Intelligent diagnosis method for composite fault of permanent magnet synchronous motor based on digital twinning

Abstract

The invention discloses an intelligent diagnosis method for composite faults of a permanent magnet synchronous motor based on digital twinning, relates to the technical field of fault diagnosis, and solves the technical problem of low diagnosis accuracy caused by the fact that fusion of multi-source information is not considered when a twinning model is adopted for fault diagnosis in the prior art. The method comprises the following steps: acquiring prophet data through a physical entity layer and transmitting the prophet data into a twin data layer; virtual current is generated through the virtual model layer and transmitted to the twinborn data layer; the twin data layer performs time-space synchronization on the received data to construct a virtual current signal reference library; the application layer makes a difference between predicted current output by the virtual model layer and current collected by the physical entity layer in real time, whether an electrical fault occurs or not is preliminarily judged in combination with a virtual current signal reference library, and corresponding fault features are input into a classification network according to a judgment result for fault diagnosis; the electromechanical composite fault diagnosis of the permanent magnet synchronous motor is realized by combining a multi-source information fusion technology.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to an intelligent diagnosis method for composite faults of permanent magnet synchronous motors based on digital twins. Background Art

[0002] Permanent magnet synchronous motors (PMSMs), with their high efficiency, high power density, and excellent dynamic performance, have become core power equipment in modern industry, widely used in CNC machine tools, electric vehicles, wind power generation, and industrial robots. However, due to their complex electromechanical coupling characteristics and the harsh operating environment of industrial sites (such as electromagnetic interference, sudden load changes, and mechanical wear), these motors are prone to electromagnetic faults (such as inter-turn short circuits and permanent magnet demagnetization) or mechanical faults (such as bearing wear and rotor eccentricity) during long-term operation, and may even cause electromagnetic-mechanical combined failures. If these faults are not diagnosed in a timely manner, they will not only lead to reduced equipment performance but may also trigger a chain reaction, paralyzing the production system and causing huge economic losses and safety risks.

[0003] Traditional fault diagnosis methods rely primarily on single signal source analysis, such as detecting electromagnetic anomalies through current signals or identifying mechanical defects using vibration signals. However, the high-noise environment of industrial sites (such as inverter switching noise and load fluctuation interference) can severely contaminate current signals, making feature extraction difficult; and while vibration signals are sensitive to mechanical faults, they can hardly reflect the characteristics of purely electrical faults. In addition, the coupling effect of electromagnetic and mechanical faults further confuses signal characteristics, making it difficult for single signal analysis methods to accurately distinguish fault types, resulting in a high misjudgment rate. For example, a rotor eccentricity fault may cause both current harmonics and vibration anomalies, but traditional methods often misjudge them as independent electrical or mechanical faults, resulting in the failure of maintenance strategies.

[0004] In recent years, the rise of digital twin technology has provided new insights for fault diagnosis in complex equipment. By constructing high-fidelity virtual models of physical entities, this technology enables real-time data interaction and dynamic simulation, effectively overcoming the limitations of actual sensor data. However, existing research, such as the Chinese patent "An Intelligent Diagnosis Method for Motor Faults Based on Digital Twins" (Patent Application No.: CN202510048247.9, Publication No.: CN119475103B), focuses on digital twin modeling of a single fault type or solely utilizes twin models for condition monitoring, failing to fully explore their potential for multi-source information fusion and coupled fault diagnosis. Summary of the Invention

[0005] In order to solve the problems existing in the above-mentioned prior art, the present invention provides an intelligent diagnosis method for composite faults of permanent magnet synchronous motors based on digital twins, which solves the technical problem that the prior art does not consider the fusion of multi-source information when using twin models for fault diagnosis, resulting in low diagnostic accuracy.

[0006] A digital twin-based intelligent diagnosis method for composite faults of permanent magnet synchronous motors, comprising:

[0007] Step 1: Establish a digital twin model of the permanent magnet synchronous motor, including a physical entity layer, a virtual model layer, a twin data layer, and an application layer. The virtual model layer includes a prediction model, and the application layer includes a classification model.

