinity, ensuring that the magnetic field is confined within the computational domain and does not undergo non-physical divergence. The weight coefficients for the boundary penalty term.

[0037] During neural network training, there is no need to pre-calculate magnetic field labels using high-fidelity finite element simulation. The system performs large-scale random sampling within the motor's expected operating range and preset fault parameter space, directly generating massive amounts of data containing only physical properties (…). and The input image samples are distributed as follows. These samples are fed into the network, and the loss function described above is minimized. This drives the network weight updates. This training method, directly driven by physical laws, enables the network to autonomously learn and incorporate the fundamental laws of electromagnetic fields. The specific training steps are as follows:

[0038] In the comprehensive state parameter space of the motor Large-scale random sampling is performed, and a set of parameters is obtained for each sampling. The corresponding dual-channel input image (material permeability distribution map and excitation source current density distribution map) is generated by physical field visualization.

[0039] The sampled input image is fed into the Conv PINN network, and the network outputs the magnetic vector potential of each pixel in the computational domain. Predicted values. A physical guidance loss function is constructed based on the weak-form energy functional of the axisymmetric static magnetic field. The magnetic vector potential output by the network is calculated using a finite-difference convolution operator based on a Yee grid. Spatial coordinates The partial derivatives are used to calculate the loss value by substituting them into the loss function formula.

[0040] The loss function is minimized using a gradient descent algorithm (such as the Adam optimizer), and all weight parameters of the Conv PINN network are updated using the backpropagation algorithm.

[0041] Repeat the above steps until the loss function converges or the preset number of iterations is reached. After training, the network weights are fixed and can be used as a digital twin of the motor for rapid magnetic field prediction in the online diagnosis stage. This replaces the traditional method of relying on table lookup and interpolation based on pre-computed databases, providing an immediate and accurate theoretical benchmark for online fault diagnosis.

[0042] At this point, the real-time health status can be input into the image. The input is fed into a trained physical guidance neural network model, which is essentially a digital twin of the input motor. This model outputs, via forward propagation, the theoretical magnetic field prediction value corresponding to the motor's fault-free state under the current real-time operating conditions, serving as the health baseline magnetic field for fault diagnosis. The model's final output is the predicted three-axis health magnetic field vector at the sensor array location. .

[0043] Step 3, predict the magnetic field vector value The residual vector is obtained by subtracting the measured magnetic field vector from the residual vector. .

[0044] The method for obtaining the measured magnetic field vector in this embodiment is as follows: A magnetic field vector array containing... is deployed around the motor casing. A static array of three-axis magnetic field sensors; raw magnetic field signals are acquired through the three-axis magnetic field sensor array. For the original magnetic field signal Bandpass filtering is performed to remove high-frequency noise and extremely low-frequency drift; and environmental background magnetic field interference is canceled out using active compensation units or reference sensor data to obtain a clean measured magnetic field vector. .

[0045] Calculate the predicted value of the magnetic field vector obtained in step 2. With the measured magnetic field vector Real-time residual vector between : ; It mainly contains abnormal magnetic field components caused by early faults, which is the basis for subsequent fault feature extraction and precise imaging diagnosis.

[0046] Step 4: Determine whether the residual vector is significant based on the given threshold, and start fault inversion when the residual is significant.

[0047] When the residual vector The amplitude exceeds the preset threshold At that time, that is The system triggers a fault inversion process. Threshold Specifically, the setting can be adaptively adjusted based on the noise level of the motor during normal operation (e.g., 3 times the standard deviation of the residual).

[0048] Step 5, iteratively map the distribution graph and A perturbation is added, and the gradient of the loss function with respect to the added perturbation is calculated using a digital twin. The added perturbation is then updated using the gradient until the reconstructed predicted magnetic field vector matches the measured magnetic field vector. The added perturbation is then used to obtain the fault imaging map.

