A method for recognizing inter-turn short circuit fault of permanent magnet synchronous motor

Abstract

This invention relates to the field of permanent magnet motor fault diagnosis, specifically to a method for identifying inter-turn short-circuit faults in permanent magnet synchronous motors. The method includes: 1) constructing an experimental platform for inter-turn short-circuit faults in permanent magnet synchronous motors and building a high-dimensional feature representation space; 2) collecting and preprocessing multi-modal operating signals during motor operation, dynamically tracking the energy distribution changes of the fault signals, and fusing different modal features to form a high-dimensional joint feature tensor; 3) inputting the joint feature tensor into a constructed deep model to extract spatial features, temporal features, and key feature weights; 4) outputting the motor operating state category at the classification layer, performing lightweight optimization on the trained model, and combining transfer learning and domain adaptation mechanisms to achieve adaptive diagnosis between different motor platforms. This method achieves dynamic fusion of multi-modal features, highlighting key features and weakening redundant information, thereby significantly improving diagnostic accuracy and model interpretability.

Description

Technical Field

[0001] This invention relates to the field of permanent magnet motor fault diagnosis, and more specifically to a method for identifying inter-turn short-circuit faults in permanent magnet synchronous motors. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs), as a type of high-efficiency, reliable, and fast-response motor, have been widely used in recent years in fields such as new energy vehicles, wind power generation, rail transportation, aerospace, robotics, and high-end manufacturing. Compared with traditional induction motors, PMSMs have advantages such as high power density, high energy efficiency, and excellent control performance, thus gradually becoming a core component of the next generation of electric drive systems. However, with increasingly complex application scenarios and equipment operating under high loads and multiple conditions for extended periods, the safety and reliability issues of motors are becoming increasingly prominent. Among these, inter-turn short circuits (ITSCs) in the stator windings are one of the most common and highly dangerous potential faults, directly affecting system stability and service life.

[0003] Inter-turn short circuit faults refer to the damage to the insulation layer between some turns of a phase coil in the stator winding of a motor, leading to a local short circuit between conductors. This type of fault can arise from various factors, such as insulation aging caused by prolonged high-temperature operation, manufacturing defects creating weak points in the insulation, mechanical vibration causing winding friction and wear, and electromagnetic stress and overvoltage impacts causing insulation breakdown. Once an inter-turn short circuit occurs, abnormal currents and strong electromagnetic imbalances will appear locally inside the motor, leading to increased stator copper losses, localized overheating, enhanced torque pulsation, and even adverse effects on components such as rotor magnets and bearings. If not detected and addressed in a timely manner in the early stages, inter-turn short circuits can quickly evolve into more serious phase-to-phase short circuits, ground faults, or even motor burnout, causing complete machine shutdown and safety accidents, resulting in huge economic losses.

[0004] To avoid such problems, researchers and engineers have proposed various methods for diagnosing motor faults. Traditional methods primarily rely on signal analysis and mathematical modeling. For example, detection methods based on electrical signals determine the presence of inter-turn short circuits by monitoring the distortion characteristics of the motor's three-phase current, zero-sequence current, or voltage; flux linkage observation-based methods calculate the actual flux linkage using a mathematical model of the motor and compare it with the ideal flux linkage to identify anomalies; and spectral analysis-based methods characterize fault conditions by detecting characteristic frequency components in the current or voltage using tools such as the Fast Fourier Transform (FFT). While these methods have achieved some success in early research, their diagnostic capabilities remain limited, mainly in the following aspects:

[0005] (1) It is not sensitive to mild or early inter-turn short circuits. In the early stage of insulation degradation, the voltage and current waveforms change by very small amplitudes, making them difficult to identify accurately;

[0006] (2) It is highly dependent on changes in environment and operating conditions. Under different loads, speeds, or temperatures, external noise and harmonic interference can easily mask fault characteristics;

[0007] (3) Manual feature extraction relies on experience and lacks adaptability and universality. Parameters need to be redesigned for different motor models.

[0008] (4) The algorithm has high complexity and insufficient real-time performance, making it difficult to deploy directly on embedded platforms or edge computing systems.

[0009] Therefore, developing an intelligent diagnostic method that can operate stably under multiple operating conditions, has self-learning capabilities, and high real-time performance has become an important research direction in the field of motor condition monitoring. Summary of the Invention

[0010] To address the problems of insufficient accuracy in early fault detection, poor anti-interference capability, and difficulty in real-time deployment in existing technologies, the purpose of this invention is to provide a diagnostic method for inter-turn short-circuit fault detection in permanent magnet synchronous motors that differs from existing diagnostic methods that rely solely on a single current signal. This method emphasizes the diversity and complementarity of signal sources, ensuring stability and reliability under complex operating conditions. In the data preprocessing stage, multi-scale time-frequency analysis and feature enhancement methods are employed to preserve the timing information of motor operation. Furthermore, feature alignment and noise suppression improve the distinguishability of the data. This method utilizes CLA and multi-source feature collaborative analysis to detect inter-turn short-circuit faults in permanent magnet synchronous motors.

[0011] To achieve the above objectives, the technical solution adopted by the present invention is: a method for identifying inter-turn short-circuit faults in a permanent magnet synchronous motor, comprising the following steps:

[0012] Step 1: Construct an experimental platform for inter-turn short-circuit faults of permanent magnet synchronous motors, introduce a multi-source signal collaborative analysis mechanism, and model the multi-mode operating signals of permanent magnet synchronous motors under different operating conditions to construct a high-dimensional feature representation space.