[0008] Step 2: Use the physical layer to collect prophetic data, real-time motor-related parameters and real-time current. The prophetic data includes current signals, vibration signals, speed signals and torque signals under different working conditions.

[0009] Step 3: Use the prophet data to create a data set to train the prediction model and classification model;

[0010] Step 4: The twin data layer extracts the difference characteristics between the virtual current generated by the prediction model and the actual current collected by the physical entity layer to form twin data, which is transmitted to the physical entity layer and the virtual model layer respectively, so that the relevant status of both can be updated in real time. The virtual current generated by the prediction model and the prophetic data collected by the physical entity layer are then synchronized in time and space to build a virtual current signal reference library containing iabc* waveforms under different working conditions.

[0011] Step 5: The predicted current is subtracted from the real-time current through the application layer, and then combined with the virtual current signal reference library to preliminarily determine whether an electrical fault has occurred. If an electrical fault has occurred, the fault characteristics are extracted based on the current signal and input into the classification network. If no electrical fault has occurred, the fault characteristics are extracted by fusing the current and vibration signals, and input into the classification network. Finally, the classification network performs fault diagnosis.

[0012] The physical entity layer inputs the motor's electrical parameters, such as stator flux, inductance, and phase resistance, into the virtual model layer and uses them as input signals to train and verify the current prediction model. At the same time, the virtual model layer can also map and feedback relevant fault information to the physical entity layer.

[0013] Furthermore, the configuration of the physical entity layer includes: deploying permanent magnet synchronous motor entities, high-precision sensors and data acquisition systems, collecting three-phase current, vibration signals, torque, etc., marking working condition labels (normal / fault type), and collecting the voltage u on the d and q axes in normal mode and fault mode under different working conditions. d 、u q The collected signals can be used for training the prediction model in the virtual model layer and subsequent fault diagnosis.

[0014] Furthermore, the virtual model layer is provided with a prediction model, which adopts an architecture design that integrates deep neural networks with the physical characteristics of the motor. The input parameters of the prediction model are motor-related parameters: one type is real-time operating parameters, including motor speed ω and electromagnetic torque Te; the other type is motor body parameters, including stator flux ψ d / ψ q 、Inductor L d / L q and stator resistance Rs.

[0015] Furthermore, the prediction model includes a convolution module, a bidirectional LSTM module and an output decoding layer. The convolution module adopts a three-level residual structure, and each level includes causal convolution, batch normalization and gated activation units;

[0016] The mathematical expression of causal convolution is:

[0017]

[0018] The calculation formula for batch normalization is:

[0019]

[0020] The activation function of the gated activation unit is:

[0021]

[0022] The bidirectional LSTM module consists of two independent LSTM layers, the forward and backward ones. The output decoding layer implements dimensionality transformation through a fully connected layer, and the final output uses the Tanh activation function:

[0023]

[0024] Furthermore, the twin data layer performs spatiotemporal synchronization on the received data, including: using a combination of dynamic time warping and Kalman filtering to achieve precise timing alignment of physical signals and virtual signals; for steady-state or slowly changing working conditions, an improved dynamic time warping algorithm is preferably used to find the optimal matching path by constructing a cost matrix containing timing constraints; for dynamic working conditions with sudden changes or high-speed operation, it switches to the Kalman filter compensation mode based on the motor motion equation, uses the rotor dynamics model to predict the signal evolution trend, and eliminates the phase deviation caused by speed fluctuations through the measurement update link, especially when the load suddenly changes, the synchronization error can be further compressed; the two synchronization strategies achieve smooth switching through an adaptive weight mechanism. When the speed change rate is detected to exceed the threshold, the transition from DTW to Kalman filtering is completed in a short time, and linear interpolation is used during the period to ensure timing continuity.

