A synchronous motor fault diagnosis method based on MSK-CNN and multi-source electromechanical information fusion

Abstract

The present invention provides a synchronous motor fault diagnosis method based on MSK-CNN and multi-source electromechanical information fusion. The method collects phase voltage, rotor vibration, and stator vibration signals of the synchronous motor in real time under normal operating conditions and different rotor winding inter-turn short-circuit conditions, and divides them into training and test sets. The method also determines the optimal MSK-CNN model. The raw data of phase voltage, rotor vibration, and stator vibration are respectively input into the MSK-CNN model for training. After passing through the multi-scale feature extraction layer, the fault features extracted by the multi-scale kernel are fused and used as the output of each MSK-CNN sub-model. The method also fuses the features extracted from the three signals to achieve the diagnostic purpose. The present invention provides an end-to-end fault diagnosis solution for synchronous generators, eliminating the need for additional signal processing and expert experience, eliminating the influence of manually extracted features, and avoiding the occurrence of missed judgments, misjudgments, and other phenomena caused by single sensor failures. The accuracy of synchronous generator fault diagnosis is significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of synchronous motor fault diagnosis, and in particular to a synchronous motor fault diagnosis method based on MSK-CNN and multi-source electromechanical information fusion. Background Art

[0002] Synchronous motors are a critical link in energy conversion in power systems and are an essential component of the system. Their operating conditions are highly variable and their structures are complex. They are susceptible to external environmental interference, material aging, and other factors, leading to frequent failures. Failure to promptly detect and address these faults can deteriorate the operating condition of the generator bearings, exacerbate stator and rotor vibration, and induce magnetization of the main shaft, forcing the unit to shut down, threatening the safe operation of the generator and the entire power grid. Therefore, researching synchronous motor fault diagnosis methods and improving their accuracy are of great practical significance for the safe operation and maintenance of power plants and power grids.

[0003] Common electrical faults in synchronous motors include rotor inter-turn short-circuit faults, stator inter-turn short-circuit faults, and air gap eccentricity faults. Traditional fault diagnosis methods extract effective fault-related features from the raw signal from different perspectives, and then input these extracted features into some shallow machine learning algorithms to achieve fault diagnosis. However, the feature extraction and classification steps of these traditional methods are designed and executed separately, both of which affect the final diagnostic performance. Moreover, feature extraction requires the establishment of precise physical models and relies on expert prior knowledge. It can only be used for specific units and faults and cannot be updated online and well generalized to new diagnostic areas. With the rapid development of the machinery manufacturing industry, motor models are becoming more precise, complex, and automated, making it more difficult to design a specific and appropriate feature extractor for each unit.

[0004] In recent years, the rapid development of deep learning has provided new insights into intelligent fault diagnosis. Convolutional neural networks, as a fundamental model of deep learning, can automatically extract features from raw signals, reducing reliance on empirical evidence. This addresses the limitations of traditional methods, which rely on expert expertise, are time-consuming, and have poor generalization capabilities. These intelligent diagnostic algorithms are rarely used on synchronous motors, and most rely on a single sensor for diagnosis. A single signal provides only partial information about the fault, and single-signal acquisition is susceptible to factors such as weak fault signatures, sensor failure, and environmental variability, leading to misdiagnosis and missed detections. Furthermore, due to the large number of rotating components in motors, the acquired signals are often time series with strong periodicity and deep correlations. Hidden within this periodicity lies a wealth of valuable information and correlations. Simultaneously considering the relationships between signals at different scales can better utilize all the information in the signal and improve fault diagnosis accuracy. Summary of the Invention

[0005] The purpose of the present invention is to provide a synchronous motor fault diagnosis method based on MSK-CNN and multi-source electromechanical information fusion, which can capture the fault characteristics of the original signal from multiple scales, improve the feature learning ability, and avoid the occurrence of misjudgment, missed judgment and other phenomena caused by single sensor failure, thereby improving the fault diagnosis accuracy.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A synchronous motor fault diagnosis method based on MSK-CNN and multi-source electromechanical information fusion includes the following steps:

[0008] S1, using a data acquisition device to collect phase voltage, rotor vibration and stator vibration signals of the synchronous motor in real time under different operating conditions: normal operation, rotor winding inter-turn short circuit, stator inter-turn short circuit and static eccentricity;

[0009] S2, slicing the collected phase voltage, rotor vibration, and stator vibration signals with a set unit length, and setting them as training sets and test sets according to the set ratio;

[0010] S3: Establish the MSK-CNN model, determine the network structure and convolution kernel size of the MSK-CNN model, use the training set for training, adjust the MSK-CNN network structure and hyperparameters according to the loss function and accuracy changes in the iterative process, train again, and repeat the above process to determine the optimal MSK-CNN model;

[0011] In step S4, the raw data of phase voltage, rotor vibration, and stator vibration are input into the optimal MSK-CNN model respectively. After passing through the multi-scale feature extraction layer, the fault features extracted by the multi-scale kernel are fused and used as the output of each MSK-CNN sub-model.