[0013] Step 2: Collect and preprocess the multi-modal operation signals of the motor during operation, perform short-time Fourier transform to extract time-frequency features, perform stationary wavelet transform to obtain time-frequency and multi-scale features, perform joint analysis in the time and frequency domains, dynamically track the energy distribution changes of the fault signal, and fuse different modal features to form a high-dimensional joint feature tensor.

[0014] Step 3: Input the joint feature tensor into the constructed deep model that includes convolutional neural network, long short-term memory network and attention mechanism, organically combine spatial feature extraction, temporal feature modeling and key feature weight adaptive allocation, and extract spatial features, temporal features and key feature weights;

[0015] Step 4: The classification layer outputs the motor operating state category, performs lightweight optimization on the trained model, and combines transfer learning and domain adaptation mechanisms to achieve adaptive diagnosis between different motor platforms.

[0016] The above-mentioned method for identifying inter-turn short-circuit faults in permanent magnet synchronous motors, in step 1, the multi-mode operating signals include stator current, flux linkage, temperature, and vibration signals, and step 1 includes:

[0017] A mathematical model of the permanent magnet synchronous motor under normal operation and inter-turn short-circuit fault conditions is established, and its stator voltage equation is expressed in the dq coordinate system as follows: ,in, Stator voltage, For stator current, For stator resistance, For inductance, Electric angular velocity, For permanent magnet flux linkage;

[0018] If an inter-turn short circuit occurs in one phase, the current expression can be modified as follows: ,in, This represents the phase current under inter-turn short-circuit fault conditions. This represents the phase current under normal operating conditions. This is the severity factor of the fault;

[0019] The current is recorded synchronously with a unified sampling frequency and time. Magnetic Link ,temperature With vibration signal It outputs a multi-channel signal matrix. It serves as the direct input for subsequent signal processing and feature extraction modules.

[0020] In the above-mentioned method for identifying inter-turn short-circuit faults in permanent magnet synchronous motors, step 2 involves performing a short-time Fourier transform on the current and flux linkage signals to extract time-frequency features. ,in, Represents the time-frequency distribution function. Represents the original signal. Indicates the time-shifting window function. Represents a time variable. Represents frequency variables. This represents the window function shift. Represents the complex exponential basis functions;

[0021] Perform stationary wavelet transform on the temperature and vibration signals. , , , ,in, Indicates signal The results of the stationary wavelet transform, This represents the original temperature signal or vibration signal. Denotes the wavelet approximation coefficients of the j-th layer. Represents the wavelet detail coefficients of the j-th layer. For the j-th layer low-pass filter, For the j-th layer high-pass filter, Indicates the wavelet decomposition level. This represents the index of the coefficient in the time series after wavelet transform. This is the summation index used during convolution operations in the stationary wavelet transform process;

[0022] The two-dimensional time-spectrum graph generated by the short-time Fourier transform is concatenated and fused with the multi-scale feature vectors of the stationary wavelet transform to construct a multimodal joint feature: This forms a high-dimensional joint feature tensor, in which This represents the two-dimensional time-frequency feature matrix extracted by STFT. This represents the multi-scale feature vector extracted by SWT. This represents the joint high-dimensional feature tensor after splicing.

[0023] In the aforementioned method for identifying inter-turn short-circuit faults in permanent magnet synchronous motors, in step 3, the convolutional neural network extracts two-dimensional time-frequency features through multi-layer convolution and pooling operations, and automatically learns the spatial distribution pattern of the signal. ,in, This represents the output feature map of the convolutional layer. Represents the convolution kernel weight matrix. Indicates the input feature map, denoted as the bias term, i and j represent the position indices of the output feature map of the convolutional layer in two spatial dimensions, respectively, and m and n represent the summation indices of the convolutional kernel in the corresponding dimensions during the convolution operation;

[0024] Long Short-Term Memory (LTSM) networks retain the temporal features of motor operation over a long period, model the temporal correlation of feature sequences, capture time-series dependencies, and avoid gradient vanishing. The state equation is as follows: ,in, Indicates the output of the forget gate. This represents the input features at time t. This indicates the hidden state in the previous moment. Indicates the current hidden state. This indicates the current state of the memory unit. Indicates the memory state of the previous moment. Indicates the input gate output. Indicates the output of the output gate. This represents the Sigmoid activation function. This represents the Tanh activation function. Indicates candidate memory units. Represents the weight matrix of each gate. , Indicates the bias terms for each gate. This represents the weight matrix from the hidden state of the previous time step to the current gate structure;

[0025] Attention mechanisms focus on key time segments and weight these segments to enhance significant fault characteristics. , , ,in, This represents the attention weight at time t. This indicates the score for attention. This represents the feature vector at time t of the LTSM output. Represents the attention score vector. Indicates the length of the time series. This represents the attention-weighted vector.

[0026] The aforementioned method for identifying inter-turn short-circuit faults in permanent magnet synchronous motors uses attention-weighted vectors. Inputting into a fully connected layer yields the logits vector. , The class probabilities are obtained through Softmax. In step 4, the classification layer uses the Softmax function to output multiple fault categories. ,in, This represents the predicted probability of the i-th class. This represents the network input of class i. Indicates the total number of categories. This represents the index of the Softmax normalized summation. Indicates the classification layer weights. This indicates the classification layer bias.

