Motor air gap unevenness diagnosis method based on dynamic focusing channel attention

Abstract

The invention discloses a motor air gap unevenness diagnosis method based on dynamic focusing channel attention. The method comprises the following steps: converting a high-frequency current signal into a two-dimensional current image by adopting a Grubrum angle field; a collaborative comparison reconstruction constraint is introduced between the generative network and the parameter estimation network, an unsupervised collaborative comparison reconstruction module is constructed, and the definition constraint of the two-dimensional current image is enhanced; a dynamic focusing channel attention module is adopted to dynamically fuse global and local feature information, a dependency relationship between features is captured through a correlation matrix, and feature channel weights are distributed in a self-adaptive mode. According to the method, the problem of insufficient current image constraint is solved by constructing the collaborative comparison reconstruction module and the dynamic focusing channel attention module, more accurate feature weight distribution is realized, and the accuracy of motor air gap unevenness diagnosis is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence deep learning fault diagnosis, and specifically relates to a motor air gap unevenness diagnosis method based on dynamic focusing channel attention. Background Art

[0002] With the continuous improvement of industrial automation and the increasingly stringent requirements for energy efficiency, motors are indispensable equipment in modern industry and daily life. The performance and reliability of motors have received more attention. The uneven air gap fault of the motor refers to the uneven air gap between the stator and the rotor of the motor at different positions. During the long-term operation of the motor, the motor is affected by mechanical vibration, temperature changes, bearing wear and other factors, which will cause the position of the rotor to shift, thereby causing uneven air gap. Uneven air gap is one of the common faults of the motor, which will cause the motor to have problems such as increased vibration and increased noise. Severe uneven air gap will also cause the motor winding to heat up and even burn the motor. Accurate diagnosis of uneven air gap faults of the motor is of great significance to ensure the normal operation of the motor and extend its service life. Summary of the invention

[0003] The purpose of the present invention is to provide a method for diagnosing uneven air gap of a motor based on dynamic focusing channel attention, by diagnosing the fault of uneven air gap of the motor and taking timely measures to repair it, thereby extending the service life of the motor.

[0004] In order to solve the above technical problems, the present invention provides a motor air gap unevenness diagnosis method based on dynamic focusing channel attention, comprising: The Gram angle field is used to convert high-frequency current signals into two-dimensional current images; A collaborative contrast reconstruction module is used to enhance the clarity constraints of two-dimensional current images; The dynamic focus channel attention module is used to dynamically fuse global and local feature information and adaptively assign feature channel weights; AdaBoost classifier is used to diagnose the uneven air gap fault of the motor.

[0005] Optionally, the converting the high-frequency current signal into a two-dimensional current image by using the Gram angle field includes: The high-frequency current signal is normalized, and then the normalized high-frequency current signal is converted to a polar coordinate system, and the Gram angle difference field matrix is ​​calculated, and finally the calculated Gram angle difference field matrix is ​​converted into a two-dimensional current image.

[0006] Optionally, the adopting a collaborative contrast reconstruction module to enhance the clarity constraint of the two-dimensional current image includes: The two-dimensional current image is used as the input of the collaborative contrast reconstruction module and sent to the designed end-to-end current image enhancement network generator and parameter estimation network generator respectively to obtain the enhanced current image, characteristic parameters and current change trend characteristic map. Forward reconstruction and loss calculation are performed first, and then reverse loss calculation is performed to combine into collaborative contrast loss. By continuously adjusting the network parameters to minimize the loss, the enhancement and optimization of the two-dimensional current image is achieved.

[0007] Optionally, the dynamically focusing channel attention module is used to dynamically fuse global and local feature information and adaptively assign feature channel weights, including: The enhanced two-dimensional current image is used as the input of the dynamic focusing channel attention module. First, global average pooling is performed to convert it into a channel descriptor. The diagonal matrix is ​​used to capture the dependencies between all channels as global information. The band matrix is ​​used for local channel interaction to obtain local information. The global information obtained by the diagonal matrix and the local information obtained by the band matrix are combined. The correlation matrix of the two at different granularities is obtained through cross-correlation operation. Then, row and column vectors are extracted from the correlation matrix as global current features and local current features. A learnable factor is introduced to dynamically fuse the global current features and local current features to obtain the dynamic focusing features after fusion.