[0012] In S5, in the optimal MSK-CNN model, the fault features extracted from the phase voltage, rotor vibration, and stator vibration signals are aggregated and input into the same Flatten layer for flattening. They are then passed to the Softmax output layer through the fully connected layer to finally obtain the probability value of each operating condition. The operating condition corresponding to the maximum probability value is considered to be the final fault diagnosis result.

[0013] Furthermore, in step S1, the rotor winding inter-turn short-circuit operating conditions include two types of rotor winding inter-turn short-circuit: 3% and 6%, the stator inter-turn short-circuit operating conditions include two types of stator inter-turn short-circuit: 3% and 6%, and the static eccentricity operating conditions include two types of static eccentricity: 3% and 6%.

[0014] Furthermore, in step S1, the data acquisition device includes a synchronous motor phase voltage signal detection device, a synchronous motor rotor vibration acceleration sensor, and a stator vibration acceleration sensor.

[0015] Furthermore, in step S1 and step S2, the setting ratio is 8:2.

[0016] Furthermore, in step S3, the optimal MSK-CNN model specifically includes: an input layer, a multi-scale kernel feature extraction layer, a fully connected layer, and an output layer; wherein the multi-scale kernel feature extraction layer is composed of multiple CNN sub-blocks with different convolution kernel sizes, and the CNN sub-block includes two alternating convolution layers CL, a normalized batch processing BN layer, an activation function ReLu layer, and a pooling layer PL. The convolution layer CL includes convolution layer CL1-1, convolution layer CL1-2, convolution layer CL2-1, convolution layer CL2-2, convolution layer CL3-1, and convolution layer CL3-2. The pooling layer PL includes pooling layer PL1-1, pooling layer PL1-2, pooling layer PL2-1, pooling layer PL2-2, pooling layer PL3-1, and pooling layer PL3-2.

[0017] Furthermore, the original signal size of the input layer input is 500×1;

[0018] The convolution kernel size of the convolution layer CL1-1 is set to 3×1, the stride is set to 1, the number of convolution kernels is 16, and the activation function is selected as the ReLU function;

[0019] The pooling layer PL1-1 adopts the maximum pooling method, the pooling area width is 2×1, and the number of convolution kernels is 16;

[0020] The convolution kernel size of the convolution layer CL1-2 is set to 3×1, the stride is set to 1, the number of convolution kernels is 32, and the activation function is selected as the ReLU function;

[0021] The pooling layer PL1-2 adopts the maximum pooling method, the pooling area width is 2×1, and the number of convolution kernels is 32;

[0022] The convolution kernel size of the convolution layer CL2-1 is set to 5×1, the stride is set to 1, the number of convolution kernels is 16, and the activation function is the ReLU function;

[0023] The pooling layer PL2-1 adopts the maximum pooling method, the pooling area width is 2×1, and the number of convolution kernels is 16;

[0024] The convolution kernel size of the convolution layer CL2-2 is set to 5×1, the stride is set to 1, the number of convolution kernels is 32, and the activation function is the ReLU function;

[0025] The pooling layer PL2-2 adopts the maximum pooling method, the pooling area width is 2×1, and the number of convolution kernels is 32;

[0026] The convolution kernel size of the convolution layer CL3-1 is set to 7×1, the stride is set to 1, the number of convolution kernels is 16, and the activation function is the ReLU function;

[0027] The pooling layer PL3-1 adopts the maximum pooling method, the pooling area width is 2×1, and the number of convolution kernels is 16;

[0028] The convolution kernel size of the convolution layer CL3-2 is set to 7×1, the stride is set to 1, the number of convolution kernels is 32, and the activation function is the ReLU function;

[0029] The pooling layer PL3-2 adopts the maximum pooling method, the pooling area width is 2×1, and the number of convolution kernels is 32;

[0030] The number of neurons in the fully connected layer is set to 512, and the activation function is the ReLU function;

[0031] The output layer adopts a Softmax classifier, and the output categories are 7, corresponding to seven operating conditions: normal, rotor inter-turn short circuit 3% and 6%, stator inter-turn short circuit 3% and 6%, and static eccentricity 3% and 6%.