[0027] In the aforementioned method for identifying inter-turn short-circuit faults in permanent magnet synchronous motors, step 4 involves lightweight optimization using a strategy combining structural pruning and parameter quantization to reduce the number of model parameters and computational complexity. After lightweight optimization, the model is directly deployed on embedded systems or edge computing terminals. Symmetric integer quantization is used to quantize the FP32 weights to INT8, reducing model storage overhead and computational complexity. , ,in, This represents the quantized weight value. This represents the original floating-point weight value. Represents the set of weights. Indicates the quantization step size. Represents the floor function;

[0028] Obtain the lightweight diagnostic model ,in This is the set of parameters obtained after pruning and quantization, and then fine-tuning. This represents the function mapping relationship of the entire deep diagnostic network;

[0029] To address the differences in motor models or operating conditions, a transfer learning mechanism is introduced to perform transfer learning on the lightweight model. The optimization objective is as follows: ,in Indicates source domain monitoring loss, Indicates the number of fault categories. Represents the source domain sample label vector. This represents the ratio of the monitoring loss to the domain alignment loss;

[0030] , This represents the MMD field alignment loss. Indicates the number of samples in the source domain. Indicates the number of samples in the target domain. Represents source domain samples, Represents the target domain sample. This represents a feature extraction network. Represents the kernel mapping function;

[0031] After domain alignment and a small number of target domain iterations, a cross-platform model is obtained: in, This represents the set of model parameters after transfer learning, which serves as the final model deployed on the target platform, with any input from the target platform. After processing, input the data directly into the model: (F) Output the motor fault category or fault level on the target platform.

[0032] The beneficial effects of this invention's method for identifying inter-turn short-circuit faults in permanent magnet synchronous motors are as follows: The invention proposes a joint architecture of CLA (CNN-LSTM and Attention mechanism), which utilizes convolutional neural networks (CNN) to extract multimodal time-frequency features and recurrent neural networks (LSTM) to model temporal dependencies. Furthermore, the attention mechanism focuses attention on key information, adaptively weighting features from different modalities and time periods. This invention achieves dynamic fusion of multimodal features for the first time, highlighting key features and reducing redundant information, thereby significantly improving diagnostic accuracy and model interpretability.

[0033] Furthermore, this invention incorporates an optimization strategy combining lightweight design and transfer learning in its implementation. By employing methods such as network pruning and parameter quantization, the model size is compressed, enabling it to run on low-power embedded platforms or edge computing devices. Simultaneously, transfer learning and domain adaptation mechanisms are utilized to allow the model to quickly adapt to different types and specifications of permanent magnet synchronous motors without requiring large-scale retraining. This design overcomes the limitations of existing technologies, which generally suffer from high computational resource consumption and insufficient generalization ability.

[0034] Therefore, this invention can not only achieve early and accurate detection of inter-turn short circuit faults in permanent magnet synchronous motors, but also has high robustness, strong portability and good interpretability, and can be widely used in high reliability fields such as new energy vehicles, wind power generation and rail transit. Attached Figure Description

[0035] Figure 1 This is a diagram illustrating the overall system structure in an embodiment of the present invention.

[0036] Figure 2 This is a comparison of the effectiveness of transfer learning in the target region in this embodiment of the invention;

[0037] Figure 3 This illustrates the impact of lightweight measurement on model accuracy in this embodiment of the invention.

[0038] Figure 4 This is the fault classification confusion matrix in this embodiment of the invention;

[0039] Figure 5 The figure shows the experimental results in an embodiment of the present invention. Detailed Implementation

[0040] To enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention will be described below in conjunction with specific embodiments and accompanying drawings.

[0041] In recent years, the rapid development of deep learning and artificial intelligence has provided new solutions for motor fault diagnosis. Researchers have begun to apply models such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), and autoencoders to operational signal analysis. These models possess automatic feature extraction and end-to-end learning capabilities, reducing reliance on human experience. CNNs excel at extracting local time-frequency features of signals, RNNs can capture the dynamic dependencies of time series, and autoencoders can be used for signal reconstruction and noise suppression. These methods have demonstrated high recognition accuracy and robustness in experiments. However, limitations remain in engineering deployment, such as insufficient information due to single signal dependencies, large model parameter count affecting real-time performance, weak generalization ability across motor platforms, and insufficient interpretability of results. Therefore, how to achieve lightweight design and transferability while maintaining model accuracy has become a key research direction in the field of intelligent diagnostics. However, several problems still need to be solved in practical applications: First, most existing studies rely only on a single signal mode (such as current or vibration) and fail to fully integrate multi-source information such as current, magnetic flux, temperature, and vibration, thus limiting the comprehensiveness and robustness of the diagnosis; Second, deep learning models are often complex in structure and have a large number of parameters, resulting in slow inference speed and high computational resource requirements, making it difficult to deploy them directly on embedded or edge computing devices in industrial sites; Third, different motor models and operating conditions vary greatly, resulting in insufficient model generalization ability and requiring frequent retraining; Fourth, some deep learning methods lack interpretability, making it difficult for maintenance personnel to directly understand and trust the diagnostic results.

[0042] In summary, inter-turn short-circuit fault detection in permanent magnet synchronous motors (PMSMs) is a crucial research direction for ensuring the safe and reliable operation of these motors. While existing traditional methods can identify some fault characteristics under laboratory conditions, they suffer from insufficient sensitivity, poor robustness, and weak real-time performance under complex real-world operating conditions. Emerging deep learning methods have shown advantages in feature extraction and classification, but further improvements are needed in areas such as multi-source feature collaborative analysis, lightweight models, real-time deployment, and result interpretability. Therefore, there is an urgent need to propose a new detection method that can integrate multi-source information and leverage the powerful feature learning capabilities of deep learning models to achieve early, accurate, and real-time detection of inter-turn short-circuit faults, adaptable to different operating conditions and motor platforms. This is precisely the core problem that this invention aims to solve and the direct driving force behind the subsequent invention.

