Asynchronous motor fault diagnosis method based on electric signals and efficient mixed attention

Abstract

An asynchronous motor fault diagnosis method based on electric signals and high-efficiency mixed attention comprises the steps that a high-efficiency mixed attention mechanism is adopted, and a one-dimensional convolutional neural network and the high-efficiency mixed attention mechanism are combined to form a convolutional high-efficiency mixed attention module to extract spatial features of different view fields and capture long-range dependence of signal features; a time sequence efficient mixed attention module is proposed by combining a bidirectional long-short-term memory network and an efficient mixed attention mechanism, and fault recognition of a motor is realized through a classifier. The invention designs an asynchronous motor fault diagnosis method based on an electric signal and an efficient mixed attention mechanism, innovatively provides the efficient mixed attention mechanism, and combines the efficient mixed attention mechanism with a convolutional neural network and a bidirectional long-short term memory network. The comprehensive long-range dependent feature capture of the signals and the weight redistribution of the homologous signals under different channels are realized, so that the difference between the fault signal sensitivities of the asynchronous motor is reduced, and the fault diagnosis precision of the asynchronous motor is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of asynchronous motor fault diagnosis, and particularly relates to a method for asynchronous motor fault diagnosis based on electrical signals and efficient hybrid attention. Background Art

[0002] In modern industrial applications, asynchronous motors, as indispensable and crucial power equipment, are extremely widely and deeply applied in various complex mechanical systems and production processes. They are like the heart of the industrial field, providing continuous power support for various equipment. However, with the rapid development of production technology and the continuous improvement of process complexity, the industrial system has put forward more stringent and urgent requirements for the performance and reliability level of asynchronous motors. In order to ensure the continuous, stable, and efficient operation of these key devices, thus guaranteeing the smooth progress of the entire production line and further improving production efficiency, the fault detection and condition monitoring of asynchronous motors are particularly crucial and important. Through advanced detection techniques and means, we can timely discover potential fault hazards of the motor and prevent the adverse effects of sudden failures on production. At the same time, real-time monitoring of the motor's operating state can help us better understand its working conditions and provide strong data support for optimization adjustment and maintenance. Therefore, strengthening the fault detection and condition monitoring of asynchronous motors is not only a necessary means to improve equipment reliability and production efficiency but also an important measure to ensure industrial production safety and stability.

[0003] According to the Chinese patent with the patent number: CN118758607A, an intelligent detection method and device for asynchronous motor bearings are disclosed, which relate to the technical field of intelligent fault detection and include: receiving a control instruction of the asynchronous motor; controlling the asynchronous motor to obtain N bearing monitoring data streams; extracting the nth bearing monitoring data stream corresponding to the nth bearing; performing a health deviation analysis on the nth bearing to obtain the nth bearing health deviation matrix; performing a fault detection on the nth bearing to obtain the nth bearing fault detection result; obtaining an auxiliary component monitoring data set of the asynchronous motor, and performing a fault correlation compensation on the nth bearing fault detection result to generate the nth bearing fault detection report. Through this application, the technical problems of strong singularity and low accuracy in the bearing fault detection of existing asynchronous motors can be solved, the technical goal of multi-parameter comprehensive fault diagnosis can be achieved, and the technical effects of improving the accuracy of bearing fault detection of asynchronous motors and reducing the false alarm rate can be achieved.

[0004] However, when diagnosing faults in existing asynchronous motors, they rely too much on experts' prior knowledge to implement feature extraction processing, and cannot achieve end-to-end fault diagnosis, making it difficult to meet the fault diagnosis requirements of asynchronous motors in a big data environment. At the same time, due to the large differences in fault sensitivity and susceptibility to interference of asynchronous motor signals, existing deep learning models also cannot effectively extract the effective fault features of electrical signals. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the deficiencies that when diagnosing faults in existing asynchronous motors, they rely too much on experts' prior knowledge to implement feature extraction processing, and cannot achieve end-to-end fault diagnosis, making it difficult to meet the fault diagnosis requirements of asynchronous motors in a big data environment. At the same time, due to the large differences in fault sensitivity and susceptibility to interference of asynchronous motor signals, existing deep learning models also cannot effectively extract the effective fault features of electrical signals.