[0032] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: the synchronous motor fault diagnosis method based on MSK-CNN and multi-source electromechanical information fusion provided by the present invention collects the phase voltage, rotor vibration, and stator vibration signals of the synchronous motor under normal and different fault states online, and inputs them into the MSK-CNN model respectively, obtains complementary and rich diagnostic information in parallel from multiple scales, and fuses the features extracted based on the three signals to achieve the diagnosis purpose; the MSK-CNN model established by the present invention integrates multi-scale learning, and learns effective fault features in parallel at different scales, thereby obtaining richer and complementary diagnostic information, improving the fault diagnosis performance, and having good robustness; different types of signals have different recognition degrees of synchronous motor faults. The present invention fuses the evidence theory of multi-source electromechanical information to avoid misjudgment, missed judgment, and other phenomena caused by single sensor failure, thereby further improving the fault diagnosis accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0034] Figure 1Schematic diagram of a flow chart of a synchronous motor fault diagnosis method based on MSK-CNN and multi-source electromechanical information fusion according to an embodiment of the present invention;

[0035] Figures 2a-2c These are respectively confusion matrix diagrams of fault diagnosis based on phase voltage, rotor vibration, and stator vibration according to an embodiment of the present invention;

[0036] Figure 3 This is a graph showing the fault diagnosis accuracy based on MSK-CNN and multi-source electromechanical information fusion according to an embodiment of the present invention;

[0037] Figure 4 This is a confusion matrix diagram of fault diagnosis based on multi-source electromechanical information fusion according to an embodiment of the present invention;

[0038] Figure 5a-5b These are t-SNE feature visualization diagrams of the original signal and the fully connected layer output signal in accordance with an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0040] The purpose of the present invention is to provide a synchronous motor fault diagnosis method based on MSK-CNN and multi-source electromechanical information fusion, which can capture the fault characteristics of the original signal from multiple scales, improve the feature learning ability, and avoid the occurrence of misjudgment, missed judgment and other phenomena caused by single sensor failure, thereby improving the fault diagnosis accuracy.

[0041] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] Synchronous motors have the characteristics of complex structure, high voltage level and heavy load. Affected by the operating environment, they are prone to various faults. Common electrical faults of synchronous motors include rotor winding inter-turn short circuit fault, eccentricity fault and stator inter-turn short circuit fault. Due to the complex coupling relationship between the various parts inside the motor, the cause of the fault and the phenomenon often show a one-to-many, many-to-one, and many-to-many mapping relationship. Common fault characteristic quantities include vibration, current, etc. However, the characteristics reflected by the fault are often not unique, and a single signal cannot fully reflect the fault state of the object under test and is easily interfered by environmental factors. The above problems can be solved by using information fusion technology. Electromechanical signals belong to heterogeneous information and have strong complementarity. Better diagnostic results can be obtained than similar information fusion. Therefore, the present invention selects rotor vibration, stator vibration and phase voltage as the original input signals to jointly diagnose synchronous motor faults.

[0043] like Figure 1 As shown, the synchronous motor fault diagnosis method based on MSK-CNN and multi-source electromechanical information fusion provided by the present invention includes the following steps:

[0044] S1, using a data acquisition device to collect phase voltage, rotor vibration, and stator vibration signals of the synchronous motor in real time under seven operating conditions: normal operation, rotor winding inter-turn short circuit 3% and 6%, stator inter-turn short circuit 3% and 6%, and static eccentricity 3% and 6%;

[0045] The data acquisition device includes a synchronous motor phase voltage signal detection device, a synchronous motor rotor vibration acceleration sensor, and a stator vibration acceleration sensor. Various sensors are installed on the synchronous motor and connected to the analysis device. The parameters of the synchronous motor are shown in Table 1.

[0046] S2, the collected phase voltage, rotor vibration and stator vibration signals are sliced ​​with a unit length of 5 cycles, and the training set and test set are divided into 8:2;

[0047] S3: Establish the MSK-CNN model, determine the network structure and convolution kernel size of the MSK-CNN model, use the training set for training, adjust the MSK-CNN network structure and hyperparameters according to the loss function and accuracy changes during the iteration process, train again, repeat the above process, determine the optimal MSK-CNN model, and use the test set to verify the effect of the optimal MSK-CNN model;

[0048] In step S4, the raw data of phase voltage, rotor vibration, and stator vibration are input into the optimal MSK-CNN model respectively. After passing through the multi-scale feature extraction layer, the fault features extracted by the multi-scale kernel are fused and used as the output of each MSK-CNN sub-model, facilitating the subsequent fusion of multi-scale fault features extracted based on different signals.