[0043] Example 1

[0044] like Figures 1-5 As shown, a method for identifying inter-turn short-circuit faults in a permanent magnet synchronous motor includes the following:

[0045] I. Motor Modeling and Fault Mechanism.

[0046] An experimental platform for inter-turn short-circuit faults in permanent magnet synchronous motors (PMSMs) was constructed. A multi-source signal collaborative analysis mechanism was introduced, and multi-mode operating signals of the PMSM under different operating conditions were modeled to construct a high-dimensional feature representation space.

[0047] Current motor fault detection methods often rely on a single signal, such as stator current or vibration measurements. This approach has limited feature dimensions and struggles to capture early signs of mild inter-turn short circuits. To improve detection sensitivity, this embodiment introduces a multi-source signal collaborative analysis mechanism at the mechanistic level, jointly modeling current, flux linkage, temperature, and vibration information to construct a high-dimensional feature representation space, thus overcoming the limitations of single-signal detection. Therefore, a mathematical model of the permanent magnet synchronous motor (PMSM) under normal operation and inter-turn short circuit fault conditions is first established. Its stator voltage equation is expressed in the dq coordinate system as follows: ,in, Stator voltage, For stator current, For stator resistance, For inductance, Electric angular velocity, It is a permanent magnet flux linkage.

[0048] If an inter-turn short circuit occurs in one phase, the current expression can be modified as follows: ,in, This represents the phase current under inter-turn short-circuit fault conditions. This represents the phase current under normal operating conditions. The severity factor is determined by changing the fault severity factor. It can simulate inter-turn short-circuit faults of varying degrees. With... As the voltage increases, the distortion of the stator current increases significantly, and the harmonic content and unbalanced magnetic pull also increase accordingly. This invention utilizes these abnormal characteristics to achieve fault detection.

[0049] During the operation of a permanent magnet synchronous motor, the stator currents of phases A, B, and C are measured in real time by connecting current sensors in series in the three-phase stator winding circuits. The current sensors are preferably Hall effect current sensors or high-precision current transformers, and their outputs are connected to the analog input channel of a data acquisition card.

[0050] During motor operation, the stator flux linkage value is obtained by integrating the stator voltage equation using the collected stator three-phase current signal and motor stator voltage signal, combined with stator resistance parameters and current feedback signal.

[0051] Temperature sensors are pre-embedded near the stator windings or in the stator slots of the permanent magnet synchronous motor to detect changes in winding temperature in real time. The output signal is processed by a signal conditioning circuit and then input into the data acquisition system.

[0052] A vibration acceleration sensor is fixedly installed on the surface of the permanent magnet synchronous motor housing or bearing seat to collect mechanical vibration signals during motor operation. During motor operation, the vibration acceleration sensor converts the mechanical vibration into an electrical signal, which is amplified and filtered before being input into a data acquisition system. Continuous sampling is performed at a preset sampling frequency to obtain a vibration acceleration time series. This signal is used to reflect the electromagnetic force imbalance caused by inter-turn short circuits and the resulting abnormal vibration characteristics.

[0053] The current collected by the experimental platform built through the above steps Magnetic Link ,temperature With vibration signal All signals are recorded synchronously with a uniform sampling frequency and time, and output as a multi-channel signal matrix. This matrix serves as the direct input to subsequent signal processing and feature extraction modules, providing the original data foundation for subsequent time-frequency feature extraction and feature fusion.

[0054] II. Signal Processing and Feature Extraction.

[0055] The system collects and preprocesses multimodal operating signals during motor operation, performs short-time Fourier transform (STFT) to extract time-frequency features, performs stationary wavelet transform (SWT) to obtain time-frequency and multi-scale features, performs joint analysis in the time and frequency domains, dynamically tracks the energy distribution changes of fault signals, and fuses different modal features to form a high-dimensional joint feature tensor.

[0056] In the signal processing stage, the current collected by the experimental platform built in step 1... Magnetic Link ,temperature With vibration signal All samples were recorded synchronously with a uniform sampling frequency and time, and preprocessed. Short-Time Fourier Transform (STFT) was then used to extract time-frequency features. ,in, Represents the time-frequency distribution function. Represents the original signal. Indicates the time-shifting window function. Represents a time variable. Represents frequency variables. This represents the window function shift. It represents a complex exponential basis function, which can be used for joint analysis in the time and frequency domains to dynamically track changes in the energy distribution of fault signals.

[0057] This process performs joint analysis in the time and frequency domains, enabling dynamic tracking of energy distribution changes in fault signals. Compared to existing methods based solely on FFT, the STFT in this embodiment maintains both time and frequency resolution, making it more sensitive to early fault harmonic variations.

[0058] To further separate features across different frequency bands, this embodiment also introduces stationary wavelet transform (SWT): , , , ,in, Indicates signal The results of the stationary wavelet transform, This represents the original temperature signal or vibration signal. Denotes the wavelet approximation coefficients of the j-th layer. Represents the wavelet detail coefficients of the j-th layer. For the j-th layer low-pass filter, For the j-th layer high-pass filter, Indicates the wavelet decomposition level. This represents the index of the coefficient in the time series after wavelet transform. This is the summation index used during convolution operations in the stationary wavelet transform process, and is used to traverse the approximation coefficients of the previous layer.

[0059] SWT can capture the gradual trend of temperature rise and magnetic flux change in the low-frequency band and locate the transient impact when a fault is triggered in the high-frequency band.