[0025] Furthermore, the extraction of differential features between actual current and virtual current through the residual attention network to form twin data includes enhanced processing through the residual attention network. The residual attention network adopts a cascade processing architecture. First, the time-frequency analysis module is used to perform multi-resolution analysis on the input signal. In the time-frequency transformation stage, the noise-free virtual current signal generated by the virtual model layer is used as an ideal reference benchmark. A short-time Fourier transform is performed by designing a Hanning window function with speed adaptive characteristics. The parameters of the Hanning window function are dynamically optimized according to the real-time speed to ensure that transient characteristics are accurately captured while maintaining frequency resolution.

[0026] Furthermore, during the time-frequency transformation process, the spectral characteristics of the virtual signal are encoded into the network weights as prior knowledge, guiding the model to automatically suppress noise components that do not conform to the electromagnetic laws of the motor during feature extraction. During the processing, the spatial attention module continuously compares the time-frequency distribution differences between the actual signal and the virtual signal, retains effective fault features through the residual learning mechanism, and filters out random interference. The output end of the residual attention network adopts a gated residual connection structure to adaptively fuse the enhanced features with the original signal.

[0027] Furthermore, the step 5 includes:

[0028] Step 5.1: Receive the predicted current signal i from the virtual model layer abc* The actual current signal i collected by the physical layer abc , generates the current residual Δi through real-time differential operation abc , and conduct preliminary screening of electrical faults based on preset threshold conditions;

[0029] Step 5.2: When it is detected that the statistical characteristics of Δiabc exceed the safety threshold, the electrical fault flag is immediately triggered, and the residual signal is directly input into the pre-trained classification network for diagnosis, and then the process proceeds to step 5.4. When it is detected that the statistical characteristics of Δiabc do not exceed the safety threshold, the process proceeds to step 5.3.

[0030] Step 5.3: If no obvious anomaly is detected, the multimodal feature fusion phase begins. The current residual signal is adaptively weighted and fused with the vibration signal processed through wavelet packet decomposition. The vibration signal focuses on capturing mechanical fault characteristics. The fusion process uses a dynamic weight adjustment strategy to automatically optimize the contribution ratio of current and vibration features based on the current signal-to-noise ratio and fault indicator status. This generates a highly discriminative composite feature vector that is input into the pre-trained classification network for diagnosis.

[0031] Step 5.4: The pre-trained classification network performs detailed diagnosis. The network output corresponds to the probability distribution of 10 typical fault types. Finally, the specific fault category is determined through the Softmax decision layer to achieve the purpose of mechatronic coupling fault diagnosis.

[0032] The beneficial effects of the present invention include:

[0033] 1. This invention innovatively proposes an intelligent fault diagnosis method based on a digital twin architecture, combined with multi-source information fusion technology, to realize electromechanical composite fault diagnosis of permanent magnet synchronous motors.

[0034] 2. Based on the established digital twin model of the permanent magnet synchronous motor, the predicted current output by the virtual model layer and the actual current output by the physical entity layer are synchronized in time and space. The noise-free characteristics of the digital twin model are used to enhance the actual signal characteristics, solving the problem that traditional methods are difficult to accurately extract current characteristics under strong noise conditions.

[0035] 3. Based on the residual between the predicted current and the actual current, the present invention preliminarily determines whether an electrical fault has occurred. It further designs a weighted adaptive fusion function to fuse the current residual signal with the vibration signal, which can achieve accurate diagnosis and identification of electromechanical composite faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Schematic diagram of the digital twin model of the permanent magnet synchronous motor involved in the embodiment of the present application.

[0037] Figure 2 This is a schematic diagram of the virtual model layer involved in an embodiment of the present application.

[0038] Figure 3 This is a flowchart of permanent magnet synchronous motor fault diagnosis based on digital twin technology involved in an embodiment of the present application.

[0039] Figure 4 This is a schematic diagram of a classification network involved in an embodiment of the present application. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of this application.