[0049] In S5, in the optimal MSK-CNN model, the fault features extracted from the phase voltage, rotor vibration, and stator vibration signals are aggregated and input into the same Flatten layer for flattening. They are then passed to the Softmax output layer through the fully connected layer to finally obtain the probability value of each operating condition. The operating condition corresponding to the maximum probability value is considered to be the final fault diagnosis result.

[0050] Table 1 Synchronous motor parameters

[0051]

[0052] The step S2 is a data preprocessing process:

[0053] One-dimensional convolutional neural networks can extract spatial correlation features from data. The raw rotor vibration, stator vibration, and phase voltage signals collected by various sensors are continuously changing data points with unclear spatial relationships. Therefore, they need to be converted into a data structure suitable for convolutional neural networks. Slicing is performed using a unit length of 5 cycles (500 sampling points). The raw data points for each operating state are divided into 3000 signal wave data points. 2400 samples are randomly selected as the training set, and 600 samples are used as the test set. Each operating state is labeled as shown in Table 2.

[0054] Table 2 Sample data labels

[0055]

[0056] In step S3, the MSK-CNN network structure corresponding to the optimal MSK-CNN model is shown in Table 3, specifically including: an input layer, a multi-scale kernel feature extraction layer, a fully connected layer, and an output layer; wherein the multi-scale kernel feature extraction layer is composed of multiple CNN sub-blocks with different convolution kernel sizes, and the CNN sub-block includes two alternating convolution layers CL, a normalized batch processing BN layer, an activation function ReLu layer, and a pooling layer PL. The convolution layer CL includes convolution layer CL1-1, convolution layer CL1-2, convolution layer CL2-1, convolution layer CL2-2, convolution layer CL3-1, and convolution layer CL3-2. The pooling layer PL includes pooling layer PL1-1, pooling layer PL1-2, pooling layer PL2-1, pooling layer PL2-2, pooling layer PL3-1, and pooling layer PL3-2.

[0057] Table 3 MSK-CNN network structure

[0058]

[0059]

[0060] The original signal size of the input layer is 500×1;

[0061] The convolution kernel size of the convolution layer CL1-1 is set to 3×1, the stride is set to 1, the number of convolution kernels is 16, and the activation function is selected as the ReLU function;

[0062] The pooling layer PL1-1 adopts the maximum pooling method, the pooling area width is 2×1, and the number of convolution kernels is 16;

[0063] The convolution kernel size of the convolution layer CL1-2 is set to 3×1, the stride is set to 1, the number of convolution kernels is 32, and the activation function is selected as the ReLU function;

[0064] The pooling layer PL1-2 adopts the maximum pooling method, the pooling area width is 2×1, and the number of convolution kernels is 32;

[0065] The convolution kernel size of the convolution layer CL2-1 is set to 5×1, the stride is set to 1, the number of convolution kernels is 16, and the activation function is the ReLU function;

[0066] The pooling layer PL2-1 adopts the maximum pooling method, the pooling area width is 2×1, and the number of convolution kernels is 16;

[0067] The convolution kernel size of the convolution layer CL2-2 is set to 5×1, the stride is set to 1, the number of convolution kernels is 32, and the activation function is the ReLU function;

[0068] The pooling layer PL2-2 adopts the maximum pooling method, the pooling area width is 2×1, and the number of convolution kernels is 32;

[0069] The convolution kernel size of the convolution layer CL3-1 is set to 7×1, the stride is set to 1, the number of convolution kernels is 16, and the activation function is the ReLU function;

[0070] The pooling layer PL3-1 adopts the maximum pooling method, the pooling area width is 2×1, and the number of convolution kernels is 16;

[0071] The convolution kernel size of the convolution layer CL3-2 is set to 7×1, the stride is set to 1, the number of convolution kernels is 32, and the activation function is the ReLU function;

[0072] The pooling layer PL3-2 adopts the maximum pooling method, the pooling area width is 2×1, and the number of convolution kernels is 32;

[0073] The number of neurons in the fully connected layer is set to 512, and the activation function is the ReLU function;

[0074] The output layer adopts a Softmax classifier, and the output categories are 7, corresponding to seven operating conditions: normal, rotor inter-turn short circuit 3% and 6%, stator inter-turn short circuit 3% and 6%, and static eccentricity 3% and 6%.