[0008] Optionally, the use of an AdaBoost classifier to implement fault diagnosis of uneven air gap of a motor includes: The dynamic focusing features and the corresponding motor air gap status labels are used as training samples and input into the AdaBoost classifier. First, the parameters of the AdaBoost algorithm are initialized, and the weak classifier is trained based on the current weight distribution. Then, the error rate on the training samples is calculated, and the sample weight is adjusted according to the error rate and the weak classifier weight is determined until the predetermined number of iterations is reached. Finally, all weak classifiers are weighted and combined into a strong classifier to realize the fault diagnosis of uneven motor air gap.

[0009] The present invention provides a motor air gap unevenness diagnosis method based on dynamic focusing channel attention, the method comprises: using Gram angle field to convert high-frequency current signal into two-dimensional current image; introducing collaborative contrast reconstruction constraint between generation network and parameter estimation network, constructing unsupervised collaborative contrast reconstruction module, and enhancing the clarity constraint of two-dimensional current image; using dynamic focusing channel attention module to dynamically fuse global and local feature information, capturing dependency between features through correlation matrix, and adaptively allocating feature channel weights; the present invention makes up for the problem of insufficient current image constraint by constructing collaborative contrast reconstruction module and dynamic focusing channel attention module, realizes more accurate feature weight allocation, and effectively improves the accuracy of motor air gap unevenness diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the embodiments of the present invention or the technical solutions of 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, other drawings can be obtained based on these drawings without paying creative work.

[0011] Figure 1 A schematic diagram of a flow chart of a method for diagnosing uneven air gap in a motor provided by an embodiment of the present invention; Figure 2 A schematic diagram of a dynamic focusing channel attention module of a motor air gap unevenness diagnosis method provided by an embodiment of the present invention; Figure 3 A schematic diagram of the AdaBoost classifier flow of the motor air gap unevenness diagnosis method provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] In order to enable those skilled in the art to better understand the scheme of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than 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 work are within the scope of protection of the present invention.

[0013] like Figure 1 As shown, Figure 1 A flow chart of a method for diagnosing uneven air gap of a motor provided in an embodiment of the present invention, wherein the method for diagnosing uneven air gap of a motor specifically includes four contents.

[0014] S1: The high-frequency current signal is converted into a two-dimensional current image using the Gram angle field.

[0015] It should be noted that the present invention uses the Gram angle field to convert the high-frequency current signal into a two-dimensional current image because the use of the Gram angle field can better retain the time dependence and potential connection characteristics of the high-frequency current signal, so that the pixel distribution in the two-dimensional current image corresponds to the temporal change of the signal, which helps to capture the complex relationship in the current signal. The method specifically includes three steps.

[0016] S11: Normalize the high-frequency current signal so that its amplitude is scaled to the range of (-1, 1). The formula is as follows:

[0017] Where y is the normalized current signal value, x is the original current signal value, and A is the entire high-frequency current signal sequence.

[0018] S12: Convert the normalized current signal value to an angle θ in polar coordinates, calculate the sine value of the angle difference between the current signal values, and generate a Gram angle difference field matrix GADF, the formula of which is as follows:

[0019] In the formula, θ i is the angle corresponding to the i-th current signal value, θ j is the angle corresponding to the jth current signal value.

[0020] S13: Map each element of the calculated Gram angle difference field matrix to a range of grayscale values, thereby generating a two-dimensional current image.

[0021] Based on the above discussion, in an optional embodiment of the present invention, the method of converting the high-frequency current signal into a two-dimensional current image using the Gram angle field specifically includes: In an optional embodiment of S1, during the operation of the motor, a Hall effect current sensor is used to collect the phase current of the three-phase motor, and a high-frequency square wave voltage injection method is adopted. In the first half cycle and the second half cycle of the injected square wave, the high-frequency current signal is extracted by directly subtracting the sampled signal.

[0022] S2: Enhance the clarity constraints of 2D current images using collaborative contrast reconstruction module.

[0023] It should be noted that the present invention adopts a collaborative contrast reconstruction module because the constraints in the collaborative contrast reconstruction process can suppress the random noise in the current image, highlight the true distribution characteristics of the current, and guide the network to more accurately extract features related to the current and avoid extracting interfering features.