[0049] This embodiment proposes and employs a physically guided alternating optimization strategy to perturb the input image. According to its physical meaning, it can be decomposed into a perturbation of the material's magnetic permeability distribution. and the perturbation of the excitation source current density distribution During optimization, the gradient is calculated using the automatic differentiation function of the neural network framework, and an adaptive moment estimator (Adam) is used for iterative updates. In practice, a step-by-step iterative alternating optimization strategy is employed to update the gradient. and The specific update rules are as follows: Real-time residual vector As core observation data, the weights of the trained Conv PINN model are fixed and treated as a deterministic, differentiable function. ,satisfy The goal of diagnosis is to find one (or a set) of input images that indicate a healthy state. Physical disturbance This makes the perturbed image After forward propagation through the network, its output magnetic field It can fit the measured magnetic field to the maximum extent. Therefore, this embodiment transforms the diagnosis into a constrained optimization problem: ; ; in, This is the fault inversion loss function. To efficiently solve this problem, this invention uses a fixed network... Under the given weights, its automatic differentiation capability is utilized to calculate the gradient of the fault inversion loss function with respect to the optimization variables (i.e., the pixel values ​​of the input image): ; Then, an adaptive moment estimation optimizer (Adam) is used for iterative updates. In each iteration... In this context, the update rule for the optimization variable is: ; in, The learning rate; gradient In the The first-order moment estimate after bias correction in the next iteration is used to characterize the exponentially weighted average of the gradient of the fault inversion loss function; gradient In the The second-order moment estimate after bias correction in the next iteration is used to characterize the exponentially weighted average of the squared gradient; The table contains numerical stability terms to avoid zero denominators. The deviation correction refers to the process of correcting the first and second moment estimates in the initial iteration stage, since these estimates are typically initialized to zero and thus tend to shift towards zero. This correction aims to improve the accuracy of parameter updates in the initial iteration stage.

[0050] This step 5 is performed iteratively until the convergence condition is met. ( (Assuming a preset threshold), the optimized perturbation image at this point This is the fault imaging image obtained from the final diagnosis.

[0051] This fault imaging map is actually a two-dimensional physical parameter disturbance distribution heat map that maps one-to-one with the physical structure of the motor (stator, rotor, permanent magnet). It can intuitively display the fault location (such as demagnetization area, short circuit point) and quantify the severity.

[0052] Step 6: Locate the fault and determine its severity based on the fault imaging diagram.

[0053] Based on the fault imaging map, the fault location and severity are determined, specifically: If the physical perturbation obtained through optimization is a permeability perturbation of the material. The amplitude of the permeability disturbance of the material is calculated, and the absolute value of the disturbance at each grid position is compared with the corresponding preset disturbance threshold; when At that time, the corresponding mesh is identified as a region of abnormal magnetic properties in the material; among which, The preset permeability perturbation threshold; If the physical disturbance obtained through optimization is the excitation source current density disturbance... The amplitude of the excitation source current density disturbance is calculated, and the absolute value of the disturbance at each grid position is compared with the corresponding preset disturbance threshold; when At that time, the corresponding grid is identified as the region of abnormal excitation source current density; among which, This is the preset excitation source current density perturbation threshold.

[0054] Based on the spatial coordinates of the abnormal region, the abnormal region is mapped to the corresponding position in the two-dimensional physical model of the motor to determine the location of the fault; and the degree of fault is determined based on the amplitude of the physical disturbance within the abnormal region.

[0055] When an abnormal region exceeding the disturbance threshold appears in the material permeability disturbance channel, it is determined that there is an abnormality in the material magnetic properties; when an abnormal region exceeding the disturbance threshold appears in the excitation source current density disturbance channel, it is determined that there is an abnormality in the winding current; the fault type is determined based on the physical field channel where the abnormality occurs.

[0056] Example 2

[0057] This embodiment provides a permanent magnet motor fault imaging diagnostic system based on magnetic field inverse inversion, including: The signal acquisition module is used to: acquire signals deployed on the casing of the permanent magnet motor. Using signals from a triaxial magnetic field sensor, a measured magnetic field vector is constructed; The physics field image construction module is used to: construct a material permeability distribution map of the motor by acquiring parameter vectors that comprehensively characterize the motor's operation and health status. and excitation source current density distribution diagram ; The motor digital twin module, constructed by training and freezing parameters of a physically guided differentiable neural network, is used to: construct a distribution map and As input, predict the output magnetic field vector. ; The residual calculation and decision module is used to: calculate the predicted value of the magnetic field vector. The residual vector is obtained by subtracting the measured magnetic field vector from the residual vector. And determine whether the residual vector is significant based on a given threshold, and initiate fault inversion when the residual is significant; The inverse imaging module is used to: iteratively map the distribution map and Add a perturbation and use the digital twin to calculate the gradient of the loss function with respect to the added perturbation. Then use the gradient to update the added perturbation until the reconstructed predicted magnetic field vector matches the measured magnetic field vector. The added perturbation is the fault imaging map. The fault determination module is used to locate and determine the severity of faults based on fault imaging images.

