Permanent magnet synchronous motor stator winding fault on-line identification method and system

Abstract

The invention discloses a permanent magnet synchronous motor stator winding fault online identification method and system, relates to the field of motor fault diagnosis, and solves the technical problem that high-precision motor fault diagnosis cannot be realized by directly extracting fault features from a three-phase current waveform in combination with model design in the prior art. The method comprises the steps of obtaining motor three-phase current data and performing preprocessing to obtain a data set; training and optimizing an SE-ResCNN model by using the data set, wherein the SE-ResCNN model comprises an input layer, a plurality of convolution modules, a residual module, an SE module, a global maximum pooling layer, an expansion layer and a SoftMax layer; converting the trained and optimized SE-ResCNN model into an ONNX format, deploying the SE-ResCNN model on edge equipment, collecting three-phase current data in real time, transmitting the three-phase current data to the edge equipment, and carrying out online diagnosis on a stator winding fault; according to the method, the fault features are directly extracted from the three-phase current waveform, the processes of other fault feature extraction and signal processing are effectively avoided, and global and local fault features are captured through channel fusion, so that the fault diagnosis precision is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of motor fault diagnosis, in particular to a permanent magnet synchronous motor stator winding fault online identification method and system. BACKGROUND

[0002] Permanent magnet synchronous motor (PMSM) is widely used in high-speed trains, renewable energy generation, large-scale manufacturing and other key fields due to its low loss, high efficiency, high power density and other excellent characteristics. This trend has prompted the study of PMSM fault diagnosis and reliability improvement to receive increasing attention. Previous studies in this field have shown that nearly one-third of the faults in PMSM are caused by internal stator winding failure, among which inter-turn short circuit fault (ISF), high resistance connection fault (HRC) and open circuit fault (OC) are classified as the most common fault modes in stator winding.

[0003] In the early stage of the fault, the stator fault characteristics are often weak. If not diagnosed and eliminated in time, these faults may gradually evolve into more serious secondary faults. For example, stator winding ground fault and inter-phase short circuit will cause significant damage to the motor control system. Therefore, designing an efficient stator winding fault detection and identification (FDI) strategy is crucial for subsequent fault-tolerant control and maintenance. Current motor stator winding fault diagnosis research mainly falls into three categories: signal-based methods, model-based methods and data-driven methods.

[0004] In existing research, the diagnosis methods for stator winding faults mainly fall into the following three categories: signal-based methods, model-based analysis methods and data-driven methods. Among them, the signal-based method collects various fault signals, including stator current, zero sequence voltage signals and motor torque signals, and then extracts fault features using specific signal processing methods. However, this method is too dependent on operating conditions and is difficult to apply in actual industries. The model-based diagnosis method is based on the mathematical logical relationship between the parameters of the permanent magnet motor to establish an accurate fault motor model. The difference between the output of the fault model and the actual output of the motor is calculated, and the difference is compared with the predefined diagnostic threshold to detect motor faults. Although the model-based method is easy to analyze, the diagnostic accuracy is highly dependent on the related parameters of the model. As the motor system becomes more complex, it is difficult to establish an accurate mathematical model to monitor and identify motor faults.

[0005] With the development of big data and artificial intelligence technology, data-driven based diagnosis methods are increasingly popular in motor health diagnosis and monitoring. The main advantage is that it uses big data and advanced algorithms to achieve higher accuracy and speed, thereby effectively reducing the risk of failure and maintenance costs. However, there are still some shortcomings: in the early stage of stator winding failure, it is difficult to accurately diagnose the winding failure due to the weak winding failure characteristics, which will affect the accuracy of the diagnosis results. Existing data-driven methods mainly focus on single fault detection, and each diagnosis method is only for a specific fault type. When the actual fault does not match the expected scenario, such specialized methods will fail, which may lead to diagnostic errors and further damage to PMSMs. In addition, most current diagnosis methods use offline operation mode, relying only on signal analysis after PMSM failure, which essentially hinders the realization of real-time monitoring and fault diagnosis. SUMMARY

[0006] In order to solve the problems existing in the prior art, the present application provides a permanent magnet synchronous motor stator winding fault online identification method and system, which solves the technical problem that the prior art cannot directly extract fault features from three-phase current waveforms to realize high-precision motor fault diagnosis in combination with model design.

[0007] A permanent magnet synchronous motor stator winding fault online identification method, comprising:

[0008] Obtaining three-phase current data of the motor under various operating conditions and preprocessing to obtain a data set;

[0009] Training and optimizing the SE-ResCNN model using the data set, the SE-ResCNN model comprising an input layer, a plurality of convolution modules, a residual module, an SE module, a global maximum pooling layer, an expansion layer, and a SoftMax layer;

[0010] The SE-ResCNN model trained and optimized is converted to ONNX format and deployed on an edge device, real-time acquisition of three-phase current data is transmitted to the edge device, and online diagnosis of stator winding faults is performed.

