Permanent magnet synchronous motor winding short circuit identification method

Abstract

The invention belongs to the field of motor fault recognition, and particularly discloses a permanent magnet synchronous motor winding short circuit recognition method, which comprises the following steps: S1, constructing a convolutional neural network model, and inputting three-phase current signals and vibration signals as input signals; s2, establishing at least two input channel groups, introducing a space attention module, performing feature extraction on the three-phase current signal and the vibration signal, and splicing the extracted features; and S3, training the convolutional neural network model by adopting the public data set. According to the invention, through a multi-input mechanism, the current and vibration signals of the PMSM are effectively separated and processed, so that the characteristics of the sample data are accurately extracted. On the basis of current and vibration signals, an attention mechanism is introduced into a convolutional neural network, and a multi-input strategy is adopted, so that the signal feature extraction and selection capability of the network is enhanced.

Description

Technical Field

[0001] The invention belongs to the technical field of motor fault identification, and in particular relates to a method for identifying a short circuit in a permanent magnet synchronous motor winding. Background Art

[0002] Permanent Magnet Synchronous Motor (PMSM) is widely used in precision control fields such as industrial control and aerospace due to its advantages such as large moment of inertia, high power density and strong robustness. PMSM performs well under normal operating conditions, but various types of damage may still occur during long-term use. The most common faults include winding short circuit, permanent magnet demagnetization and bearing damage. Winding turn-to-turn short circuit faults account for about 30% of motor faults. This type of fault is mainly caused by the deterioration of winding insulation. Once a turn-to-turn short circuit occurs, the fault current may reach twice the normal current, resulting in severe local overheating. In addition, the current becomes larger, causing the magnetic field to be amplified, and its instantaneous value is greater than the coercive force of the permanent magnet, which may cause irreversible local demagnetization faults of the permanent magnets on the rotor. The motor can still operate normally at the beginning of the fault, but if the fault is not detected and repaired in time, it will quickly spread to a larger part, causing excessive performance degradation or even downtime of the PMSM. Therefore, if these problems are not diagnosed and solved in time, the deterioration of the fault will lead to further damage to the insulation, aggravate short circuit faults and vibration, and seriously affect the performance and life of the motor.

[0003] The assessment of the health status of permanent magnet synchronous motors can generally be classified into two methods: signal analysis-based and artificial intelligence-based identification. Both strategies rely on signal processing to achieve accurate fault diagnosis. Traditional signal analysis fault diagnosis methods rely on manual feature selection, which requires high human participation and relies on expert knowledge in the field. The use of artificial intelligence technology to diagnose the state of permanent magnet synchronous motor windings requires feature extraction and algorithm selection for different application environments, which is subjective and prone to omission and loss of fault information, resulting in poor overall generalization of the algorithm. In addition, the time sequence of current and vibration information, and the convolutional neural network (CNN) are easily affected by the selection of hyperparameters, resulting in insufficient recognition accuracy. Summary of the invention

[0004] The present invention provides a method for identifying a permanent magnet synchronous motor winding short circuit, which is used to solve the problem of low motor short circuit fault identification accuracy.

[0005] In order to solve the above technical problems, the technical solution of the present invention is: the method for identifying a short circuit in a permanent magnet synchronous motor winding comprises the following steps:

[0006] S1: Construct a convolutional neural network model, with the input signals being three-phase current signals and vibration signals;

[0007] S2: Establish an input channel group and introduce a spatial attention module to extract features from three-phase current signals and vibration signals, and concatenate the extracted features;

[0008] S3: Use a public data set to train the convolutional neural network model to determine the short circuit condition of the permanent magnet synchronous motor.

[0009] In a preferred embodiment of the present invention, in step S2, the input channel group includes a plurality of independent input channels.

[0010] In a preferred embodiment of the present invention, the three-phase current signals are input in parallel, and the three-phase currents are processed through the same input channel.

[0011] In a preferred embodiment of the present invention, in step S2, different convolutional layers are arranged in different input channels.

[0012] In a preferred embodiment of the present invention, the input channel sequentially includes a first convolutional layer, a first linear rectification function layer, a first maximum pooling layer, a second convolutional layer, a second linear rectification function layer, a second maximum pooling layer and a spatial attention module.