[0024] The collaborative contrast reconstruction module is used to enhance the clarity constraint of two-dimensional current images, which specifically includes four steps.

[0025] S21: The two-dimensional current image is used as the input of the collaborative contrast reconstruction module, and the enhanced current image is obtained through the designed end-to-end current image network generator. The current distribution characteristic parameters are extracted through the parameter estimation network generator to construct the current change trend characteristic diagram.

[0026] S22: Combining the obtained enhanced current image and the current change trend feature map to forward reconstruct the original two-dimensional current image, calculating the difference loss between the forward reconstructed image and the input image, and setting a discriminator to distinguish the enhanced current image output by the end-to-end network from the input image to ensure that the generator can output high-quality enhanced current images.

[0027] S23: Reusing the model constructed above, the input image and the current change trend characteristic map are reversely reconstructed into an enhanced current image, and the loss function between the reversely reconstructed enhanced current image and the forwardly reconstructed enhanced current image is calculated.

[0028] S24: The forward reconstructed enhanced current image is used as a positive sample, and the reverse reconstructed enhanced current image and the input image are used as negative samples to construct a contrast constraint for current image enhancement. At the same time, the input image is used as a positive sample, and the forward reconstructed enhanced current image and the reverse reconstructed enhanced current image are used as negative samples to construct another contrast constraint for current image enhancement. The two-dimensional current image enhancement is optimized by continuously adjusting the network parameters to minimize the loss function.

[0029] Based on the above discussion, in an optional embodiment of the present invention, the definition constraint of the two-dimensional current image enhanced by the collaborative contrast reconstruction module specifically includes: In an optional embodiment of S2, during the training phase of the network, adversarial loss, content consistency loss, clarity consistency loss, structural similarity loss and collaborative contrast constraint are used as optimization objectives of the collaborative contrast reconstruction module.

[0030] S3: A dynamic focusing channel attention module is used to dynamically fuse global and local feature information and adaptively assign feature channel weights.

[0031] It should be noted that the present invention introduces a dynamic focusing channel attention module because compared with the traditional channel attention mechanism, the dynamic focusing channel attention module can effectively control the number of model parameters while maintaining information interaction capabilities through the combination of band matrices and diagonal matrices, thereby avoiding excessive computational burden caused by too many parameters. Through adaptive fusion strategies, it can reduce redundant cross-correlation operations and improve computational efficiency.

[0032] like Figure 2 As shown in the figure, a dynamic focusing channel attention module is used to dynamically fuse global and local feature information and adaptively allocate feature channel weights, which specifically includes three steps.

[0033] S31: The enhanced 2D current image is used as the input of the dynamic focus channel attention module and converted into a channel descriptor through a global average pooling operation.

[0034] S32: A diagonal matrix is ​​used to capture the dependencies between all channels as global information, and a band matrix is ​​used for local channel interaction to obtain local information. The global information obtained by the diagonal matrix and the local information obtained by the band matrix are combined, and the correlation matrix of the two at different granularities is obtained through a cross-correlation operation. Then, row and column vectors are extracted from the correlation matrix as global current features and local current features, and a learnable factor is introduced for dynamic fusion. The weight W is calculated as follows:

[0035] In the formula, σ is the sigmoid activation function, b is the learnable factor, and U g is the global current characteristic, U n It is the local current characteristic.

[0036] S33: Multiply the calculated feature channel weights with the global and local current features to obtain dynamic focusing features.

[0037] Based on the above discussion, in an optional embodiment of the present invention, the dynamic focusing channel attention module is used to dynamically fuse global and local feature information and adaptively allocate feature channel weights, specifically including: In an optional embodiment of S3, an Adam optimizer is used for network optimization, a learning rate is set to 0.0001, a batch size is set to 32, and a momentum decay exponent is 0.9.

[0038] S4: AdaBoost classifier is used to diagnose the uneven air gap fault of the motor.

[0039] It should be noted that the present invention introduces the AdaBoost classifier because the AdaBoost classifier iteratively trains a series of weak classifiers. The training and integration process of each weak classifier is relatively simple and suitable for processing large-scale data sets. Finally, these weak classifiers are combined into a strong classifier. This integrated learning method can effectively improve the accuracy of fault diagnosis and make the diagnosis result of motor air gap uneven fault more reliable.