[0011] Further, the residual module realizes a double-branch architecture, one branch of which maintains the original input features through identity mapping, while the complementary branch performs nonlinear feature transformation through convolution operation. The SE module then recalibrates the feature maps output by the residual module channel by channel, with the SE module adaptively weighting each channel according to the learned importance.

[0012] Further, the SE module includes a squeeze module that reduces the feature dimension using average pooling, a fully connected layer and a ReLU layer that parameterize the channel selection mechanism and select important features; finally, the initial features are multiplied by the optimized feature weights to obtain enhanced features.

[0013] U * =AvgPooling(U)∈R 1×1×C

[0014] wherein AvgPooling(·) is average global pooling;

[0015] Then, the compressed features will be re-weighted.

[0016] V=Sigmoid(f2(Relu(f1·U * )))∈R 1×1×C

[0017] wherein f1(·) represents the first FC layer, f2(·) represents the second FC layer, Relu(·) and Sigmoid(·) represent the Relu activation function and the Sigmoid activation function respectively; afterwards, the feature weight is re-distributed by matrix multiplication:

[0018]

[0019] wherein x=1,2,3…,H,y=1,2,3…,W,z=1,2,3…,C

[0020] Further, the cross-entropy function is used as the cost function in the training optimization process, which is defined as:

[0021]

[0022] wherein J(w,b) is the cost function, M is the number of samples, K is the number of fault types, U m,k is the true label of the mth sample, U* m,k is the probability of the mth sample predicted by the model belonging to class k.

[0023] Further, the edge device is a domestic RK3588 edge device.

[0024] Further, a fault simulation platform is built to obtain three-phase current data of the motor under various operating conditions, the fault simulation platform includes a permanent magnet synchronous motor and a fault setting board, the fault setting board realizes stator winding faults of different fault degrees through air switches and stator resistors based on equivalent circuit diagrams of fault types, the fault types include stator winding inter-turn short circuit, high resistance connection fault, and winding open circuit fault, and three-phase current data of the permanent magnet synchronous motor is subjected to the fault setting board to obtain specific motor fault signals under different fault states.

[0025] Further, for the stator winding inter-turn short circuit fault, the permanent magnet synchronous motor is taken as the test object, the stator winding adopts a special winding process, the tap interfaces are reserved at the key positions of each phase winding, the taps are led out at the 5th, 10th and 15th turns of the A, B and C three-phase windings, the external resistance R f is connected in the A, B and C three-phase windings, and the neutral point forms a short circuit loop; for the stator winding high resistance connection fault, the resistance R d is connected in the A, B and C three-phase windings, and the neutral point forms a short circuit loop; for the stator winding high resistance connection fault, the resistance R

[0026] A permanent magnet synchronous motor stator winding fault online identification system, comprising a data acquisition and processing module, an edge device and an upper computer, the data acquisition and processing module is used for real-time acquisition of motor three-phase current data and pretreatment, the edge device is deployed with a trained and optimized SE-ResCNN model, the edge device and the upper computer are used for online diagnosis of stator winding faults based on the pretreated three-phase current data.

[0027] The beneficial effects of the present application include:

[0028] 1. The proposed diagnostic method directly extracts fault features from three-phase current waveforms, effectively avoiding the process of other fault feature extraction and signal processing;

[0029] 2. The SE attention mechanism is combined with the residual CNN branch, and the global and local fault features are captured through channel fusion, so that the fault diagnosis accuracy is improved;

[0030] 3. While ensuring the diagnosis accuracy and reasoning speed, the edge device is used to realize real-time diagnosis by combining model conversion and migration technology. In addition, by using the domestic RK3588 chip along with the model conversion and migration technology, the designed network realizes efficient edge deployment, ensures high diagnosis accuracy and real-time processing capability. The experimental results on the PMSM driving system platform show that the average diagnosis accuracy is more than 98%, and the diagnosis speed reaches the industrial application standard, verifying the effectiveness of the method, which shows that the method can meet the real-time fault detection industrial requirements. In addition, the method is also applicable to other models, and has important reference value for industrial applications. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 It is an equivalent current diagram for the inter-turn short circuit fault involved in the embodiments of the present application.

[0032] Figure 2 It is an equivalent circuit diagram for the high resistance connection fault involved in the embodiments of the present application.

[0033] Figure 3A control chart of the PMSM driver and the fault current waveform involved in the embodiments of the present application.

[0034] Figure 4 A schematic diagram of the normalization process involved in the embodiments of the present application.

[0035] Figure 5 A structure diagram of the SE-ResCNN diagnostic model involved in the embodiments of the present application.

[0036] Figure 6 A structure diagram of the SE module involved in the embodiments of the present application.