[0075] In step S4, the original data of phase voltage, rotor vibration, and stator vibration after data preprocessing are respectively input into the MSK-CNN model for training to obtain the specific classification of fault diagnosis based on the single signals of phase voltage, rotor vibration, and stator vibration, which is represented by a confusion matrix.

[0076] like Figures 2a-2c Figure 2 shows the confusion matrix for fault diagnosis based on phase voltage, rotor vibration, and stator vibration. It can be seen that a single signal is not very effective for synchronous motor fault diagnosis. For each operating condition of the synchronous motor, a large number of samples are misclassified.

[0077] In step S5, the fault features extracted from each signal are aggregated and input into the same Flatten layer for flattening, and then passed to the Softmax output layer through the fully connected layer to finally obtain the probability value of each operating state, among which the largest category is regarded as the final fault diagnosis result.

[0078] After the input signal is processed by the convolution layer and the pooling layer and extracted as high-level information features, a fully connected layer is set to expand the high-level feature vector into a one-dimensional vector; the output layer uses the Softmax classifier to solve the classification problem and obtain the final fault diagnosis result. Its expression is:

[0079]

[0080] Where θ is the parameter set of the training model; M is the total number of training set samples; (X m ,Y m ) is the training set sample of the model; D is the number of synchronous motor operating state categories; 1{Y m =d} is an indicator function. When the bracket value is true, the function value is 1, otherwise it is 0.

[0081] Figure 3 This is a graph showing the fault diagnosis accuracy based on MSK-CNN and multi-source electromechanical information fusion, as shown in Figure 3 During the model training process, the accuracy of both the training set and the test set increased rapidly, and converged to 100% and 99.64% respectively after 12 iterations. This shows that the proposed method has excellent anti-interference performance and is effective in feature mining and learning of synchronous motor fault data.

[0082] Figure 4The figure below shows a confusion matrix diagram for fault diagnosis based on multi-source electromechanical information fusion. The accuracy of fault diagnosis is significantly improved after multi-source information fusion, reaching almost 100% accuracy, except for individual sample classification errors. This demonstrates the strong complementarity of electromechanical signals and that multi-source electromechanical information fusion can achieve better diagnostic results.

[0083] Figure 5a-5b The figure shows a t-SNE feature visualization, where each sample is visualized as a point. Different colors represent different fault types, and samples belonging to the same class have the same color. It can be seen that the different health states of the original input signal overlap significantly, indicating that the feature information of the original signal is difficult to distinguish. However, the different fault features extracted by the output layer of the model established by this invention can be easily distinguished. This shows that the feature maps extracted by the proposed framework are better features for fault diagnosis, thus verifying the effectiveness of the present invention.

[0084] The present invention provides a synchronous motor fault diagnosis method based on MSK-CNN and multi-source electromechanical information fusion. First, the established MSK-CNN model integrates multi-scale learning and learns effective fault features in parallel at different scales, thereby obtaining richer and complementary diagnostic information, improving fault diagnosis performance, and having good robustness; second, different types of feature information have different recognition levels for synchronous motor faults. The present invention integrates the evidence theory of multi-source electromechanical information, avoiding misjudgment, missed judgment, and other phenomena caused by single sensor failure, further improving the fault diagnosis accuracy.