[0041] Example 1

[0042] The following is combined with Figure 1 The specific embodiments of the present invention are described in detail;

[0043] An intelligent diagnosis method for composite faults of permanent magnet synchronous motors based on digital twins. First, Figure 1 As shown in the figure, a digital twin model of a permanent magnet synchronous motor is established. The specific structure includes:

[0044] 1) Physical layer: mainly includes permanent magnet synchronous motor, motor control system and data acquisition device;

[0045] 2) Virtual model layer: mainly composed of the constructed neural network prediction model;

[0046] 3) Twin data layer: This layer aligns the real data output by the physical entity layer with the virtual data output by the virtual simulation layer in time and space. The noise-free nature of the digital twin model is leveraged to enhance the actual signal characteristics, enabling real-time interaction between the virtual model layer and the physical entity layer, thereby updating the twin model in real time.

[0047] 4) Application layer: The predicted current output by the virtual model layer is subtracted from the current collected in real time by the physical entity layer to preliminarily determine whether an electrical fault has occurred. The fault characteristics are then extracted by fusing the current and vibration signals and input into the classification network for fault diagnosis.

[0048] Specifically, the physical entity layer transmits the prophetic data used to train the digital twin model (including the current signal, vibration signal, speed signal, torque signal, etc. under the motor's health status) to the twin data layer;

[0049] The virtual model (twin model) layer transmits virtual data to the twin data layer. The twin data layer synchronizes the received data in time and space, eliminates feature misalignment caused by differences in sensor sampling frequencies, and uses the noise-free simulation characteristics of the digital twin model to enhance the actual signal characteristics. It also constructs a virtual current signal reference library (such as iabc* waveforms for normal and load sudden changes) as a comparison benchmark for the actual collected signals.

[0050] The residual attention network extracts the difference characteristics between the actual current and the virtual current to form twin data, which are transmitted to the physical entity layer and the virtual model layer respectively, so that the relevant status of the two layers can be updated in real time and the digital twin model can be continuously updated.

[0051] The physical entity layer inputs the motor's electrical parameters, such as stator flux, inductance, and phase resistance, into the virtual model layer and uses them as input signals for training and verification of the prediction model. At the same time, the virtual model layer can also map and feedback relevant fault information to the physical model layer.

[0052] Step 1: Physical layer configuration

[0053] Deploy permanent magnet synchronous motor entities, high-precision sensors, and data acquisition systems to collect three-phase current, vibration signals, torque, etc., and mark operating condition labels (normal / fault type).

[0054] By collecting the voltage u on the d and q axes in normal mode and fault mode under different working conditions d 、u q The signal and the electrical angular velocity signal of the motor rotor are used as prophetic data for training the prediction model in the virtual model layer.

[0055] At the same time, the virtual model can output the current at time t+1 in real time and transmit it to the twin data layer. A virtual current signal benchmark library is built in the twin data layer, which includes iabc* waveforms for normal / load mutation and other working conditions. By utilizing the noise-free simulation characteristics of the virtual model layer, a pure, ideal current waveform that conforms to the electromagnetic laws of the motor is generated as a comparison benchmark for the actual collected signal. Based on this, an early warning can be given as to whether the type of fault that will occur in the motor is an electrical fault.

[0056] First, the phase current signal i is collected in real time by a high-precision current sensor installed on the three-phase winding of the motor. a ,i b ,i c At the same time, the rotor position angle θ is measured by an absolute encoder, and the sampling frequency is not less than 10kHz to ensure the control bandwidth.