[0037] Figure 7 A schematic diagram of the FC layer and the AMP layer involved in the embodiments of the present application.

[0038] Figure 8 A flowchart of the model training algorithm of the method involved in the embodiments of the present application.

[0039] Figure 9 A confusion matrix of the simulation verification diagnostic result of the SE-ResCNN involved in the embodiments of the present application.

[0040] Figure 10 A visualization of the SE-ResCNN model classification result based on t-SNE involved in the embodiments of the present application.

[0041] Figure 11 A schematic diagram of the stator winding fault setting involved in the embodiments of the present application.

[0042] Figure 12 A stator winding fault experimental platform involved in the embodiments of the present application.

[0043] Figure 13 Diagnostic results of the 2nd data set in different networks involved in the embodiments of the present application; (a) CNN (b) SE-CNN (c) Res-CNN (d) SE-ResCNN.

[0044] Figure 14 Diagnostic results of the four networks under different working conditions involved in the embodiments of the present application.

[0045] Figure 15 Experimental diagnostic results under different working conditions involved in the embodiments of the present application; (a) Model 1, (b) Model 2.

[0046] Figure 16 A flowchart of the online diagnostic strategy involved in the embodiments of the present application. DETAILED DESCRIPTION

[0047] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0048] Example 1

[0049] The following is in conjunction with the appendix Figures 1-16 Specific embodiments of the present invention will be described in detail;

[0050] The fault mechanisms of the stator windings were analyzed and modeled, including inter-turn short circuit faults, high-resistance connection faults, and open circuit faults.

[0051] like Figure 1 The diagram shown is the equivalent circuit diagram for an inter-turn short-circuit fault. The number of short-circuit turns is defined as n, and the total number of turns in one phase is N. The fault severity factor is thus defined as u = n / N. f It can be described as flowing through the fault resistor R f The fault current. When an ISF fault occurs in the PMSM, the voltage equation V f It can be represented as:

[0052]

[0053] Where R sf i sf V f e sf Represented as

[0054]

[0055] i sf =[i a i b i c i f ] T

[0056] V f =[V a V b V c ,0] T

[0057]

[0058] In the formula: V f It is the phase voltage matrix of the stator winding, R sf Let i be the resistance matrix. sf For the current matrix, e sf It is the back electromotive force matrix of the three-phase stator winding, R a1 R is the resistance of stator winding a1. a2 R is the resistance of stator winding a2. b R is the resistance of stator winding b; c Let i be the resistance of the stator winding c. a i is the current flowing through the A-phase stator winding. b i is the current flowing through the B-phase stator winding. c i is the current flowing through the C-phase stator winding. f R is the current flowing through the short-circuited conductor; f V is the short-circuit resistance value, and V0 is the zero-sequence voltage; V a V is the voltage through the A-phase stator winding. b V is the voltage through the B-phase stator winding. c e is the voltage through the C-phase stator winding. a1 e is the back electromotive force of stator winding a1. a2 e is the back electromotive force of stator winding a2. b e is the back electromotive force of stator winding b. c e is the back electromotive force of stator winding c. f This is the back electromotive force on the short-circuited conductor.

[0059] In addition, inductor L f The expression is:

[0060]

[0061] When a permanent magnet synchronous motor (PMSM) experiences an inter-turn short circuit (ISF) fault, the expression for the electromagnetic torque is:

[0062]

[0063] Among them, La1, La2, L b and L c M represents the self-inductance of stator windings a1, a2, b, and c. j,k (j,k={a1,a2,b,c}) represents the mutual inductance between stator windings j and k; ω r This represents the mechanical angular velocity of the motor.

[0064] like Figure 2 The diagram shown is the equivalent circuit diagram for a high-resistance connection fault (HRC). For a surface-mounted permanent magnet synchronous motor, the three-phase stator windings have the same resistance value, and the inductance between the three-phase stator windings is also the same. Rx (x = a, b, c) equals R, M y (y = ab, bc, ca), likewise. HRC can be equivalently simplified as a form of a certain winding string in series with an additional power supply.

[0065] When a high resistance connection fault (HRC) fault occurs, the equation of the fault permanent magnet synchronous motor A phase is represented as:

[0066]

[0067] where, u a , u b and u c are the three-phase voltages of the stator winding, u0 is the voltage between the DC midpoint of the inverter and the neutral point center of the stator winding, i a , i b and i c are the three-phase stator currents, u xo (x = a, b, c) is the difference between u x and u0. R is the stator phase resistance, ΔR is the fixed resistance of the additional high resistance, L x (x = a, b, c) is the stator phase self-inductance of A, B and C phases, and M is the mutual inductance between the stator windings.