[0006] The technical solution adopted to solve the above technical problem is as follows:

[0007] A fault diagnosis method for asynchronous motors based on electrical signals and efficient hybrid attention, comprising:

[0008] S1: Under the working conditions where the input speed of the motor is set to 1750 r / min, the load current is adjusted to 0.7 A, and the system sampling frequency is 50 Hz, the current signals of asynchronous motors in six different health states are collected, namely healthy state, stator winding, bearing fault, rotor imbalance, rotor eccentricity, and rotor bar breakage;

[0009] S2: Adopt an efficient hybrid attention mechanism to reduce the differences in fault sensitivity of asynchronous motor electrical signals;

[0010] S3: Combine a one-dimensional convolutional neural network and an efficient hybrid attention mechanism to form a convolutional efficient hybrid attention module to extract spatial features in different visual fields and capture long-range dependencies of signal features, while suppressing the influence of irrelevant components within the same-source signals under different channels;

[0011] S4: Combine a bidirectional long short-term memory network and an efficient hybrid attention mechanism to propose a temporal efficient hybrid attention module to further realize the re-extraction of signal temporal features and the adaptive fusion of features in different channels;

[0012] S5: Realize the fault identification of the motor through a classifier.

[0013] Further, the specific content of S2 includes: The self-attention mechanism associates the features of different positions of a single sequence signal to calculate the feature representation of the same sequence. The calculation formula of the self-attention mechanism is as follows:

[0014]

[0015] Among them, Q refers to the self-prompt, that is, the feature vector generated by the subjective tendency. K refers to the non-self-prompt, which emphasizes the prominent features of the thing itself. V represents the features of the thing itself, and L represents the dimension of the key. The softmax function is used to normalize the attention weights so that their sum is 1.

[0016] Furthermore, the S2 specifically includes: The calculation process of the image recognition structure is as follows: The feature F undergoes global pooling F sq (·) for dimensionality reduction, and then passes through two fully connected layers F scale (·) to generate weights for each feature channel; the generated channel weights are multiplied by F channel by channel.

[0017] Through the above steps, the importance of each feature channel is automatically obtained and weighted, so as to enhance the useful features and suppress the features that are useless for the current task, realizing the adaptive adjustment of different channels and the improvement of network accuracy.

[0018] Furthermore, the S2 specifically includes: The efficient hybrid attention mechanism combines the advantages of the self-attention mechanism and the network for enhancing channel feature correlation in the image recognition structure, reduces the difference in fault sensitivity of the asynchronous motor electrical signal, and thus improves the feature extraction ability of the diagnostic model.

[0019] The calculation formula of the efficient hybrid attention mechanism is as follows:

[0020]

[0021] s = a(w2 · ReLU(w1 · z))

[0022] EHAM = s · V

[0023] In the formula, V is the input signal, SAM is the input signal with self-attention weights, h and w are the two dimensions of the signal respectively; z is the signal compressed by global pooling; w1 and w2 are two fully connected layers, which can restore the signal z compressed by global pooling to the original number of channels; s is the channel weight excited by the channel attention mechanism, and EHAM is the input signal with hybrid attention weights.

[0024] Furthermore, the S4 specifically includes: The convolutional neural network includes: convolutional layer, pooling layer, fully connected layer, Softmax layer, batch normalization and activation layer, and the output equation of this network is as follows:

[0025]

[0026] Among them, x t-w+1 is the input signal, k wis the network weight, b w is the network bias, and ReLU(·) is the activation function.

[0027] The pooling layer is mainly used to select features and reduce data dimensionality to reduce the training parameters of the model and improve the convergence speed. Batch normalization is used to improve the generalization performance of the model. The activation function is used to enhance the non-linear representation ability of the model and improve the computational efficiency of the model. The feature representation after the signal is processed by the feature selection of the convolutional layer and the pooling layer is mapped to the target sample label space by the fully connected layer, and finally the Softmax layer is used to calculate the corresponding label probability, thereby realizing fault diagnosis.

[0028] Further, the S5 specifically includes: The bidirectional long short-term memory network copes with the uncertainty problem of sensor signal acquisition caused by the lack of reference to future moment information by learning historical moment information and future moment information simultaneously. The expression of the bidirectional long short-term memory network is as follows:

[0029]

[0030] In the above formula, w t corresponds to the forward state weight of the bidirectional long short-term memory network, v t corresponds to the backward state weight, and b t is the bias.

[0031] The forward state of the bidirectional long short-term memory network and the backward state do not share parameters, and only act jointly on the result h t through different hidden layer connections to obtain richer signal features.