[0085] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A synchronous motor fault diagnosis method based on MSK-CNN and multi-source electromechanical information fusion, characterized in that: The following steps are involved: S1, using a data acquisition device to collect phase voltage, rotor vibration and stator vibration signals of the synchronous motor in real time under different operating conditions: normal operation, rotor winding inter-turn short circuit, stator inter-turn short circuit and static eccentricity; S2, slicing the collected phase voltage, rotor vibration, and stator vibration signals with a set unit length, and setting them as training sets and test sets according to the set ratio; S3, establish the MSK-CNN model, determine the network structure and convolution kernel size of the MSK-CNN model, use the training set for training, adjust the MSK-CNN network structure and hyperparameters according to the loss function and accuracy changes in the iterative process, train again, repeat the above process to determine the optimal MSK-CNN model, the optimal MSK-CNN model specifically includes: input layer, multi-scale kernel feature extraction layer, fully connected layer, output layer; wherein the multi-scale kernel feature extraction layer consists of multiple convolution kernels of different sizes The CNN sub-block includes two alternating convolutional layers CL, a normalized batch processing BN layer, an activation function ReLu layer, and a pooling layer PL. The convolutional layer CL includes convolutional layer CL1-1, convolutional layer CL1-2, convolutional layer CL2-1, convolutional layer CL2-2, convolutional layer CL3-1, and convolutional layer CL3-2. The pooling layer PL includes pooling layer PL1-1, pooling layer PL1-2, pooling layer PL2-1, pooling layer PL2-2, pooling layer PL3-1, and pooling layer PL3-2. In step S4, the raw data of phase voltage, rotor vibration, and stator vibration are input into the optimal MSK-CNN model respectively. After passing through the multi-scale feature extraction layer, the fault features extracted by the multi-scale kernel are fused and used as the output of each MSK-CNN sub-model. In S5, in the optimal MSK-CNN model, the fault features extracted from the phase voltage, rotor vibration, and stator vibration signals are aggregated and input into the same Flatten layer for flattening. They are then passed to the Softmax output layer through the fully connected layer to finally obtain the probability value of each operating condition. The operating condition corresponding to the maximum probability value is considered to be the final fault diagnosis result.

2. The synchronous motor fault diagnosis method based on MSK-CNN and multi-source electromechanical information fusion according to claim 1 is characterized in that: In step S1, the rotor winding inter-turn short-circuit operating conditions include two types of rotor winding inter-turn short-circuit: 3% and 6%, the stator inter-turn short-circuit operating conditions include two types of stator inter-turn short-circuit: 3% and 6%, and the static eccentricity operating conditions include two types of static eccentricity: 3% and 6%.

3. The synchronous motor fault diagnosis method based on MSK-CNN and multi-source electromechanical information fusion according to claim 1 is characterized in that: In step S1, the data acquisition device includes a synchronous motor phase voltage signal detection device, a synchronous motor rotor vibration acceleration sensor, and a stator vibration acceleration sensor.

4. The synchronous motor fault diagnosis method based on MSK-CNN and multi-source electromechanical information fusion according to claim 1 is characterized in that: In step S1 and step S2, the setting ratio is 8:

2.

5. The synchronous motor fault diagnosis method based on MSK-CNN and multi-source electromechanical information fusion according to claim 4 is characterized in that: The original signal size of the input layer is 500×1; The convolution kernel size of the convolution layer CL1-1 is set to 3×1, the stride is set to 1, the number of convolution kernels is 16, and the activation function is selected as the ReLU function; The pooling layer PL1-1 adopts the maximum pooling method, the pooling area width is 2×1, and the number of convolution kernels is 16; The convolution kernel size of the convolution layer CL1-2 is set to 3×1, the stride is set to 1, the number of convolution kernels is 32, and the activation function is selected as the ReLU function; The pooling layer PL1-2 adopts the maximum pooling method, the pooling area width is 2×1, and the number of convolution kernels is 32; The convolution kernel size of the convolution layer CL2-1 is set to 5×1, the stride is set to 1, the number of convolution kernels is 16, and the activation function is the ReLU function; The pooling layer PL2-1 adopts the maximum pooling method, the pooling area width is 2×1, and the number of convolution kernels is 16; The convolution kernel size of the convolution layer CL2-2 is set to 5×1, the stride is set to 1, the number of convolution kernels is 32, and the activation function is the ReLU function; The pooling layer PL2-2 adopts the maximum pooling method, the pooling area width is 2×1, and the number of convolution kernels is 32; The convolution kernel size of the convolution layer CL3-1 is set to 7×1, the stride is set to 1, the number of convolution kernels is 16, and the activation function is the ReLU function; The pooling layer PL3-1 adopts the maximum pooling method, the pooling area width is 2×1, and the number of convolution kernels is 16; The convolution kernel size of the convolution layer CL3-2 is set to 7×1, the stride is set to 1, the number of convolution kernels is 32, and the activation function is the ReLU function; The pooling layer PL3-2 adopts the maximum pooling method, the pooling area width is 2×1, and the number of convolution kernels is 32; The number of neurons in the fully connected layer is set to 512, and the activation function is the ReLU function; The output layer adopts a Softmax classifier, and the output categories are 7, corresponding to seven operating conditions: normal, rotor inter-turn short circuit 3% and 6%, stator inter-turn short circuit 3% and 6%, and static eccentricity 3% and 6%.

Images