[0068] In addition, the open circuit (OC) fault can be effectively modeled by making ΔR approach infinity, thereby forcing the phase current to asymptotically approach zero. This modeling method serves two key goals. First, it physically characterizes the open circuit fault by the infinite impedance resulting from circuit discontinuity, and establishes a controlled experimental benchmark for evaluating the performance of the diagnostic system under complete current interruption conditions. Intentionally constructing this extreme fault scenario can strictly verify the ability of the network to detect and characterize a complete open circuit fault under theoretical ideal conditions.

[0069] After that, the verification of the above fault analysis is realized in MATLAB / Simulink, and the PMSM drive system adopts the field-oriented control (FOC) strategy, as shown in Figure 3 . To obtain three-phase current data under various operating conditions. Current waveform analysis shows that stator winding faults will cause significant three-phase current distortion, which adversely affects the stability of motor operation. Early fault conditions, especially ISF and HRC, show similar weak features, which pose a challenge for reliable differentiation. Without timely identification and targeted intervention, these early faults will gradually degenerate into OC faults, which may lead to catastrophic PMSM damage. The simulation results prove the accuracy of the fault model and provide theoretical support for the subsequent construction of an experimental platform. And provide data support for subsequent permanent magnet synchronous motor fault diagnosis and classification.

[0070] By analyzing the simulation model of stator winding fault and accurately characterizing the fault features, the application proposes an online identification method for permanent magnet synchronous motor stator winding fault, which realizes accurate diagnosis and identification of stator winding inter-turn short circuit, high resistance connection fault and winding open circuit fault through multi-branch residual connection and channel attention mechanism. The model has the following three characteristics: by introducing residual connection, the network can efficiently learn global features, avoid gradient disappearance problem of deep network, and improve the model's ability to capture overall features; a multi-branch parallel CNN structure is adopted, each branch focuses on local feature extraction at different levels, and the sensitivity of the model to local features is enhanced; through channel attention mechanism, the features of different branches are dynamically weighted and fused, realizing adaptive optimization and fusion of cross-branch features, and improving the diagnostic accuracy.

[0071] Specifically, an online identification method for permanent magnet synchronous motor stator winding fault includes the following steps:

[0072] (1) Data preprocessing

[0073] In most cases, the collected three-phase current data is essentially continuous one-dimensional time series data. However, the current amplitude of different faults may not belong to the same unit dimension under various loads and speeds, and abnormal samples may be generated, which may have an adverse effect on training the SE-ResCNN model. Before inputting the model, the original current signal will be preprocessed by standardization, in which the min-max normalization scales the fault data to the range of [-1, 1]. Its expression is as follows:

[0074]

[0075] In the formula, max(x) is the maximum value of the data, and min(x) is the minimum value of the data

[0076] This transformation also reduces the amplitude distortion caused by operating conditions and improves the gradient optimization efficiency during network training, as shown in Figure 4 First, the three-phase current signals of the motor collected in the simulation model are normalized to the range of [-1, +1] by the min-max normalization method, which eliminates the influence of working condition changes on diagnostic accuracy and performance. Then, the data samples are divided into training set, cross-validation set and test set according to a certain proportion, which are used for training, model parameter design and evaluation of model diagnostic accuracy and performance, respectively.

[0077] (2) Model structure design

[0078] The proposed SE-ResCNN model is as follows Figure 5As shown, it is realized by applying an input layer, four convolution modules, a channel attention mechanism, a global max pooling and an expansion layer, and a SoftMax layer. Each convolution module contains a convolution layer, a batch normalization layer, an activation function layer and a max pooling layer. Overall, compared with the traditional regular CNN model, the significant advantage of the proposed model lies in the introduction of residual module, channel attention mechanism and global max pooling layer. Therefore, through sufficient feature extraction, the diagnostic accuracy will be improved, and through improving the gradient flow and reducing the number of parameters, the training speed can be improved.

[0079] During network training, the increase in structural complexity in deep networks can cause gradient vanishing during backpropagation, thereby impairing model convergence and training stability. To solve this problem, the present embodiment adopts a residual network architecture with skip connections to establish a direct path for input information across layers, thereby mitigating information degradation. This design effectively alleviates gradient vanishing and training difficulties, accelerating network optimization. In a deep network, the output is obtained by a direct nonlinear transformation, expressed as

[0080] y = F(x) = W2gσ(W1gx + b1) + b2

[0081] where F(x) is the output of the current layer, x is the input, W1 and W2 are weight matrices, b1, b2 are bias terms, and σ is the activation function.

[0082] In the residual module, a direct path is introduced, in which the input and output are directly combined by skip connection. This architecture effectively prevents information loss while improving training efficiency and model performance. The expression of skip connection is

[0083] y = F(x) + x1

[0084] Since the residual connection realizes a dual-branch architecture, one branch preserves the original input features through identity mapping, while the complementary branch performs nonlinear feature transformation through convolution operation. The feature maps from the two branches are concatenated along the channel dimension. Subsequently, the feature maps are recalibrated channel by channel through the SE module, which adaptively weights each channel according to the learned importance. This adaptive weighting mechanism enhances discriminative feature channels while suppressing less relevant ones, thereby optimizing feature representation and improving network robustness.