[0057] Then the three-phase current signal is converted into the stationary coordinate system by Clarke transformation. α ,i β , and then combine the real-time rotor position information to obtain the direct axis current i in the rotating coordinate system through Park transformation d and the quadrature axis current i q . The measured i d and i q With the target current i d * and i q * Compare and calculate the error and

[0058] Due to the cross-coupling effect of the motor, the electrical angular velocity ω must be considered when calculating the final voltage. e and motor parameters (such as inductance L d , L q and permanent magnet flux ψ f ). Therefore, the d-axis voltage u d The calculation formula is:

[0059] u d=K p,d ·e d +K i,d ∫e d dt-ωL q i q

[0060] q-axis voltage u q The calculation formula is:

[0061] u q =K p,q ·e q +K i,q ∫e q dt+ω(L d i d +ψ f )

[0062] Among them, ω e is the electrical angular velocity (calculated by encoder differential), L d , L q is the motor parameter (factory calibration value), ψ f is the permanent magnet flux (0.1-0.3Wb). K p,d , K i,d and K p,q , K i,q These are the PI parameters for the d-axis and q-axis, respectively, and are usually determined through experimental debugging or automatic tuning methods.

[0063] u d 、u q The signal is obtained by the above method, and the electrical angular velocity of the motor rotor is calculated by the encoder differential, λ d ,λ q is the magnetic flux on the motor dq axis, L d , L q is the inductance of the motor on the dq axis, which is used as the input of the prediction model in the virtual model layer.

[0064] Step 2: Modeling the virtual model layer

[0065] The virtual model layer is equipped with a prediction model, which adopts the architecture design of integrating deep neural network with the physical characteristics of the motor. The input parameters of the prediction model are motor-related parameters: one is the real-time working condition parameters, including motor speed ω and electromagnetic torque Te; the other is the motor body parameters, including stator flux ψ d / ψ q 、Inductor L d / L q and stator resistance Rs.

[0066] All input parameters need to be standardized and preprocessed before entering the network.

[0067] For the speed parameter, the maximum and minimum value normalization method is used:

[0068]

[0069] where ω min Set to 0rpm, ω max Take 1.2 times the rated speed of the motor. The torque parameter is scaled logarithmically:

[0070]

[0071] The motor parameters are encoded in percentage deviation based on the nominal value. Take the d-axis inductance as an example:

[0072]

[0073] Among them L d,nom To achieve the design value, the preprocessed 7-dimensional input vector is first subjected to feature expansion through a fully connected encoding layer.

[0074] The prediction model consists of three parts: a convolutional module, a bidirectional LSTM module, and an output decoding layer. The convolutional module adopts a three-level residual structure, and each level contains causal convolution, batch normalization, and gated activation units.

[0075] Predictive network model Figure 2 As shown, the mathematical expression of causal convolution is:

[0076]

[0077] The calculation formula for batch normalization is:

[0078]

[0079] The activation function of the gated activation unit is:

[0080]

[0081] The bidirectional LSTM module consists of two independent LSTM layers, the forward and backward ones. The output decoding layer implements dimensionality transformation through a fully connected layer, and the final output uses the Tanh activation function:

[0082]

[0083] Step 3: Use the twin data layer to precisely align the physical signal with the virtual information;

[0084] In the twin data layer, dynamic time warping and Kalman filtering are combined to achieve precise timing alignment of physical and virtual signals.

[0085] The optimal synchronization strategy is automatically selected based on the characteristics of different working conditions: for steady-state or slowly varying working conditions, an improved dynamic time warping algorithm is preferentially used to find the optimal matching path by constructing a cost matrix containing timing constraints.

[0086] For dynamic working conditions with sudden changes or high-speed operation, the system switches to the Kalman filter compensation mode based on the motor motion equation, uses the rotor dynamics model to predict the signal evolution trend, and eliminates the phase deviation caused by speed fluctuations through the measurement update link, especially when the load suddenly changes, the synchronization error can be further compressed.

[0087] The two synchronization strategies achieve smooth switching through an adaptive weight mechanism. When the speed change rate is detected to exceed the threshold, the transition from DTW to Kalman filtering is completed in a short time, and linear interpolation is used during this period to ensure timing continuity.