[0032] The beneficial effects of the present invention are as follows:

[0033] Since the present invention designs a fault diagnosis method for asynchronous motors based on electrical signals and an efficient hybrid attention mechanism, innovatively proposes an efficient hybrid attention mechanism, and combines it with a convolutional neural network and a bidirectional long short-term memory network to achieve comprehensive long-range dependent feature capture of signals and reallocation of the weights of homologous signals under different channels, thereby reducing the difference in the sensitivity of asynchronous motor fault signals and improving the accuracy of asynchronous motor fault diagnosis. Description of the Drawings

[0034] Figure 1 is the fault diagnosis framework for asynchronous motors;

[0035] Figure 2 is the schematic diagram of the efficient hybrid attention mechanism;

[0036] Figure 3 is the schematic diagram of the self-attention mechanism structure;

[0037] Figure 4 are the training and test accuracy curves;

[0038] Figure 5 are the visualizations of feature extraction for different modules. Specific implementation manners

[0039] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0040] It should be noted that the following detailed descriptions are all illustrative and are intended to provide further descriptions of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0041] It should be noted that the terms used herein are only for describing specific implementation manners and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless otherwise clearly specified in the context, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0042] Embodiment 1

[0043] As Figure 1 , this embodiment provides a method for diagnosing asynchronous motor faults based on electrical signals and an efficient hybrid attention mechanism, including:

[0044] S1: Under the working conditions where the input speed of the motor is set to 1750 r / min, the load current is adjusted to 0.7 A, and the system sampling frequency is 50 Hz, the current signals of six different health states of the asynchronous motor are collected, namely the healthy state, stator winding, bearing fault, rotor imbalance, rotor eccentricity, and rotor bar breakage;

[0045] S2: As Figure 2 , an efficient hybrid attention mechanism is adopted to reduce the fault sensitivity difference of asynchronous motor electrical signals;

[0046] S3: The one-dimensional convolutional neural network and the efficient hybrid attention mechanism are combined to form a convolutional efficient hybrid attention module to extract spatial features in different visual fields and capture long-range dependencies of signal features, while suppressing the influence of irrelevant components within the same-source signals in different channels;

[0047] S4: The bidirectional long short-term memory network and the efficient hybrid attention mechanism are combined to propose a temporal efficient hybrid attention module to further realize the re-extraction of signal temporal features and the adaptive fusion of different channel features;

[0048] S5: Implement motor fault identification through a classifier.

[0049] S2 specifically includes:

[0050] The self-attention mechanism associates the features of different positions of a single sequence signal to calculate the feature representation of the same sequence. The calculation formula of the self-attention mechanism is as follows:

[0051]

[0052] Among them, Q refers to the self-prompt, that is, the feature vector generated by the subjective tendency. K refers to the non-self-prompt, which emphasizes the prominent features of the thing itself. V represents the features of the thing itself, and L represents the dimension of the key. The softmax function is used to normalize the attention weights so that their sum is 1. The structure of the self-attention mechanism is as Figure 3 shown.

[0053] S2 specifically includes:

[0054] The calculation process of the image recognition structure is as follows: The feature F is reduced in dimension through global pooling F sq (·), and then passes through two fully connected layers F scale (·) to generate weights for each feature channel; the generated channel weights are multiplied by F channel by channel.

[0055] Through the above steps, the importance of each feature channel is automatically obtained and weighted, so as to enhance the useful features and suppress the features that are useless for the current task, realizing the adaptive adjustment of different channels and the improvement of network accuracy.

[0056] S2 specifically includes:

[0057] The efficient hybrid attention mechanism combines the advantages of the self-attention mechanism and the image recognition structure network to enhance channel feature correlation, reduces the fault sensitivity difference of the asynchronous motor electrical signal, and thus improves the feature extraction ability of the diagnostic model.

[0058] The calculation formula of the efficient hybrid attention mechanism is as follows:

[0059]

[0060] s = σ(w2·ReLU(w1·z))

[0061] EHAM = w = s·V

[0062] Wherein, V is the input signal, SAM is the input signal with self-attention weights, h and w are the two dimensions of the signal respectively; z is the signal compressed by global pooling; w1 and w2 are two fully connected layers, which can restore the signal z compressed by global pooling to the original number of channels; s is the channel weight excited by the channel attention mechanism, and EHAM is the input signal with hybrid attention weights.