[0013] In a preferred embodiment of the present invention, the convolutional neural network model also includes a third convolutional layer, a third linear rectification function layer, a third maximum pooling layer, a Flatten layer, a fourth convolutional layer, a fourth linear rectification function layer, a fifth convolutional layer, a fifth linear rectification function layer and an output layer, which are sequentially arranged at the output end of the input channel group.

[0014] In a preferred embodiment of the present invention, in step S2, the spatial attention module is obtained by formula 1:

[0015]

[0016] Among them, F is the input feature map, AvgPool(F) is the average pooling, MaxPool(F) is the maximum pooling, f is the applied convolution, and σ is the activation function.

[0017] In a preferred embodiment of the present invention, the input channel group includes four input channels.

[0018] In a preferred embodiment of the present invention, in step S2, the spatial attention module includes a maximum pooling and weighted layer, an average pooling and weighted layer, a convolutional layer, and a logistic function activation layer, and the three-phase current signal or vibration signal is processed by the spatial attention module to generate a spatial attention map.

[0019] In a preferred embodiment of the present invention, the public data set in step S3 is obtained through an acquisition system, and the acquisition system includes a load controller, a PMSM, a current sensor, and a vibration sensor.

[0020] The technical solution provided by the present invention has the following advantages compared with the prior art:

[0021] The present invention effectively separates and processes the current and vibration signals of the PMSM through a multi-input mechanism, thereby accurately extracting the features of the sample data and realizing accurate diagnosis of the PMSM winding state. Based on the current and vibration signals, an attention mechanism is introduced into the convolutional neural network, and a multi-input strategy is adopted to enhance the network's ability to extract and select signal features. This method achieves high-precision identification of PMSM short-circuit faults, demonstrating its effectiveness and application potential in complex fault diagnosis scenarios. This method has obvious advantages over other artificial intelligence fault diagnosis technologies:

[0022] 1) By setting up a dedicated processing channel for each time series signal, the unique information in each signal is fully mined, which enhances the flexibility, learning efficiency and interpretability of the model, making it particularly suitable for processing tasks involving complex multiple signals.

[0023] 2) Designing independent convolutional layers for different signals can more accurately extract their respective features; at the same time, by allowing signals of the same type to share convolutional layers, the complexity of the model and training time are reduced.

[0024] 3) The spatial attention mechanism is introduced into the convolutional neural network to enhance the model's global extraction capability of time series signal features and improve the recognition accuracy of the model by effectively utilizing the order information of the sequence data. In addition, the spatial attention mechanism also improves the generalization ability of the model, reduces the dependence on the operator's skills, and enables the model to be quickly and efficiently applied to different models of motors. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, drawings of other embodiments can be obtained based on these drawings without paying creative work.

[0026] Figure 1 It is a schematic diagram of a convolutional neural network model in a method for identifying a short circuit in a permanent magnet synchronous motor winding according to an embodiment of the present invention;

[0027] Figure 2It is a simulation diagram of the loss of Mobi model training in a method for identifying a short circuit in a permanent magnet synchronous motor winding described in one embodiment of the present invention;

[0028] Figure 3 It is a schematic diagram of the accuracy of identification model training in a method for identifying a short circuit in a permanent magnet synchronous motor winding described in one embodiment of the present invention;

[0029] Figure 4 It is a schematic diagram of a CNN confusion matrix in a method for identifying a short circuit in a permanent magnet synchronous motor winding described in one embodiment of the present invention;

[0030] Figure 5 It is a schematic diagram of the attention convolutional neural network confusion matrix in a permanent magnet synchronous motor winding short circuit identification method described in one embodiment of the present invention. DETAILED DESCRIPTION

[0031] For ease of understanding, a method for identifying a short circuit in a permanent magnet synchronous motor winding is described below in conjunction with embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0032] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation and positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0033] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected" and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0034] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. The preferred embodiments of the present invention are given in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thoroughly understood.

[0035] The present invention discloses a method for identifying a short circuit in a permanent magnet synchronous motor winding, which comprises the following steps:

[0036] S1: Construct a convolutional neural network model with the input signals being three-phase current signals and vibration signals.