[0085] The structure of the SE block is shown as Figure 6 The input is U = R H×W×C , the feature dimension is reduced by using average pooling in the squeeze module, and the input is changed to U* = R 1×1×CNext, the channel selection mechanism is parameterized and important features are selected through two fully connected (FC) layers and one ReLU layer. Finally, the initial features are multiplied by optimized feature weights to obtain enhanced features.

[0086] U * =AvgPooling(U)∈R 1×1×C

[0087] AvgPooling(·) is the average global pooling. Then, the compressed features will be reweighted.

[0088] V = Sigmoid(f2(Relu(f1·U)) * )))∈R 1×1×C

[0089] Where f1(·) represents the first fully connected (FC) layer, f2(·) represents the second FC layer, and ReLU(·) and Sigmoid(·) represent the ReLU activation function and the Sigmoid activation function, respectively. Then, the feature weights are redistributed through matrix multiplication:

[0090]

[0091] Where x = 1, 2, 3…, H, y = 1, 2, 3…, W, z = 1, 2, 3…, C

[0092] like Figure 7 As shown, a fully connected (FC) layer first expands a multi-dimensional feature map into a one-dimensional vector, and then maps this one-dimensional vector to the output node using a weight matrix. Each node in the network represents a parameter. As the size of the input feature map increases, the length of the expanded vector increases significantly. This leads to an exponential increase in the number of parameters in the fully connected layer. Therefore, an AMP layer is used instead of a FC layer. By merging other processes, no parameters are required, reducing the computational load and thus improving training efficiency.

[0093] (3) Fault Diagnosis

[0094] The diagnostic process based on SE-ResCNN is as follows: Figure 8 The process is implemented as follows: Before training, the relevant parameters of the convolutional layers in the SE-ResCNN model, such as weights and biases, are randomly initialized. Then, data from the training set is fed into the model, which propagates forward through layers to compute the predicted value U* for each sample. The prediction is then calculated using a loss function between the classification layer output U* and the true label U. Specifically, the cross-entropy function is used as the cost function, defined as:

[0095]

[0096] where J(w, b) is the cost function, M is the number of samples, K is the number of fault types, U m,k is the true label of the mth sample, U m,k is the probability of the mth sample belonging to class k predicted by the model. Then, it is determined whether the number of iterations has reached the maximum epoch (i > M). If so, the training process is terminated. Otherwise, the backpropagation process continues and the next round of training is performed. During this process, the choice of optimization algorithm is crucial to the effectiveness and efficiency of network training.

[0097] Therefore, the AdamW algorithm with decoupled weight decay is adopted as the optimizer for training the SE-ResCNN model. This algorithm is an improved algorithm based on the L2 regularization Adam algorithm. Therefore, this method reduces the adverse effects of weight decay on bias parameters while improving the efficiency and efficacy of model training. Finally, the test set is input into the trained model to evaluate the accuracy and generalization ability of the trained SE-ResCNN model. The test results show that the model has the ability to identify different fault types and different fault locations of the stator winding.

[0098] (4) Simulation verification:

[0099] A fault model of PMSM is established in the Matlab / Simulink environment, and a three-phase current dataset for network training is generated. The simulation systematically records the three-phase current data under various stator winding fault conditions, and the corresponding fault labels are listed in Table 1.

[0100] Table 1 Fault categories and corresponding labels

[0101]

[0102] Afterwards, the data under different fault conditions are extracted, and the model is trained and tested. Specifically, the PMSM control speed is 300 r / min, the load torque is 5 N.m, and the sampling frequency is set to 10 kHz. Considering that the motor cannot run for a long time under fault conditions, the simulation model is only set to a sampling time of 2 s, so each raw data set collected is a two-dimensional matrix of [20000, 3]. By introducing the Min-Max normalization function as described earlier, the collected data set is normalized to [-1, 1]. Then the number of samples is expanded using random overlap sampling, with a window size of 1000 and a step size of 50. Therefore, the fault data set of each condition is divided into 381 samples of size [1000, 3], each sample has 10 labels, and the total data set is 3810. Before starting to input the data set to the network architecture, the 2-D time series sample [1000, 3] is reconstructed into a three-channel 1-D sample [1000, 1] as an input feature map. Among them, 60% of the total sample data is used as the training set, 20% is used as the validation set, and the remaining 20% is used as the test set. The effectiveness of the proposed method is then verified, and the confusion matrix and diagnostic accuracy of the test set are analyzed, as shown in Figure 9 , where the horizontal coordinate represents the predicted fault label and the vertical coordinate represents the true fault class label. There are 768 test samples, and the red nodes represent the true value and the blue nodes represent the predicted value of the sample. Through the calculation of all test samples, the fault diagnosis accuracy can reach 99.74%. Specifically, among all samples, only two healthy samples are misclassified as HRC samples. This misclassification may be due to the fact that the three-phase current has not yet stabilized when the motor is in the initial start-up state. The results show that the method has strong superior detection and recognition ability for the three typical stator winding faults.