[0088] The signal is then enhanced using a residual attention network, which employs a cascaded processing architecture. First, a time-frequency analysis module performs multi-resolution analysis on the input signal. During the time-frequency transformation phase, the noise-free virtual current signal generated by the prediction model in the virtual model layer serves as an ideal reference for constructing a virtual current signal reference library. A speed-adaptive Hanning window function is designed for short-time Fourier transform (SFT). The window function parameters are dynamically optimized based on the real-time speed, ensuring accurate capture of transient characteristics while maintaining frequency resolution.

[0089] In the short-time Fourier transform (SFT), the spectral characteristics of the virtual signal are encoded as prior knowledge into the network weights, guiding the model to automatically suppress noise components that do not conform to the motor's electromagnetic laws during feature extraction. During processing, the spatial attention module continuously compares the time-frequency distribution differences between the actual and virtual signals. Using a residual learning mechanism, it retains valid fault features while filtering out random interference.

[0090] The output of the residual attention network uses a gated residual connection structure to adaptively fuse the enhanced features with the original signal. This structural design ensures the complete preservation of fault information while fully leveraging the noise-free nature of the digital twin model.

[0091] Step 4: The application layer performs electrical fault diagnosis based on physical and virtual signals.

[0092] The application layer includes classified networks such as Figure 4As shown, the classification network includes an input layer, a vibration branch, a current branch, a diagnosis branch and an output layer. The input layer inputs the vibration signal into the vibration branch and the current signal into the current signal. Both the vibration branch and the current branch include a one-dimensional convolution layer, a batch normalization layer, a GELU activation layer and a pooling layer. The one-dimensional convolution kernel of the vibration branch is 5, and the one-dimensional convolution kernel of the current branch is 3. The diagnosis branch receives the features input by the vibration branch and the current branch. If the vibration branch has all zero inputs, it is determined that the vibration signal is missing. The diagnosis branch includes dynamic feature splicing, a convolution layer, a global pooling layer and a fully connected layer. The output layer determines the fault type according to the output features of the diagnosis branch.

[0093] The specific process of making the classification network training data set is as follows:

[0094] The three-phase stator current and axial and radial vibration signals of the motor are collected from a real motor, and a fault label is given. Then, a preprocessing is performed to obtain a data set. The fault label includes the fault location and fault type.

[0095] For the vibration signal branch, preprocessing begins by time-aligning the axial and radial vibration data. Samples are then segmented using a fixed-length sliding window. The window length should cover the characteristic period of typical motor faults. Random overlapping sampling with a 50% overlap ratio is used to expand the sample size. The segmented vibration signals undergo min-max normalization, mapping the amplitude to the [-1, 1] range to eliminate the influence of varying operating conditions. Preprocessed vibration samples require a dedicated zero-input detection module, corresponding to the "zero-input missing" processing module in the network.

[0096] Preprocessing of the current signal branch also requires time window segmentation. The window setting must be synchronized with the vibration signal to ensure that the samples of the two branches are aligned in the time dimension. The residual signal is also normalized to the range [-1, 1] using the same method. The preprocessed vibration and current samples are divided into training, validation, and test sets in a ratio of 7:1.5:1.5. This division ensures that each fault category is evenly distributed across the datasets. The resulting dataset must be able to support parallel processing of the two branches of the network: the vibration branch uses a one-dimensional convolution kernel with K=5 to extract features, while the current branch uses a convolution kernel with K=3, and multimodal feature fusion is implemented in the dynamic feature concatenation layer. The dataset also needs to contain a certain proportion of abnormal samples, such as all-zero vibration signals, to enable the network to learn to handle abnormal conditions such as sensor failure.

[0097] Fault diagnosis process at the application layer, such as Figure 3 As shown, first, the predicted current signal i is received from the virtual model layer. abc* The actual current signal i collected by the physical layer abc, generates the current residual Δi through real-time differential operation abc , and conduct preliminary screening of electrical faults based on preset threshold conditions.