[0037] The one-dimensional convolutional neural network model greatly improves the efficiency and effectiveness of sequence data processing through its unique one-dimensional convolution structure and highly flexible network design, while significantly reducing the complexity of traditional feature extraction. By deeply learning the complex patterns in the data, the one-dimensional convolutional neural network model can provide accurate analysis results in a variety of tasks such as prediction, classification, and sequence modeling.

[0038] In order to more accurately capture the unique information features contained in each signal, a special convolution layer is designed for current and vibration signals in the present invention.

[0039] S2: Establish an input channel group and introduce the spatial attention module (SAM) to extract features of the three-phase current signal and vibration signal, and concatenate the extracted features.

[0040] In step S2, the input channel group includes a plurality of independent input channels. Figure 1 As shown, the input channel group includes four input channels. Different convolutional layers are arranged in different input channels.

[0041] The three-phase current signals are input in parallel, that is, the three-phase currents are processed through the same input channel to simplify the model structure and shorten the training time.

[0042] Although the three-phase current signals are input in parallel, they share the same set of convolutional layers. This enables the model to analyze the characteristics of each input signal more deeply, thereby significantly improving the accuracy and efficiency of feature extraction, while also ensuring the complexity of the model to a certain extent. When processing different types of signals, considering the differences in their respective scales, frequencies, and noise levels, processing them separately can ensure the effective extraction of the key features of each signal and avoid being interfered by the noise of other signals.

[0043] Specific reference Figure 1 , the input channel includes the first convolution layer, the first linear rectification function layer, the first maximum pooling layer, the second convolution layer, the second linear rectification function layer, the second maximum pooling layer and the spatial attention module in sequence.

[0044] The convolutional neural network model also includes a third convolution layer, a third linear rectification function layer, a third maximum pooling layer, a Flatten layer, a fourth convolution layer, a fourth linear rectification function layer, a fifth convolution layer, a fifth linear rectification function layer and an output layer, which are sequentially arranged at the output end of the input channel group.

[0045] In the construction of deep learning models, independent processing of different time series signals can significantly reduce mutual interference between signals and improve the robustness of the model. By customizing exclusive convolution processing processes for various types of signals, the model can more deeply grasp the inherent complexity and structural characteristics of each signal, thereby enhancing the ability to analyze signals and improving overall performance. Compared to mixing all signals and inputting them into the same convolutional network, processing the signals separately makes the use of network parameters more efficient and targeted. The processing path of each signal can be optimized for its specific characteristics, ensuring the effective use of network parameters during feature extraction and further optimizing model performance.

[0046] In step S2, the spatial attention module is obtained by formula 1:

[0047]

[0048] Among them, F is the input feature map, AvgPool(F) is the average pooling, MaxPool(F) is the maximum pooling, f is the applied convolution, and σ is the activation function.

[0049] The spatial attention module includes a maximum pooling and weighting layer, an average pooling and weighting layer, a convolutional layer, and a logistic function activation layer. The three-phase current signal or vibration signal is processed by the spatial attention module to generate a spatial attention map in order to highlight the key areas in the features.

[0050] The introduction of the spatial attention module can focus on those feature combinations with decisive expressive significance. By using weight vectors, higher weights are assigned to those state features with significant classification capabilities, thereby highlighting the relative importance of these state features in the entire time series signal in the model.

[0051] In this improved framework, the SAM module is placed before the convolutional layer, aiming to embed a new mechanism in the traditional feature extraction process to give the network the ability to identify the interdependencies between features. In this way, the model can not only evaluate the importance of a single feature, but also gain insight into the relative position and importance of each feature in the entire sequence. The essence of this method is to give higher weights to key decision-making features while reducing attention to information that has less impact on the task, thereby improving the classification efficiency and accuracy of the model.

[0052] Specifically, assuming that the original time series signal obtained through the input layer is represented as X, and the feature autocorrelation weight matrix calculated by the SAM module is W, then the role of the SAM module can be described as applying different weights to each feature to highlight those features that are more important to the task objective. The advantage of introducing the SAM module is that it does not introduce additional training parameters, and by dynamically adjusting the weight distribution of the input features, it effectively enhances the model's ability to capture key information, thereby improving the learning efficiency and overall performance of the entire network.