[0103] To verify the classification ability of the designed SE-ResCNN, the data distribution of the 10 learned features in different layers is mapped to a three-dimensional space by the (t-SNE) algorithm, as shown in Figure 10 . From the results, it can be seen that the original input signal shows the most severe feature overlap. After processing through the first convolutional module, some fault features begin to form clusters, although the boundaries between them are still quite blurred. When propagated through the second and third convolutional modules, the degree of confusion between the 10 fault categories is greatly reduced, and their respective clusters become more dense. Although there is still some category mixing, the subsequent SE attention module successfully achieves complete separation of all fault labels. This progress proves the excellent diagnostic ability of the designed network architecture.

[0104] (5) Online deployment

[0105] Figure 16 The implementation of the proposed online diagnosis of stator winding faults is summarized. First, seeFigure 4 The raw data is normalized. At the same time, as shown in Figure 5 The framework of SE-ResCNN is designed in Pytorch. On this basis, the structure and hyperparameters of SE-ResCNN are automatically optimized by using the Optuna framework to minimize the loss function. Then, by using model conversion and migration technology, the trained network model is converted into ONNX format and deployed on RK 3588 edge device. Finally, the three-phase current data is transmitted to the edge computing device in real time, and the online diagnosis of stator winding fault is carried out.

[0106] In order to verify the performance of the proposed online diagnosis method, in Table 2, 100 diagnosis tests are carried out for each fault condition under the conditions of group 1, group 2 and group 3. Then the average inference latency and diagnosis accuracy are recorded and analyzed. It is worth noting that the real-time measured data sampling frequency is 10 kHz and the sampling time is 0.1 seconds. After each group of data is collected, it is transmitted to the edge device for real-time diagnosis, and the corresponding diagnosis results are summarized in Table 3. Due to electromagnetic interference between the motor and the current clamp, the diagnosis accuracy of some fault labels is affected, resulting in distortion of current measurement. In addition, the transient current irregularity during motor start-up can cause misclassification errors. Overall, in addition to misdiagnosis caused by external factors, the proposed online diagnosis method shows strong real-time performance, with an average diagnosis accuracy of more than 98% under all verified fault conditions.

[0107] Table 2 Fault data conditions Conditions

[0108]

[0109] Table 3 Online diagnosis accuracy and diagnosis time

[0110]

[0111] In another embodiment, a fault simulation platform is built to obtain three-phase current data of the motor under various operating conditions, the fault simulation platform includes a permanent magnet synchronous motor and a fault setting board, the fault setting board realizes different degrees of stator winding faults through air switches and stator resistors based on equivalent circuit diagrams of fault types, the fault types include stator winding inter-turn short circuit and high resistance connection fault, and the three-phase current data of the permanent magnet synchronous motor is obtained through the fault setting board to obtain specific motor fault signals.

[0112] Through the analysis of the fault mechanism, due to the complex internal structure of the stator winding of the permanent magnet synchronous motor, the stator winding fault often presents nonlinear characteristics in actual operation, and it is difficult to accurately reproduce the dynamic evolution process of the real fault through theoretical calculation or finite element simulation. In addition, potential faults such as winding insulation degradation and partial discharge are more difficult to trigger controllably under conventional experimental conditions, which brings challenges to fault mechanism research and diagnostic algorithm verification. In view of this problem, a high-precision permanent magnet motor stator winding fault simulation platform is designed and built. Figure 11 The permanent magnet synchronous motor stator winding fault simulation diagram built is shown, and according to the simplified equivalent circuit diagram in the first part, the fault setting board is set, and the stator winding fault of different fault degrees is realized through air switch and stator resistance. The realized fault circuit reproduces the simulation circuit described above, where Rf represents the resistance of the short-circuit branch. The short-circuit ratio u is set to 5%, 10% and 15%, and Rd represents the high resistance value during the fault condition. The platform adopts a modular architecture, realizes flexible simulation of stator winding inter-turn short circuit and high resistance connection fault through fixed value resistance and air switch, and supports accurate adjustment of fault degree (such as short-circuit turn ratio). The experimental motor adopts a special winding process and reserves fault injection points at key positions to ensure controllability and repeatability of fault introduction.