[0063] S4 specifically includes:

[0064] The convolutional neural network includes: a convolutional layer, a pooling layer, a fully connected layer, a Softmax layer, batch normalization, and an activation layer. The output equation of this network is as follows:

[0065]

[0066] Among them, x t-w+1 is the input signal, k w is the network weight, b w is the network bias, and ReLU(·) is the activation function.

[0067] The pooling layer is mainly used to select features and reduce data dimensionality to reduce the training parameters of the model and improve the convergence speed. Batch normalization is used to improve the generalization performance of the model. The activation function is used to enhance the non-linear representation ability of the model and improve the computational efficiency of the model. The feature representation of the signal after feature selection processing by the convolutional layer and the pooling layer is mapped to the target sample label space using the fully connected layer, and finally the corresponding label probability is calculated using the Softmax layer, thereby realizing fault diagnosis.

[0068] S5 specifically includes:

[0069] The bidirectional long short-term memory network copes with the uncertainty problem of sensor signal acquisition caused by the lack of reference to future moment information by learning historical moment information and future moment information simultaneously. The expression of the bidirectional long short-term memory network is as follows:

[0070]

[0071] In the above formula, w t corresponds to the forward state weight of the bidirectional long short-term memory network, v t corresponds to the backward state weight, and b t is the bias.

[0072] The forward state of the bidirectional long short-term memory network and the backward state do not share parameters, and only act jointly on the result h t through different hidden layer connections to obtain richer signal features.

[0073] In a specific embodiment: The hardware platform is an Intel(R) Core(TM) i9-9900K CPU @ 3.60GHz processor, and the software platform is the TensorFlow framework combined with the deep learning Keras database. All experiments were conducted under the same random seed, using Adam as the optimizer for training, with the learning rate set to 0.001 and the batch size to 128.

[0074] To avoid interference from random errors, the proposed model was run 10 times on the above-mentioned electrical signal test dataset, and the average test accuracy was taken as the final diagnostic accuracy of the model

[17] . After 10 runs of the model, the final diagnostic accuracy was 98.17%, and its iteration curve is as Figure 4 shown.

[0075] As Figure 4 shown, the proposed model has good convergence effect and fast convergence speed, and the fault recognition accuracy of the training samples and test samples of the electrical signal dataset both reach over 98%. This experimental result proves that the fault diagnosis model proposed in this paper can effectively use electrical signals to identify and classify motor faults. To further reveal the feature mining performance of different modules of the above model, the distribution neighborhood embedding algorithm technology was used to visually display the feature mining results of the proposed model, with different colors representing different fault types. The visualization results of feature extraction of different modules are as Figure 5 shown.

[0076] It can be seen from Figure (a) that the spatial features of different visual fields of the motor electrical signals extracted by the convolutional network module gather together and cannot be effectively separated. It can be seen from Figure (b) that after passing the filtered spatial features through the efficient hybrid attention module, the long-range dependence of the signal features is captured, and the influence of irrelevant components within the homologous signals under different channels is suppressed, realizing the efficient separation of features. The feature extraction results of Figure (a) and (b) prove that the efficient hybrid attention mechanism proposed in this paper can effectively reduce the fault sensitivity difference of asynchronous motor electrical signals, thereby improving the feature re-separation ability of the network. However, the separation ability of single spatial feature extraction for the mechanical fault signal features in the motor electrical signals is limited, and there is still a phenomenon of aggregation of a small amount of mechanical fault features (such as bearing faults and rotor imbalance), which affects the diagnostic accuracy of the model. Therefore, it is necessary to introduce a bidirectional long short-term memory network to further extract the temporal features of the signals. It is found from Figure (c) and Figure (d) that combining the bidirectional long short-term memory network and the efficient hybrid attention mechanism further realizes the re-separation of mechanical fault signal features and effectively improves the feature extraction ability of the model.