[0053] In summary, by incorporating the SAM module into the convolutional neural network, the model can significantly enhance its recognition accuracy for specific fault types such as winding short circuit when processing current and vibration signals. This comprehensive network design aims to effectively capture and utilize the information in the current and vibration signals to achieve robust and accurate recognition of PMSM short circuit faults.

[0054] S3: Use public datasets to train the convolutional neural network model to determine the short-circuit condition of the permanent magnet synchronous motor.

[0055] In one embodiment of the present invention, a public data set released by the Korea Advanced Institute of Science and Technology is used. Combined with Table 1, the data set includes vibration and current data of a three-phase permanent magnet synchronous motor with a stator fault. The data collection system is composed of a load controller, a PMSM, a current sensor, and a vibration sensor. The data set is constructed by setting artificial short circuits on the coil and between the coil turns.

[0056] Table 1 PMSM parameters

[0057]

[0058] In the present invention, a deep learning model of a multi-input attention convolutional neural network is used to identify various short-circuit faults occurring in the windings of PMSM. In order to optimize the classification effect, the output layer uses one-hot encoding to represent the prediction results, thereby improving the performance of the model in multi-class fault identification tasks.

[0059] During the training process, in order to improve the optimization efficiency of the neural network and reduce the impact of data distribution on the training effect, a batch normalization layer was introduced between the convolution layer and the pooling layer. This step aims to increase the sensitivity of the activation function to the input features and avoid overfitting the model on a specific data set, thereby improving the generalization performance of the model.

[0060] At the same time, the present invention also adopts a random inactivation mechanism, which reduces the model's excessive dependence on individual features and further prevents the occurrence of overfitting by randomly inactivating the activation values ​​of some neurons with a certain probability during the forward propagation process.

[0061] In this experiment, the data set was divided into training set and test set in a ratio of 7:3, and the status of PMSM was defined as three categories: healthy, turn-to-turn short circuit, and coil short circuit. The probability of random deactivation was set to 0.2, and the classification accuracy was selected as the evaluation index. The cross entropy was used as the loss function for 1000 rounds of training. The training process loss is as follows Figure 2 As shown, the success rate is Figure 3 As shown in the figure, it can be seen that both the traditional CNN and the attention convolutional neural network can achieve stable convergence.

[0062] Compared with traditional CNN, although the attention convolutional neural network has a larger loss fluctuation in the early stage of training, it shows a faster convergence speed. In addition, the accuracy of the attention convolutional neural network in identifying PMSM short-circuit faults is as high as 98.28%, exceeding the 97.78% of CNN. Although the optimal accuracy has not yet been achieved, there is still room for improvement in recognition accuracy by further extending the training time. By comparing the network training process before and after the introduction of the spatial attention module, it can be seen that the spatial attention module can reduce the number of training rounds required for the network to reach the optimal state without adding additional training parameters, thereby enhancing the generalization ability of the network.

[0063] When the two models obtained through training were evaluated using the test set, the accuracy reached 80.29% and 98.54% respectively. This result further confirms that the introduction of the spatial attention module can significantly improve the accuracy and generalization performance of the network, thereby enhancing the model's recognition effect on the PMSM winding state.

[0064] The evaluation accuracy reflects the overall performance of the discriminant network on the current and vibration signal data sets. In order to avoid the problem that the skewed data generated during the random partitioning of the data set obscures the actual recognition, the present invention uses a confusion matrix for a more detailed analysis.

[0065] like Figure 4 and Figure 5 The two confusion matrices shown show that the recall rate of the CNN network in identifying PMSM short-circuit faults is generally lower than that of the attention convolutional neural network network. This performance weakness hidden in the overall data may cause a greater risk of misdiagnosis in fault diagnosis. In contrast, the attention convolutional neural network network generally has an accuracy rate of more than 98% in identifying various types of short-circuit faults, which is consistent with the overall performance of the network obtained by evaluating the accuracy.