[0113] Figure 12 For the built stator winding fault experimental platform, the PMSM used in the experimental platform is a healthy permanent magnet synchronous motor, and is directly connected to the magnetic powder brake through the shaft coupling. The PMSM parameters used for experimental verification are shown in Table 2, the drive controller is a frequency converter, which can realize the FOC control algorithm. The collected data are connected to the oscilloscope through the current clamp for visualization and storage, and are matched with the edge diagnosis device and the upper computer to realize real-time monitoring and online diagnosis of the stator winding. For the stator winding inter-turn short circuit fault, the permanent magnet synchronous motor is taken as the test object, and the stator winding adopts a special winding process, with a tap interface reserved at the key position of each phase winding. The tap is led out at the 5th, 10th and 15th turns of the A, B and C phase windings, and a short-circuit loop is formed by connecting an external resistance Rf and the neutral point. For example, when the inter-turn short-circuit ratio u is 5%, the short-circuit resistance R f The inter-turn short-circuit fault of the early stator winding is simulated by connecting the A-phase winding to the first 5% of the inter-turn short-circuit. Moreover, through the air switch, the occurrence of the short-circuit fault and the fault degree can be controlled in real time. For the high resistance connection fault of the stator winding, a fixed value resistance R dThe stator resistance organization can be replaced according to the actual situation, and the occurrence of faults is controlled in real time through air switches. In addition, considering the stator winding failure of the permanent magnet synchronous motor under high-speed or heavy load conditions, a larger fault current will be generated, which will cause irreversible damage to the motor winding and permanent magnet. Therefore, the experiment is carried out under low-speed and light load conditions to ensure the safety of the equipment. Then verify the effectiveness of the platform, collect the basic running data of the motor under healthy state, and then gradually introduce different categories of preset faults, and compare and analyze the change law of fault characteristic quantities. The experimental results show that the platform can stably generate experimental data with typical fault characteristics, providing a reliable data source for the training and testing of subsequent intelligent diagnosis algorithms. This research not only provides a new experimental method for motor fault mechanism analysis, but also lays a solid foundation for the development of online monitoring systems.

[0114] After collecting data under different working conditions, in this work, 5 groups of experiments were designed to verify the robustness of the proposed SE-ResCNN model and the three-phase current of the motor as the collected data. See Table 2 for specific experimental parameters. In addition, the sampling frequency and time are consistent with the simulation, and then the data is processed by normalization.

[0115] In order to systematically evaluate the performance of the designed network, a series of ablation experiments were performed by removing each module step by step. In order to make a fair comparison, we trained and validated four different network architectures (CNN, SE-CNN, Residual-CNN and SE-ResCNN) on the same dataset, and verified their diagnostic accuracy.

[0116] Figure 13 The test results of the four comparison networks after hyperparameter optimization on the same dataset under the second group of datasets are shown. The results show that the accuracy of the proposed SE-ResCNN model is as high as 98.90%, which is higher than other models, further proving the superiority of the proposed model. Considering the randomness of parameter initialization and the randomness of data under a single condition, the data under each experimental condition is trained and tested. The diagnostic results are shown in Figure 14 As can be seen from the average accuracy of the 5 groups of experiments, the accuracy of this model is 98.58%, which is higher than other models. As Figure 14As shown, the traditional CNN architecture presents a too simple structure, which limits its ability to comprehensively capture the stator winding fault features, resulting in a relatively low accuracy of 95.5%. Although the SE attention mechanism enhances the network training performance and accelerates the convergence, its effectiveness in accurately identifying the stator winding fault features in the test set is still limited. In contrast, the Res-CNN architecture, due to its residual connection, exhibits superior superior learning ability compared to the above-mentioned models. The integration of the SE module with the residual connection module effectively addresses the overfitting problem inherent in traditional CNNs. Experimental results show that, combined with the SE module and residual connectivity, along the merging of CNN branch and residual branch channels, not only eliminates network overfitting, but also improves training efficiency and diagnostic accuracy.

[0117] In actual motor drive systems, the stator winding fault detection and identification method should be able to adapt to different operating conditions, including different loads and speeds, without false diagnosis. Therefore, experiments were conducted under different load torques and speed conditions. As shown in Table 4, the diagnostic performance was verified using the cross-validation method. In this case, Model 1 and Model 2 were specifically designed to evaluate the generalization ability of the SE-ResCNN architecture under different operating scenarios, Model 1 verifies the generalization ability under different speeds, and Model 2 verifies the generalization ability under different load torques. The results of the diagnosis are shown in Figs. Figure 15 (a) and 15(b). As can be seen from the classification confusion matrix, the accuracy of stator fault under Model 1 is over 97.6%, and some healthy samples are misdiagnosed, which may be due to the fact that the three-phase current does not initially reach a smooth state when the motor is running. Overall, under different speeds and loads, the average diagnostic accuracy is 97.9%. This indicates that the trained SE-ResCNN model can accurately locate and identify faults under variable load and speed conditions.