[0060] The two-dimensional time spectrum generated by the short-time Fourier transform Multiscale eigenvectors of stationary wavelet transform Perform splicing and fusion to construct multimodal joint features: This forms a high-dimensional joint feature tensor, in which This represents the two-dimensional time-frequency feature matrix extracted by STFT. This represents the multi-scale feature vector extracted by SWT. This represents the joint high-dimensional feature tensor after splicing.

[0061] The model's convolutional layers first receive the tensor and perform spatial feature extraction, then LSTM captures temporal dependencies, and finally, an attention mechanism calculates keyframe weights. Here, F is the sole input to step 3, ensuring the end-to-end implementability and repeatability of the entire method.

[0062] Compared to traditional feature splicing methods, this embodiment adopts a unified normalization and attention weighting mechanism to suppress noise interference and enhance the complementarity between different modalities.

[0063] III. Deep Learning Model Design.

[0064] The joint feature tensor is input into a deep model that includes convolutional neural networks, long short-term memory networks, and attention mechanisms. Spatial feature extraction, temporal feature modeling, and key feature weight adaptive allocation are organically combined to extract spatial features, temporal features, and key feature weights.

[0065] The deep model proposed in this embodiment comprises three core components: a convolutional neural network (CNN), a long short-term memory network (LSTM), and an attention mechanism. This architecture achieves an organic combination of spatial feature extraction, temporal feature modeling, and adaptive allocation of key feature weights.

[0066] Convolutional neural networks (CNNs) extract two-dimensional time-frequency features through multiple layers of convolution and pooling operations. Compared with traditional hand-crafted features, CNNs can automatically learn the spatial distribution patterns of signals, avoiding subjectivity in feature selection.

[0067] ,in, This represents the output feature map of the convolutional layer. Represents the convolution kernel weight matrix. Indicates the input feature map, The term represents the bias term, where i and j represent the position indices of the output feature map of the convolutional layer in two spatial dimensions, and m and n represent the summation indices of the convolutional kernel in the corresponding dimensions during the convolution operation.

[0068] Long Short-Term Memory (LTTM) networks retain the temporal characteristics of motor operation over a long period, modeling the temporal correlations of feature sequences. By capturing time-series dependencies and avoiding gradient vanishing, the core of the LSTM network structure lies in its gating mechanism, with the following state equation:

[0069] ,in, Indicates the output of the forget gate. This represents the input features at time t. This indicates the hidden state in the previous moment. Indicates the current hidden state. This indicates the current state of the memory unit. Indicates the memory state of the previous moment. Indicates the input gate output. Indicates the output of the output gate. This represents the Sigmoid activation function. This represents the Tanh activation function. Indicates candidate memory units. Represents the weight matrix of each gate. , Indicates the bias terms for each gate. This represents the weight matrix from the hidden state of the previous time step to the current gate structure.

[0070] Since inter-turn short-circuit faults typically manifest as local abrupt changes, intermittent enhancements, or non-persistent anomalies in time series, relying solely on the equal weighting of all time-step features from the LSTM output can easily weaken the contribution of key time segments to the final classification result.

[0071] Therefore, an attention mechanism is introduced after the LSTM output layer to adaptively weight the hidden states at different time steps, so as to highlight the time segment features that play a key role in fault discrimination, thereby realizing a further mapping from "temporal features" to "discriminative deep representation".

[0072] The above structure can preserve the timing characteristics of motor operation for a long time and avoid the gradient vanishing problem.

[0073] An attention mechanism is introduced to focus on key time segments, and these key time segments are weighted to enhance the characteristics of significant faults. , in, This represents the attention weight at time t. This indicates the score for attention. This represents the feature vector at time t of the LTSM output. This represents the attention score vector, which improves the model's ability to respond to sudden changes in fault characteristics, enabling the model to maintain a high recognition rate under complex working conditions.

[0074] After obtaining the attention weights at each time step, the attention-weighted context feature vector is calculated. Its mathematical expression is: in, Indicates the length of the time series. This represents the attention-weighted vector. Context feature vector. It integrates information from the entire time series and significantly enhances the fault characteristics of key time segments through a weighting mechanism.

[0075] This mechanism enhances the model's ability to respond to sudden changes in fault characteristics, enabling it to maintain a high recognition rate under complex operating conditions.

[0076] IV. Model Optimization and Cross-Domain Adaptation.

[0077] The classification layer outputs the motor operating state category, performs lightweight optimization on the trained model, and combines transfer learning and domain adaptation mechanisms to achieve adaptive diagnosis across different motor platforms. The attention vector is then... Inputting into a fully connected layer yields the logits vector. , The class probabilities are obtained through Softmax.

[0078] The classification layer uses the Softmax function to output multiple fault categories:

[0079] ,in, This represents the predicted probability of the i-th class. This represents the network input of class i. Indicates the total number of categories. This represents the index of the Softmax normalized summation. Indicates the classification layer weights. This indicates the classification layer bias.

[0080] The classification layer is the core part of the model's output, while lightweight optimization compresses and accelerates the parameters of the entire deep network structure (including the classification layer).

[0081] To deploy on low-computing devices, perform the following steps on the trained model.

[0082] To enable real-time model execution on low-power platforms, this embodiment employs weight pruning and parameter quantization techniques. After quantization, the number of model parameters is reduced by approximately 70%, and the computation speed is increased by 2.8 times, allowing direct deployment in embedded systems or edge computing terminals. Quantization alters both the feature input received by the classification layer (from the pruned backbone network) and directly affects the weight matrix of the classification layer itself.