[0066] In response to the need for motor winding state diagnosis, the present invention proposes a method for identifying short-circuit windings of permanent magnet synchronous motors. The model effectively separates and processes the current and vibration signals of the PMSM through a multi-input mechanism, thereby accurately extracting the features of the sample data and realizing accurate diagnosis of the PMSM winding state. Based on the current and vibration signals, an attention mechanism is introduced into the convolutional neural network, and a multi-input strategy is adopted to enhance the network's ability to extract and select signal features. This method achieves high-precision identification of PMSM short-circuit faults, demonstrating its effectiveness and application potential in complex fault diagnosis scenarios. This method has obvious advantages over other artificial intelligence fault diagnosis technologies:

[0067] 1) By setting up a dedicated processing channel for each time series signal, the unique information in each signal is fully mined, which enhances the flexibility, learning efficiency and interpretability of the model, making it particularly suitable for processing tasks involving complex multiple signals.

[0068] 2) Designing independent convolutional layers for different signals can more accurately extract their respective features; at the same time, by allowing signals of the same type to share convolutional layers, the complexity of the model and training time are reduced.

[0069] 3) The spatial attention module is introduced into CNN to enhance the model's global extraction capability of time series signal features and improve the recognition accuracy of the model by effectively utilizing the order information of sequence data. In addition, the spatial attention module also improves the generalization ability of the model, reduces the dependence on operator skills, and enables the model to be quickly and efficiently applied to different models of motors.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. Although the present invention has been described in detail with reference to the above embodiments, a person skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some or all of the technical features can be replaced by equivalents, and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying a short circuit in a permanent magnet synchronous motor winding, characterized in that: The following steps are involved: S1: Construct a convolutional neural network model, with the input signals being three-phase current signals and vibration signals; S2: Establish an input channel group and introduce a spatial attention module to extract features from three-phase current signals and vibration signals, and concatenate the extracted features; S3: Use a public data set to train the convolutional neural network model to determine the short circuit condition of the permanent magnet synchronous motor.

2. A method for identifying a short circuit in a permanent magnet synchronous motor winding according to claim 1, characterized in that: In the step S2, the input channel group includes a plurality of independent input channels.

3. A method for identifying a short circuit in a permanent magnet synchronous motor winding according to claim 2, characterized in that: The three-phase current signals are input in parallel, and the three-phase currents are processed through the same input channel.

4. A method for identifying a short circuit in a permanent magnet synchronous motor winding according to claim 2, characterized in that: In the step S2, the convolutional layers in different input channels.

5. A method for identifying a short circuit in a permanent magnet synchronous motor winding according to claim 2, characterized in that: The input channel includes a first convolution layer, a first linear rectification function layer, a first maximum pooling layer, a second convolution layer, a second linear rectification function layer, a second maximum pooling layer and a spatial attention module in sequence.

6. A method for identifying a short circuit in a permanent magnet synchronous motor winding according to claim 4, characterized in that: The convolutional neural network model also includes a third convolution layer, a third linear rectification function layer, a third maximum pooling layer, a Flatten layer, a fourth convolution layer, a fourth linear rectification function layer, a fifth convolution layer, a fifth linear rectification function layer and an output layer, which are sequentially arranged at the output end of the input channel group.

7. A method for identifying a short circuit in a permanent magnet synchronous motor winding according to claim 1, characterized in that: In step S2, the spatial attention module is obtained by formula 1: Among them, F is the input feature map, AvgPool(F) is the average pooling, MaxPool(F) is the maximum pooling, f is the applied convolution, and σ is the activation function.

8. A method for identifying a short circuit in a permanent magnet synchronous motor winding according to claim 1, characterized in that: The input channel group includes four input channels.

9. A method for identifying a short circuit in a permanent magnet synchronous motor winding according to claim 1, characterized in that: In step S2, the spatial attention module includes a maximum pooling and weighting layer, an average pooling and weighting layer, a convolution layer, and a logistic function activation layer. The three-phase current signal or vibration signal is processed by the spatial attention module to generate a spatial attention map.

10. A method for identifying a short circuit in a permanent magnet synchronous motor winding according to claim 1, characterized in that: The public data set in step S3 is obtained through an acquisition system, which includes a load controller, a PMSM, a current sensor, and a vibration sensor.

Images