[0118] Table 4 Generalization ability verification dataset

[0119]

[0120] In addition, in order to verify the real-time diagnosis of the proposed method, by controlling Figure 11 the air switch in the fault setting, the PMSM can instantaneously experience different types of faults, and the diagnostic results are listed in Table 5. When the fault occurs instantaneously, the proposed stator winding fault detection and identification method is activated to locate and identify the fault type in real time.

[0121] Table 5 Real-time diagnostic results

[0122]

[0123] In another embodiment, a permanent magnet synchronous motor stator winding fault online identification system is provided, comprising a data acquisition and processing module, an edge device and an upper computer, the data acquisition and processing module is used for real-time acquisition of motor three-phase current data and pretreatment, the SE-ResCNN model trained and optimized is deployed in the edge device, and the edge device and the upper computer are used for online diagnosis of stator winding fault based on the pretreated three-phase current data.

[0124] The above embodiments only express the specific implementation of the present application, and the description is more specific and detailed, but it cannot be understood as a limitation on the protection scope of the present application. It should be pointed out that for ordinary skilled in the art, without departing from the technical concept of the present application, a number of modifications and improvements can be made, which belong to the protection scope of the present application.

Claims

1. A method for on-line identification of stator winding faults in permanent magnet synchronous machines, characterized in that, The method comprises the following steps: Obtain three-phase current data of the motor under various operating conditions and preprocess the data to obtain a data set; Train and optimize the SE-ResCNN model using the data set, wherein the SE-ResCNN model comprises an input layer, a plurality of convolution modules, a residual module, an SE module, a global maximum pooling layer, an expansion layer, and a SoftMax layer; Convert the SE-ResCNN model trained and optimized into an ONNX format and deploy it on an edge device, collect three-phase current data in real time and transmit it to the edge device, and perform online diagnosis of stator winding faults.

2. The method according to claim 1, characterized in that, The residual module realizes a double-branch architecture, one branch of which maintains the original input features through identity mapping, and the complementary branch performs nonlinear feature transformation through convolution operation, and then the SE module recalibrates the feature maps output by the residual module channel by channel, and the SE module adaptively weights each channel according to the learned importance.

3. The method according to claim 1, characterized in that, The SE module comprises: a squeeze module that reduces the feature dimension by using average pooling; a fully connected layer and a ReLU layer that perform a parameterized channel selection mechanism and select important features; and multiplying the initial features by the optimized feature weights to obtain enhanced features. U * = AvgPooling(U) ∈ R 1×1×C Wherein AvgPooling(·) is an average global pooling; The enhanced features will be reweighted: V = Sigmoid(f2(Relu(f1 · U * ))) ∈ R 1×1×C Wherein f1(·) represents the first FC layer, f2(·) represents the second FC layer, Relu(·) and Sigmoid(·) represent the Relu activation function and the Sigmoid activation function, respectively; then, the feature weights are redistributed through matrix multiplication: Where x = 1, 2, 3…, H, y = 1, 2, 3…, W, z = 1, 2, 3…, C 4. The method according to claim 1, characterized in that, The cross-entropy function is used as the cost function during the training and optimization process, which is defined as: where J(w, b) is the cost function, M is the number of samples, K is the number of failure types, U m,k is the true label of the mth sample, U m,k is the probability that the mth sample belongs to class k predicted by the model.

5. The method according to claim 5, characterized in that, The edge device is a domestic RK3588 edge device.

6. The method according to claim 1, characterized in that, A fault simulation platform is built to obtain three-phase current data of the motor under various operating conditions, the fault simulation platform comprises a permanent magnet synchronous motor and a fault setting board, the fault setting board realizes different fault degrees of stator winding faults through air switches and stator resistors based on the equivalent circuit diagram of the fault type, the fault type includes stator winding inter-turn short circuit, high resistance connection fault, and open circuit fault, and the three-phase current data of the permanent magnet synchronous motor is obtained through the fault setting board to obtain specific motor fault signals under different fault states.

7. The method according to claim 6, characterized in that, For the stator winding inter-turn short circuit fault, the permanent magnet synchronous motor is taken as the test object, the stator winding adopts special winding process, the tap interface is reserved at the key position of each phase winding, the external resistance R f is connected with the neutral point to form a short circuit loop; for the high resistance connection fault of the stator winding, the resistance R d of a certain value is connected in the A, B and C three-phase windings, and the stator resistance organization can be replaced according to the actual situation, and the occurrence of the fault is controlled in real time through the air switch.

8. An online identification system for stator winding faults of a permanent magnet synchronous motor, comprising a data acquisition and processing module, an edge device, and an upper computer, the data acquisition and processing module is used to collect three-phase current data of the motor in real time and preprocess the data, the edge device is deployed with the SE-ResCNN model trained and optimized in the online identification method for stator winding faults of a permanent magnet synchronous motor according to claims 1-7, and the edge device and the upper computer are used to diagnose stator winding faults based on the preprocessed three-phase current data.

Images