[0058] The implementation methods of each module in this embodiment are the same as those described in Embodiment 1.

[0059] The above embodiments are preferred embodiments of this application. Those skilled in the art can make various changes or improvements based on them. Without departing from the overall concept of this application, these changes or improvements should fall within the scope of protection claimed in this application.

Claims

1. A method for fault imaging diagnosis of permanent magnet motors based on magnetic field inverse inversion, characterized in that, include: Obtain parameter vectors to comprehensively characterize the motor's operation and health status, and construct a material permeability distribution map of the motor. and excitation source current density distribution diagram ; Distribution map and Input the digital twin of the motor to predict the output magnetic field vector The digital twin is constructed by training and freezing the parameters of a physically guided differentiable neural network. Predicted value of magnetic field vector The residual vector is obtained by subtracting the measured magnetic field vector from the residual vector. ; Determine whether the residual vector is significant based on a given threshold, and initiate fault inversion if the residual is significant; Iteratively towards the distribution map and Add a perturbation and use the digital twin to calculate the gradient of the loss function with respect to the added perturbation. Then use the gradient to update the added perturbation until the reconstructed predicted magnetic field vector matches the measured magnetic field vector. The added perturbation is the fault imaging map. The fault location and severity are determined based on the fault imaging diagram.

2. The permanent magnet motor fault imaging diagnosis method based on magnetic field inverse inversion according to claim 1, characterized in that, The parameter vector includes the motor's basic parameters, real-time operating parameters, and fault identification parameters. The basic parameters include rated power, rated speed, number of pole pairs, number of stator slots, winding connection method, remanence of permanent magnet material, stator inner diameter, air gap length, and core length. The real-time operating parameters include speed and q-axis current.

3. The permanent magnet motor fault imaging diagnosis method based on magnetic field inverse inversion according to claim 2, characterized in that, When constructing the distribution map, a two-dimensional axisymmetric computational domain including the stator, rotor, permanent magnet and air gap is constructed with the motor axis as the axis of symmetry. This domain is then discretized into a regular grid with a fixed resolution, and a physical property map with two channels is generated for each grid. The first channel is a material permeability distribution map: based on the material corresponding to each grid position, a corresponding relative permeability scalar value is assigned, forming a static material permeability distribution map. ;in, Represents the radial coordinates in a two-dimensional axisymmetric cylindrical coordinate system. Represents the axial coordinates in a two-dimensional axisymmetric cylindrical coordinate system; The second channel represents the excitation source current density: based on the rotational speed, The shaft current and winding distribution are encoded with normalized current density values ​​at the corresponding grid positions, forming a dynamic excitation source current density distribution map. .

4. The permanent magnet motor fault imaging diagnosis method based on magnetic field inverse inversion according to claim 1, characterized in that, The digital twin of the motor uses a convolutional neural network with an encoder-decoder structure, and the input is a dual-channel physical field image. and The output is a magnetic vector potential distribution map; the end of the convolutional neural network is connected to a post-processing layer, which converts the magnetic vector potential distribution map into a magnetic field vector at the sensor location through curl calculation and coordinate mapping.

5. The permanent magnet motor fault imaging diagnosis method based on magnetic field inverse inversion according to claim 1, characterized in that, When training a physically guided differentiable neural network, the loss function adopts the weak form of Maxwell's equations for static magnetic fields in axisymmetric cylindrical coordinates, i.e., minimizing the co-energy of the coupled fields; the specific loss function is as follows: ； in, This is the physical loss function used for network training; To calculate the number of sampling points within the domain; the first term The second term represents the magnetic field energy storage density. This represents the work done by the current source; For the network in the first sampling points The azimuth component of the predicted magnetic vector potential. The radial coordinates are in an axisymmetric coordinate system. , The input images are respectively and Mapped to the The material permeability and current density azimuth components at each sampling point Represents the radial coordinates in a two-dimensional axisymmetric cylindrical coordinate system. Represents the axial coordinates in a two-dimensional axisymmetric cylindrical coordinate system; The boundary constraint term loss is used to force the calculation of the magnetic vector potential at the outer boundary of the domain. It is zero; The weight coefficients for the boundary penalty term.