[0083] All weights in the CNN, LSTM, Attention, and classification layers are quantized using symmetric integer quantization, converting the FP32 weights to INT8. Weight quantization involves converting network weight parameters, represented as 32-bit floating-point numbers, into 8-bit signed integers through numerical mapping, thereby reducing model storage overhead and computational complexity without altering the physical meaning of the weights. , ,in, This represents the quantized weight value. This represents the original floating-point weight value. Represents the set of weights. This represents the quantization step size, which determines the interval for weight discretization. This represents the floor function.

[0084] Obtain the lightweight diagnostic model ,in, This is the set of parameters obtained after pruning and quantization, and then fine-tuning. It represents the function mapping relationship of the entire deep diagnostic network, including CNN module, LSTM module and attention mechanism module, which serves as the basic model for subsequent transfer learning.

[0085] To address the differences in motor models or operating conditions, this embodiment introduces a transfer learning mechanism to perform transfer learning on the lightweight model. Its optimization objective is: ,in, Indicates source domain monitoring loss, Indicates the number of fault categories. Represents the source domain sample label vector. This represents the scaling factor for the tradeoff between the supervision loss and the domain alignment loss, ranging from 0.1 to 1.0, and can be determined by the validation set.

[0086] , This represents the MMD field alignment loss. Indicates the number of samples in the source domain. Indicates the number of samples in the target domain. Represents source domain samples, Represents the target domain sample. This represents a feature extraction network. This represents the kernel mapping function.

[0087] By minimizing the source domain supervision loss, the model avoids a decline in its ability to distinguish known fault categories while performing domain alignment and parameter transfer, thereby ensuring the stability and reliability of the transfer learning process.

[0088] After domain alignment and a small number of target domain iterations, a cross-platform model is obtained: ,in, This represents the diagnostic model after transfer learning. This represents the function mapping relationship of the entire deep diagnostic network, including the CNN module, LSTM module, and attention mechanism module. This represents the set of model parameters after transfer learning. The goal is to ensure that the model retains its original classification ability while adapting to the differences in characteristics of the target platform.

[0089] Transferred model This serves as the final deployment model for the target platform. Any input from the target platform. After processing, input the data directly into the model: (F) Output the motor fault category or fault level on the target platform.

[0090] Example 2

[0091] To verify the effectiveness of the inter-turn short-circuit fault identification method for permanent magnet synchronous motors proposed in this invention, this embodiment constructs a comprehensive experimental platform for multi-modal signal acquisition, time-frequency feature construction, and deep learning diagnosis, and conducts actual tests under multiple operating conditions and multiple fault levels.

[0092] The experimental system consists of a signal acquisition module, a feature processing module, a deep learning analysis module, and a result output module. The permanent magnet synchronous motor used in the experiment has a rated power of 3kW, a rated voltage of 220V, a rated speed of 1500r / min, and is equipped with a magnetic powder brake as a loading device, with a maximum output torque of 30 N·m. To obtain multi-source operating data reflecting the electromagnetic, thermal, and mechanical states of the motor, sensors were installed on the motor casing and input terminals in this embodiment. These included a Hall current sensor (0~30 A) for measuring the stator three-phase current, a PT100 resistance temperature detector (0~150 ℃) for temperature measurement, and a PCB356A12 accelerometer for acquiring vibration response. The flux linkage signal was estimated in real time by a built-in sliding mode flux linkage observer. All sensor signals were synchronously acquired at a uniform sampling frequency of 10kHz, with a single sampling duration of 10s.

[0093] To improve the authenticity and repeatability of the data, the experiment was conducted under different loads (25%, 50%, 75%, 100%) and rotational speeds (800 r / min, 1200 r / min, 1600 r / min) for multiple samplings. Each operating condition was repeated 20 times, resulting in a total of 960 sets of multi-source raw data, which were used to construct a multimodal raw signal matrix. .

[0094] To simulate different degrees of inter-turn short-circuit faults in the motor, this embodiment reserves a controllable short-circuit terminal in phase A winding, and connects an external adjustable resistor. Controlling the degree of inter-turn short circuits according to fault factors in, The resistance corresponding to the shorted coil segment (number of turns) is represented, and three fault levels are set: mild (3% of turns shorted), moderate (8%), and severe (15%). For each fault level, 20 sets of data are collected under the above different speed and load conditions to establish an experimental dataset covering normal operation and the three fault levels.

[0095] The acquired multimodal signals first undergo a unified preprocessing procedure, including bandpass filtering (10-500Hz), Clarke-Park coordinate transformation of three-phase currents, moving average filtering of temperature signals, high-pass filtering of vibration signals (10Hz cutoff frequency), and normalization based on mean-standard deviation, to obtain a cleaned signal set. Subsequently, a short-time Fourier transform (STFT) was performed on the current and flux linkage signals to extract two-dimensional time-frequency features. A 128×128 time-frequency spectrum matrix was generated using a Hamming window configuration with a window length of 1024, an overlap ratio of 50%, and 2048 FFT points. Simultaneously, a four-level stationary wavelet transform (SWT) was performed on the temperature and vibration signals using the db4 wavelet basis to obtain multi-scale feature vectors including approximation coefficients and detail coefficients. The length is 1024. The two types of features are then used to construct a high-dimensional joint feature tensor through size matching, normalization, and multi-channel concatenation. The resulting tensor has a size of 128×128×3 and is used as a unified input for deep neural networks.