[0098] If the statistical characteristics of Δiabc (including mean offset and harmonic distortion rate) are detected to exceed the safety threshold, the electrical fault flag is immediately triggered. The residual signal is then directly input into the pre-trained classification network for diagnosis without the need for fusion with the vibration signal, thus shortening processing time.

[0099] If it is detected that the statistical characteristics of Δiabc (including mean offset, harmonic distortion rate, etc.) do not exceed the safety threshold, the multimodal feature fusion stage is entered.

[0100] In this stage, the current residual signal is adaptively weighted fused with the vibration signal processed by wavelet packet decomposition, where the vibration signal focuses on capturing the mechanical fault characteristics.

[0101] The fusion process adopts a dynamic weight adjustment strategy to automatically optimize the contribution ratio of current and vibration features according to the signal-to-noise ratio of the current signal and the fault sign status, and generate a composite feature vector with high discrimination.

[0102] The composite feature vector is then input into a pre-trained classification network for detailed diagnosis. The network output corresponds to the probability distribution of 10 typical fault types. Finally, the specific fault category is determined through the Softmax decision layer to achieve the purpose of mechatronic coupling fault diagnosis.

[0103] The above-described embodiments merely represent specific implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of protection of the present application. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the technical concept of the present application, and all such variations and improvements fall within the scope of protection of the present application.

Claims

1. An intelligent diagnosis method for composite faults of permanent magnet synchronous motors based on digital twins, characterized in that: include: Step 1: Establish a digital twin model of the permanent magnet synchronous motor, including a physical entity layer, a virtual model layer, a twin data layer, and an application layer. The virtual model layer includes a prediction model, and the application layer includes a classification model. Step 2: Use the physical layer to collect prophetic data, real-time motor-related parameters and real-time current. The prophetic data includes current signals, vibration signals, speed signals and torque signals under different working conditions. Step 3: Use the prophet data to create a data set to train the prediction model and classification model; Step 4: The twin data layer extracts the difference characteristics between the virtual current generated by the prediction model and the actual current collected by the physical entity layer to form twin data, which is transmitted to the physical entity layer and the virtual model layer respectively, so that the relevant status of both can be updated in real time. The virtual current generated by the prediction model and the prophetic data collected by the physical entity layer are then synchronized in time and space to build a virtual current signal reference library containing iabc* waveforms under different working conditions. Step 5: The application layer subtracts the predicted current from the real-time current, and then uses the virtual current signal reference library to preliminarily determine whether an electrical fault has occurred. If an electrical fault has occurred, the fault features are extracted based on the current signal and input into the classification network. If no electrical fault occurs, the fault features are extracted by fusing the current and vibration signals, input into the classification network, and finally the classification network performs fault diagnosis.

2. The intelligent diagnosis method for composite faults of permanent magnet synchronous motors based on digital twins according to claim 1 is characterized in that: The configuration of the physical entity layer includes: deploying permanent magnet synchronous motor entities, high-precision sensors and data acquisition systems, collecting three-phase current, vibration signals, torque, marking working condition labels, and collecting the voltage u on the d and q axes in normal mode and fault mode under different working conditions. d 、u q The collected signals can be used for training the prediction model in the virtual model layer and subsequent fault diagnosis.

3. The intelligent diagnosis method for composite faults of permanent magnet synchronous motors based on digital twins according to claim 1 is characterized in that: The input parameters of the prediction model are motor-related parameters: one is the real-time operating parameters, including motor speed ω and electromagnetic torque Te; the other is the motor body parameters, including stator flux ψ d / ψ q 、Inductor L d / L q and stator resistance Rs.

4. The intelligent diagnosis method for composite faults of permanent magnet synchronous motors based on digital twins according to claim 3 is characterized in that: The prediction model includes a convolution module, a bidirectional LSTM module and an output decoding layer. The convolution module adopts a three-level residual structure, and each level includes causal convolution, batch normalization and gated activation units.