[0077] From Figure 4 and Figure 5The test results can prove that the proposed method can effectively reduce the sensitivity to electrical signal faults, further extract the fault characteristics of different categories, and effectively classify the fault electrical signals of the motor.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the purpose and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. An asynchronous motor fault diagnosis method based on electrical signals and efficient hybrid attention, comprising: S1: Under the condition that the input speed of the motor is set to 1750 r / min, the load current is adjusted to 0.7 A, and the system sampling frequency is 50 Hz, the current signals of asynchronous motors in six different health states are collected, namely the healthy state, stator winding, bearing fault, rotor imbalance, rotor eccentricity, and rotor bar breakage; S2: Adopt an efficient hybrid attention mechanism to reduce the difference in fault sensitivity of asynchronous motor electrical signals; S3: Combine a one-dimensional convolutional neural network and an efficient hybrid attention mechanism to form a convolutional efficient hybrid attention module to extract spatial features in different visual fields and capture long-range dependencies of signal features, while suppressing the influence of irrelevant components within homologous signals under different channels; S4: Combine a bidirectional long short-term memory network and an efficient hybrid attention mechanism to propose a temporal efficient hybrid attention module to further realize the re-extraction of signal temporal features and the adaptive fusion of features in different channels; S5: Realize the fault identification of the motor through a classifier.

2. The asynchronous motor fault diagnosis method based on electrical signals and efficient hybrid attention according to claim 1, characterized in that The specific content of S2 includes: The self-attention mechanism correlates the features of different positions of a single sequence signal to calculate the feature representation of the same sequence. The calculation formula of the self-attention mechanism is as follows: Among them, Q refers to self-prompting, that is, the feature vector generated by subjective tendency. K refers to non-self-prompting, emphasizing the prominent features of things themselves. V represents the features of things themselves, and L represents the dimension of the key. The softmax function is used to normalize the attention weights so that their sum is 1.

3. The asynchronous motor fault diagnosis method based on electrical signals and efficient hybrid attention according to claim 2, characterized in that The specific content of S2 includes: The calculation process of the image recognition structure is as follows: The feature F undergoes global pooling F sq (·) for dimensionality reduction, and then passes through two fully connected layers F scale (·) to generate weights for each feature channel; multiply the generated channel weights channel by channel onto F. Through the above steps, automatically obtain the importance degree of each feature channel and assign weights, so as to enhance useful features and suppress features useless for the current task, and realize the adaptive adjustment of different channels and the improvement of network accuracy.

4. The asynchronous motor fault diagnosis method based on electrical signals and efficient hybrid attention according to claim 3, wherein The specific content of S2 includes: The efficient hybrid attention mechanism combines the advantages of the self-attention mechanism and the image recognition structure network in enhancing channel feature correlation, reduces the difference in fault sensitivity of asynchronous motor electrical signals, and thus improves the feature extraction ability of the diagnostic model. The calculation formula of the efficient hybrid attention mechanism is as follows: s = σ(w2·ReLU(w1·z)) EHAM = s·V In the formula, V is the input signal, SAM is the input signal with self-attention weights, h and w are the two dimensions of the signal respectively; z is the signal compressed by global pooling; w1 and w2 are two fully connected layers, which can restore the signal z compressed by global pooling to the original number of channels; s is the channel weight excited by the channel attention mechanism, and EHAM is the input signal with hybrid attention weights.

5. The asynchronous motor fault diagnosis method based on electrical signals and efficient hybrid attention according to claim 1, wherein The specific content of S4 includes: The convolutional neural network includes: a convolutional layer, a pooling layer, a fully connected layer, a Softmax layer, batch normalization, and an activation layer. The output equation of this network is as follows: Among them, x t-w+1 is the input signal, k w is the network weight, b w is the network bias, and ReLU(·) is the activation function. The pooling layer is mainly used for feature selection and data dimensionality reduction to reduce the training parameters of the model and improve the convergence speed. Batch normalization is used to improve the generalization performance of the model. The activation function is used to enhance the non-linear representation ability of the model and improve the computational efficiency of the model. The feature representation after the signal is processed by feature selection in the convolutional layer and the pooling layer is mapped to the target sample label space using the fully connected layer, and finally the corresponding label probability is calculated using the Softmax layer, thereby realizing fault diagnosis.

6. The asynchronous motor fault diagnosis method based on electrical signals and efficient hybrid attention according to claim 1, wherein The specific content of S5 includes: The bidirectional long short-term memory network copes with the uncertainty problem of sensor signal acquisition caused by the lack of reference to future moment information by learning historical moment information and future moment information simultaneously. The expression of the bidirectional long short-term memory network is as follows: In the above formula, w t corresponds to the forward state weight of the bidirectional long short-term memory network, and v t corresponds to the backward state weight, and b t is the bias. Forward state of bidirectional long short-term memory network and backward state do not share parameters. Only through different hidden layer connections, the results h t act together to obtain richer signal features.

Images