[0096] The deep learning model constructed in this embodiment consists of three parts: a convolutional neural network (CNN), a long short-term memory network (LSTM), and an attention mechanism. The CNN part consists of three convolutional layers with 16, 32, and 64 channels respectively, used to extract spatial features of the joint feature tensor. The LSTM layer contains 128 hidden units to capture temporal variation patterns. Based on this, an attention mechanism is introduced to calculate the weights at different time steps, obtain key information of the fault feature mutation locations, and thus generate the final context vector as the input to the classification layer. The model training uses the Adam optimizer with a learning rate of 0.001, a batch size of 64, and 200 training epochs. Dropout of 0.3 is added between convolutional layers to suppress overfitting. Experimental data is divided into training, validation, and test sets according to a 70% / 15% / 15% ratio, totaling 672 training sets, 144 validation sets, and 144 test sets.

[0097] To improve model efficiency, this embodiment performs lightweight processing on the trained model. First, based on the L1 norm of the convolutional kernel, 30% structural pruning is performed on the CNN channels to effectively remove redundant feature maps. Then, a linear fixed-point weight quantization method is used to compress the original 32-bit floating-point weights to 8 bits, with a quantization step size set to Δ=0.02. After lightweight processing, the model parameter size is reduced from 18.2MB to 5.1MB, a reduction of approximately 72%, and the inference time is reduced from 84 ms to 29ms, an improvement of approximately 2.9 times in inference speed, meeting the deployment requirements of embedded devices.

[0098] Experimental results show that the overall recognition accuracy of the proposed solution reached 98.7% during the testing phase, and its detection accuracy under mild inter-turn short circuit conditions was significantly higher than that of traditional methods (an improvement of approximately 12%). After introducing the attention mechanism, the model's responsiveness to key feature regions improved by approximately 15%, effectively focusing on time-frequency abrupt changes in abnormal signals. Through lightweight processing via network pruning and parameter quantization, the model parameter size was reduced by approximately 65%, and the inference speed was increased to 2.8 times that of the original model. The single fault identification latency was less than 100 milliseconds, meeting the requirements for real-time online detection.

[0099] To verify the model's cross-device adaptability, transfer learning tests were further conducted on different models of permanent magnet synchronous motors. Adaptive alignment of feature distributions was achieved by minimizing the maximum mean difference (MMD) between samples in the source and target domains. After transfer learning, the model achieved 94.6% accuracy on the new motor platform without retraining, demonstrating the good versatility and scalability of the proposed solution.

[0100] In summary, this embodiment fully verifies the stability and reliability of the proposed technical solution under multiple operating conditions and multiple platforms. The proposed deep learning diagnostic framework can not only accurately identify inter-turn short-circuit faults in the early stages, but also operate efficiently in resource-constrained embedded environments, providing an engineering-featured solution for motor health management and intelligent operation and maintenance.

[0101] This invention proposes a method for detecting inter-turn short-circuit faults in permanent magnet synchronous motors (PMSMs) based on multi-source feature collaborative analysis and deep learning. This method fuses multi-source features from motor current, flux linkage, temperature, and vibration signals to construct a joint time-frequency feature space. A deep learning architecture employing a CNN-LSTM-attention mechanism is then used to achieve high-precision identification of early-stage motor faults. Compared to traditional single-signal analysis and shallow classification models, this invention significantly improves diagnostic accuracy and noise resistance. Through lightweight design and transfer learning optimization, the model can run stably in resource-constrained embedded devices and adapt to feature distribution variations across different motor platforms, exhibiting good versatility and scalability. This method provides a new technical path for motor health management and intelligent operation and maintenance, with broad engineering application prospects in fields such as new energy vehicles, industrial servos, and wind power generation.

[0102] This invention jointly collects heterogeneous signals such as current, flux linkage, temperature, vibration, and acoustics during motor operation. After denoising, normalization, and time-frequency domain feature extraction, these signals are input into a deep learning model containing a convolutional neural network (CNN), a recurrent neural network (RNN), and an attention weighting module to achieve multi-dimensional feature joint learning and intelligent classification. This solution can identify signs of inter-turn short circuits in stator windings at an early stage, and features high detection accuracy, strong anti-interference capability, and online deployment. It is suitable for scenarios such as new energy vehicle drive systems, wind turbine generators, and rail transit traction equipment.