5. The intelligent diagnosis method for composite faults of permanent magnet synchronous motors based on digital twins according to claim 1 is characterized in that: The twin data layer performs spatiotemporal synchronization on the received data, including: using a combination of dynamic time warping and Kalman filtering to achieve precise timing alignment of physical signals and virtual signals; for steady-state or slowly changing working conditions, an improved dynamic time warping algorithm is preferentially used to find the optimal matching path by constructing a cost matrix containing timing constraints; for dynamic working conditions with sudden changes or high-speed operation, it switches to the Kalman filter compensation mode based on the motor motion equation, uses the rotor dynamics model to predict the signal evolution trend, eliminates the phase deviation caused by speed fluctuations through the measurement update link, and further compresses the synchronization error when the load suddenly changes; the two synchronization strategies achieve smooth switching through an adaptive weight mechanism. When the speed change rate is detected to exceed the threshold, the transition from dynamic time warping to Kalman filtering is completed in a short time, and linear interpolation is used during this period to ensure timing continuity.

6. The intelligent diagnosis method for composite faults of permanent magnet synchronous motors based on digital twins according to claim 1 is characterized in that: The extraction of differential features between actual current and virtual current to form twin data includes enhanced processing through a residual attention network. The residual attention network adopts a cascade processing architecture. First, a time-frequency analysis module is used to perform multi-resolution analysis on the input signal. In the time-frequency transformation stage, the noise-free virtual current signal generated by the virtual model layer is used as an ideal reference benchmark. A short-time Fourier transform is performed by designing a Hanning window function with speed adaptive characteristics. The parameters of the Hanning window function are dynamically optimized according to the real-time speed to ensure that transient features are accurately captured while maintaining frequency resolution.

7. The intelligent diagnosis method for composite faults of permanent magnet synchronous motors based on digital twins according to claim 6 is characterized in that: During the time-frequency transformation process, the spectral characteristics of the virtual signal are encoded into the network weights as prior knowledge, guiding the model to automatically suppress noise components that do not conform to the electromagnetic laws of the motor during feature extraction. During the processing, the spatial attention module continuously compares the time-frequency distribution differences between the actual signal and the virtual signal, retains effective fault features through the residual learning mechanism, and filters out random interference. The output end of the residual attention network adopts a gated residual connection structure to adaptively fuse the enhanced features with the original signal.

8. The intelligent diagnosis method for composite faults of permanent magnet synchronous motors based on digital twins according to claim 1 is characterized in that: The step 5 comprises: Step 5.1: Receive the predicted current signal i from the virtual model layer abc* The actual current signal i collected by the physical layer abc , generates the current residual Δi through real-time differential operation abc , and conduct preliminary screening of electrical faults based on preset threshold conditions; Step 5.2: When it is detected that the statistical characteristics of Δiabc exceed the safety threshold, the electrical fault flag is immediately triggered, and the residual signal is directly input into the pre-trained classification network for diagnosis, and then the process proceeds to step 5.

4. When it is detected that the statistical characteristics of Δiabc do not exceed the safety threshold, the process proceeds to step 5.

3. Step 5.3: If no obvious anomaly is detected, the multimodal feature fusion phase begins. The current residual signal is adaptively weighted and fused with the vibration signal processed through wavelet packet decomposition. The vibration signal focuses on capturing mechanical fault characteristics. The fusion process uses a dynamic weight adjustment strategy to automatically optimize the contribution ratio of current and vibration features based on the current signal-to-noise ratio and fault indicator status. This generates a highly discriminative composite feature vector that is input into the pre-trained classification network for diagnosis. Step 5.4: The pre-trained classification network performs detailed diagnosis. The network output corresponds to the probability distribution of 10 typical fault types. Finally, the specific fault category is determined through the Softmax decision layer to achieve the purpose of mechatronic coupling fault diagnosis.

Images