6. The permanent magnet motor fault imaging diagnosis method based on magnetic field inverse inversion according to claim 5, characterized in that, The magnetic vector potential of the network output is calculated using a finite-difference convolution operator based on Yee meshes. Spatial coordinates The partial derivatives of the loss function are calculated and substituted into the loss function formula to calculate the loss value. Then, the loss value is minimized using the gradient descent algorithm, and all weight parameters of the network are updated using the backpropagation algorithm.

7. The permanent magnet motor fault imaging diagnosis method based on magnetic field inverse inversion according to claim 1, characterized in that, The gradient of the inversion loss function with respect to the added perturbation is calculated using a digital twin and is expressed as: ； ； In the formula, The fault inversion loss function is used to characterize the difference between the magnetic field vector predicted by the digital twin and the measured magnetic field vector; the fault inversion loss function is calculated... Relative to disturbance gradient This yields the optimized direction for updating the perturbation; Represents a digital twin; This represents the actual dual-channel physical property diagram of the motor, derived from the actual material permeability distribution diagram of the motor. and excitation source current density distribution diagram constitute; This indicates the added perturbation; express The measured magnetic field vector at time t.

8. The permanent magnet motor fault imaging diagnosis method based on magnetic field inverse inversion according to claim 7, characterized in that, The added perturbation is updated using gradients and an adaptive moment estimation optimizer, and in each iteration... The update rule for the added perturbation is as follows: ； in, For learning rate, gradient In the The first-order moment estimate after bias correction in the next iteration is used to characterize the exponentially weighted average of the gradient of the fault inversion loss function; gradient In the The second-order moment estimate after bias correction in the next iteration is used to characterize the exponentially weighted average of the squared gradient; The table contains numerically stable terms to avoid the denominator being zero.

9. The permanent magnet motor fault imaging diagnosis method based on magnetic field inverse inversion according to claim 1, characterized in that, Based on the fault imaging map, the fault location and severity are determined, specifically: Based on the physical disturbance corresponding to the material's magnetic permeability channel The spatial distribution forms a material permeability fault imaging map. When the absolute value of the disturbance at a certain grid position in the material permeability fault imaging map is greater than a preset material permeability disturbance threshold... When this occurs, the grid location is determined as a region of abnormal magnetic properties of the material, and the degree of abnormal magnetic properties is determined based on the absolute value of the disturbance at the grid location. Based on the physical disturbance corresponding to the excitation source current density channel The spatial distribution forms a current density fault imaging map. When the absolute value of the disturbance at a certain grid position in the current density fault imaging map is greater than a preset current density disturbance threshold... When this occurs, the grid location is identified as a region of abnormal current density, and the degree of current anomaly is determined based on the absolute value of the disturbance at the grid location. The spatial coordinates of the abnormal area in the fault imaging map are mapped to the physical model of the motor to determine the specific location of the fault.

10. A permanent magnet motor fault imaging diagnostic system based on magnetic field inverse inversion, characterized in that, include: The signal acquisition module is used to: acquire signals deployed on the casing of the permanent magnet motor. Using signals from a triaxial magnetic field sensor, a measured magnetic field vector is constructed; The physics field image construction module is used to: construct a material permeability distribution map of the motor by acquiring parameter vectors that comprehensively characterize the motor's operation and health status. and excitation source current density distribution diagram ; The motor digital twin module, constructed by training and freezing parameters of a physically guided differentiable neural network, is used to: construct a distribution map and As input, predict the output magnetic field vector. ; The residual calculation and decision module is used to: calculate the predicted value of the magnetic field vector. The residual vector is obtained by subtracting the measured magnetic field vector from the residual vector. And determine whether the residual vector is significant based on a given threshold, and initiate fault inversion when the residual is significant; The inverse imaging module is used to: iteratively map the distribution map and Add a perturbation and use the digital twin to calculate the gradient of the loss function with respect to the added perturbation. Then use the gradient to update the added perturbation until the reconstructed predicted magnetic field vector matches the measured magnetic field vector. The added perturbation is the fault imaging map. The fault determination module is used to locate and determine the severity of faults based on fault imaging images.

Images