[0040] like Figure 3 As shown in the figure, the AdaBoost classifier is used to diagnose the uneven air gap fault of the motor, which specifically includes three steps.

[0041] S41: The dynamic focusing feature and the corresponding motor air gap state label are used as training samples, and the training samples are then input into the AdaBoost classifier.

[0042] S42: First, the parameters of the AdaBoost algorithm are initialized. In each round of iteration, the weak classifiers are trained according to the weight distribution of the current training samples, the error rate of the weak classifiers on the training samples is calculated, the weights of the weak classifiers are calculated by the error rate, and the weight distribution of the training samples is updated until the predetermined number of iterations is reached, and finally all the weak classifiers are weighted and combined into a strong classifier.

[0043] S43: After the motor operation data to be diagnosed is subjected to the same image processing, feature extraction and fusion, it is input into the AdaBoost classifier to realize the fault diagnosis of uneven air gap of the motor.

[0044] Based on the above discussion, in an optional embodiment of the present invention, the above-mentioned AdaBoost classifier is used to implement the fault diagnosis of uneven air gap of the motor, which specifically includes: In an optional embodiment of S4, the training set and the test set are randomly divided into a ratio of 70%, the training set is used to train the parameters of the AdaBoost classifier, and the test set is used to test the performance of the AdaBoost classifier, and the number of iterations is set to 100.

[0045] It should be noted that, in this document, relational terms such as first and second, etc. are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations.

[0046] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core ideas of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A motor air gap unevenness diagnosis method based on dynamic focusing channel attention, characterized in that: include: The Gram angle field is used to convert high-frequency current signals into two-dimensional current images; The two-dimensional current image is used as the input of the collaborative contrast reconstruction module, and the enhanced current image is obtained through the generator of the designed end-to-end current image network. The current distribution characteristic parameters are extracted through the parameter estimation network generator, and the current change trend characteristic map is constructed. The collaborative contrast learning strategy is adopted, and the collaborative contrast loss function is formed through the combined optimization of the forward reconstruction loss and the reverse reconstruction loss. The network parameters are continuously adjusted through the iterative optimization process to minimize the loss function, thereby achieving continuous enhancement and optimization of the current image. The enhanced two-dimensional current image is used as the input of the dynamic focus channel attention module. First, global average pooling is performed to convert it into a channel descriptor. A diagonal matrix is ​​used to capture the dependencies between all channels as global information. A band matrix is ​​used to perform local channel interactions to obtain local information. The global information obtained by the diagonal matrix and the local information obtained by the band matrix are combined. The correlation matrix of the two at different granularities is obtained through cross-correlation operations. Then, row and column vectors are extracted from the correlation matrix as global current features and local current features. A learnable factor is introduced to dynamically fuse the global current features and local current features, and the dynamic focus features are obtained after fusion. The dynamic focusing feature is input into the AdaBoost classifier to realize the fault diagnosis of uneven air gap in the motor.

2. The motor air gap unevenness diagnosis method as claimed in claim 1, characterized in that: The Gram angle field is used to convert high-frequency current signals into two-dimensional current images, including: The high-frequency current signal is normalized, and then the normalized high-frequency current signal is converted to a polar coordinate system, and the Gram angle difference field matrix is ​​calculated, and finally the calculated Gram angle difference field matrix is ​​converted into a two-dimensional current image.

3. The motor air gap unevenness diagnosis method as claimed in claim 1, characterized in that: The dynamic focus feature is input into the AdaBoost classifier to realize the fault diagnosis of uneven air gap of the motor, including: The dynamic focusing features and the corresponding motor air gap status labels are used as training samples and input into the AdaBoost classifier. First, the parameters of the AdaBoost algorithm are initialized, and the weak classifier is trained based on the current weight distribution. Then, the error rate on the training samples is calculated, and the sample weight is adjusted according to the error rate and the weak classifier weight is determined until the predetermined number of iterations is reached. Finally, all weak classifiers are weighted and combined into a strong classifier to realize the fault diagnosis of uneven motor air gap.

Images