[0103] The above embodiments are merely illustrative of the structural concept and features of the present invention, intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly, and should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made based on the essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying inter-turn short-circuit faults in a permanent magnet synchronous motor, characterized in that, Includes the following steps: Step 1: Construct an experimental platform for inter-turn short-circuit faults of permanent magnet synchronous motors, introduce a multi-source signal collaborative analysis mechanism, and model the multi-mode operating signals of permanent magnet synchronous motors under different operating conditions to construct a high-dimensional feature representation space. The multi-mode operating signals include stator current, flux linkage, temperature and vibration signals. Step 2: Collect and preprocess the multi-modal operation signals of the motor during operation, perform short-time Fourier transform to extract time-frequency features, perform stationary wavelet transform to obtain time-frequency and multi-scale features, perform joint analysis in the time and frequency domains, dynamically track the energy distribution changes of the fault signal, and fuse different modal features to form a high-dimensional joint feature tensor. Perform short-time Fourier transform on the current and flux linkage signals to extract time-frequency features: ,in, Represents the time-frequency distribution function. Represents the original signal. Indicates the time-shifting window function. Represents a time variable. Represents frequency variables. This represents the window function shift. Represents the complex exponential basis functions; Perform stationary wavelet transform on temperature and vibration signals , , , ,in, Indicates signal The results of the stationary wavelet transform, This represents the original temperature signal or vibration signal. Denotes the wavelet approximation coefficients of the j-th layer. Represents the wavelet detail coefficients of the j-th layer. For the j-th layer low-pass filter, For the j-th layer high-pass filter, Indicates the wavelet decomposition level. This represents the index of the coefficient in the time series after wavelet transform. This is the summation index used during convolution operations in the stationary wavelet transform process; The two-dimensional time-spectrum graph generated by the short-time Fourier transform is concatenated and fused with the multi-scale feature vectors of the stationary wavelet transform to construct a multimodal joint feature: This forms a high-dimensional joint feature tensor, in which This represents the two-dimensional time-frequency feature matrix extracted by STFT. This represents the multi-scale feature vector extracted by SWT. Represents the concatenated joint high-dimensional feature tensor Step 3: Input the joint feature tensor into the constructed deep model that includes convolutional neural network, long short-term memory network and attention mechanism, organically combine spatial feature extraction, temporal feature modeling and key feature weight adaptive allocation, and extract spatial features, temporal features and key feature weights; Step 4: The classification layer outputs the motor operating state category, performs lightweight optimization on the trained model, and combines transfer learning and domain adaptation mechanisms to achieve adaptive diagnosis between different motor platforms. Lightweight optimization employs a strategy combining structural pruning and parameter quantization to reduce the number of model parameters and lower computational complexity. After lightweight optimization, the model can be directly deployed on embedded systems or edge computing terminals. Symmetric integer quantization is used to quantize FP32 weights to INT8, reducing model storage overhead and computational complexity. , ,in, This represents the quantized weight value. This represents the original floating-point weight value. Represents the set of weights. Indicates the quantization step size. Represents the floor function; Obtain the lightweight diagnostic model ,in This is the set of parameters obtained after pruning and quantization, and then fine-tuning. This represents the function mapping relationship of the entire deep diagnostic network; To address the differences in motor models or operating conditions, a transfer learning mechanism is introduced to perform transfer learning on the lightweight model. The optimization objective is as follows: ,in Indicates source domain monitoring loss, Indicates the number of fault categories. Represents the source domain sample label vector. This represents the ratio of the monitoring loss to the domain alignment loss; , Indicates the MMD field alignment loss. Indicates the number of samples in the source domain. Indicates the number of samples in the target domain. Represents source domain samples, Represents the target domain sample. This represents a feature extraction network. Represents the kernel mapping function; After domain alignment and a small number of target domain iterations, a cross-platform model is obtained: ,in, This represents the set of model parameters after transfer learning, which serves as the final model deployed on the target platform, with any input from the target platform. After processing, input the data directly into the model: (F) Output the motor fault category or fault level on the target platform.

2. The method for identifying inter-turn short-circuit faults in a permanent magnet synchronous motor according to claim 1, characterized in that, Step 1 includes: A mathematical model of the permanent magnet synchronous motor under normal operation and inter-turn short-circuit fault conditions is established, and its stator voltage equation is expressed in the dq coordinate system as follows: ,in, Stator voltage, For stator current, For stator resistance, For inductance, Electric angular velocity, For permanent magnet flux linkage; If an inter-turn short circuit occurs in one phase, the current expression can be modified as follows: ,in, This represents the phase current under inter-turn short-circuit fault conditions. This represents the phase current under normal operating conditions. This is the severity factor of the fault; The current is recorded synchronously with a unified sampling frequency and time. Magnetic Link ,temperature With vibration signal It outputs a multi-channel signal matrix. It serves as the direct input for subsequent signal processing and feature extraction modules.

3. The method for identifying inter-turn short-circuit faults in a permanent magnet synchronous motor according to claim 2, characterized in that, In step 3, the convolutional neural network extracts two-dimensional time-frequency features through multiple convolution and pooling operations, and automatically learns the spatial distribution pattern of the signal: ,in, This represents the output feature map of the convolutional layer. Represents the convolution kernel weight matrix. Indicates the input feature map, denoted as the bias term, i and j represent the position indices of the output feature map of the convolutional layer in two spatial dimensions, respectively, and m and n represent the summation indices of the convolutional kernel in the corresponding dimensions during the convolution operation; Long Short-Term Memory (LTSM) networks retain the temporal features of motor operation over a long period, model the temporal correlation of feature sequences, capture time-series dependencies, and avoid gradient vanishing. The state equation is as follows: ,in, Indicates the output of the forget gate. This represents the input features at time t. This indicates the hidden state in the previous moment. Indicates the current hidden state. This indicates the current state of the memory unit. Indicates the memory state of the previous moment. Indicates the input gate output. Indicates the output gate output. This represents the Sigmoid activation function. This represents the Tanh activation function. Indicates candidate memory units. Represents the weight matrix of each gate. , Indicates the bias terms for each gate. This represents the weight matrix from the hidden state of the previous time step to the current gate structure; Attention mechanisms focus on key time segments and weight these segments to enhance significant fault characteristics. , , ,in, This represents the attention weight at time t. This indicates the score for attention. This represents the feature vector at time t of the LTSM output. Represents the attention score vector. Indicates the length of the time series. This represents the attention-weighted vector.

4. The method for identifying inter-turn short-circuit faults in a permanent magnet synchronous motor according to claim 3, characterized in that, Attention-weighted vector Input to the fully connected layer to obtain the logits vector. , The class probabilities are obtained through Softmax. In step 4, the classification layer uses the Softmax function to output multiple fault categories. ,in, This represents the predicted probability of the i-th class. This represents the network input of class i. Indicates the total number of categories. This represents the index of the Softmax normalized summation. Indicates the classification layer weights. This indicates the classification layer bias.

Images