Motor stator and rotor eccentricity defect feature extraction method based on multivariate deep network
Through the combination of stacking image method, Kuppman noise reduction autoencoder, multi-scale nuclear convolutional network, L-Transformer and random dynamic residual network, the defect characteristics of the eccentricity of the motor stator rotor are extracted, and the traditional fault detection cannot be prevented The problem of eccentric defect of the stator rotor achieves high-precision fault diagnosis and early warning.
Patent Information
- Application Number
- CN202411320608.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-23
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2044-09-23
AI Technical Summary
Traditional motor fault detection methods cannot effectively prevent eccentric defects of the fixed rotor, resulting in mechanical vibration and noise, which may cause early failure or unplanned production suspension, causing economic losses to the company.
The stacked image method is used to convert the time domain data of the motor stator current into grayscale image data, and noise and outliers are removed through the Kuppman denoising autoencoder, and features are extracted using a multi-scale nuclear convolutional network, combining L-Transformer and random Dynamic residual network extracts defect characteristics of motor stator eccentricity.
It realizes high-precision and robust motor fault diagnosis, which can early warning of potential faults, provides a basis for subsequent fault diagnosis and maintenance, and reduces the economic losses of motor faults.
Smart Images

Figure CN118839307B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of big data artificial intelligence motor defect feature extraction, and specifically relates to a motor stator and rotor eccentricity defect feature extraction method based on a multivariate deep network. Background Art
[0002] In the production process of motors, due to differences in operator levels and mechanical equipment accuracy, assembly defects such as stator-rotor misalignment, shaft bending and shaft misalignment are easily caused, especially stator-rotor eccentricity, which will cause unbalanced electromagnetic force during motor operation, resulting in more serious mechanical vibration and noise, and may eventually cause early failure of the motor or unplanned shutdown, causing huge economic losses to the company.
[0003] Traditional motor fault detection methods mainly rely on later vibration analysis and fault diagnosis, but this method is often not detected until the fault has occurred or worsened, and cannot effectively prevent it. In recent years, with the development of artificial intelligence technology, defect feature extraction methods based on current signals have gradually become a research hotspot. This type of method analyzes the current data during motor operation and identifies the core characteristics of stator and rotor eccentricity defects. It can provide early warning of potential motor faults and provide a basis for subsequent fault diagnosis and maintenance. Summary of the invention
[0004] The purpose of the present invention is to provide a method for extracting the defect features of motor stator and rotor eccentricity. The method extracts the defect features of motor stator and rotor eccentricity through a multivariate deep network. The extracted defect features can be used for high-precision motor fault diagnosis, which provides an effective method for extracting the defect features of motor stator and rotor eccentricity.
[0005] In order to solve the above technical problems, the present invention provides a method for extracting features of motor stator and rotor eccentricity defects, comprising:
[0006] The collected motor stator current time domain data is converted into grayscale image data using the stacked image method, and the current grayscale image data set is obtained.
[0007] The Koopman denoising autoencoder is used to reconstruct and denoise the current grayscale image dataset samples and remove outliers to obtain a noise-free current grayscale image dataset.
[0008] A multi-scale kernel convolutional network is used to extract multi-scale features of the noise-free current grayscale image dataset;
[0009] The L-Transformer network is used to perform feature adaptation on multi-scale features to obtain adaptive fusion features;
[0010] The adaptive fusion features are input into the random dynamic residual network to extract the defect features of the motor stator and rotor eccentricity.
[0011] Optionally, the collected motor stator current time domain data is converted into grayscale image data by using a stacked image method, and a current grayscale image data set is obtained by sorting, including:
[0012] The original stator current time domain data of the motor is collected when it is running. The stacked image method is used for the original current time domain data. K segments of length are sequentially intercepted from the original time domain data. 2 The segmented samples are normalized from 0 to 255, and the time domain signals are sequentially filled with pixels of the image to obtain a grayscale image with K×K pixels. The converted images are organized to form a current grayscale image dataset.
[0013] Optionally, a Koopman denoising autoencoder is used to reconstruct and denoise the current grayscale image dataset samples and remove outliers to obtain a noise-free current grayscale image dataset, including:
[0014] The samples of the current grayscale image dataset are added with random Gaussian noise and then input into the Koopman denoising autoencoder for training. The Koopman denoising autoencoder uses the reconstructed grayscale image as output and learns a mapping model from noisy data to noise-free data. The samples of the current grayscale image dataset are input into the trained Koopman denoising autoencoder again for reconstruction, denoising and outlier removal to obtain a noise-free current grayscale image dataset.
[0015] Optionally, a multi-scale kernel convolutional network is used to extract multi-scale features of the noise-free current grayscale image dataset, including:
[0016] The samples of the noise-free current grayscale image dataset are input into the multi-scale kernel convolutional network. The multi-scale feature extraction unit of the multi-scale kernel convolutional network contains convolution kernels of 6 scales: N1×N1, N2×N2, N3×N3, N4×N4, N5×N5, and N6×N6. The multi-scale kernel convolutional network extracts the multi-scale features of the noise-free current grayscale image dataset under 6 different scale convolution kernels.
[0017] Optionally, an L-Transformer network is used to perform feature adaptation on multi-scale features to obtain adaptive fusion features, including:
[0018] The multi-scale features are input into the L-Transformer network for feature adaptation. The self-attention layer in the L-Transformer network calculates the attention distribution of the multi-scale features. The feedforward layer in the L-Transformer network removes the redundant information of the multi-scale features. The self-attention layer and feedforward layer of the L-Transformer network run in parallel. Finally, the L-Transformer network outputs the adaptive fusion features.
[0019] Optionally, the adaptive fusion features are input into a random dynamic residual network to extract defect features of the stator and rotor eccentricity of the motor, including:
[0020] The adaptive fusion features are input into the random dynamic residual network, which contains N residual blocks and N gating layers. The gating layer in the random dynamic residual network weights the adaptive fusion features and inputs them into the residual block. The residual blocks are connected with probability 𝑝 jumps. The random dynamic residual network finally outputs the defect features of the stator and rotor eccentricity of the motor.
[0021] The present application discloses a method for extracting features of motor stator and rotor eccentricity defects based on a multivariate deep network, including using a stacked image method to convert the collected motor stator current time domain raw data into grayscale image data, and sorting out a current grayscale image data set; using a Koopman denoising autoencoder to reconstruct and denoise the current grayscale image data set samples and remove outliers to obtain a noise-free current grayscale image data set; using a multi-scale kernel convolution network to extract multi-scale features of the noise-free current grayscale image data set; using an L-Transformer network to perform feature adaptation on the multi-scale features to obtain adaptive fusion features; and inputting the adaptive fusion features into a random dynamic residual network to extract the motor stator and rotor eccentricity defect features. The feature extraction method proposed in the present application has high accuracy and strong robustness, and can extract defect features that are distinguished from other motor defects, providing an effective method for feature extraction of motor stator and rotor eccentricity defects. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] 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.
[0023] Figure 1 A schematic diagram of the process flow of the motor stator and rotor eccentricity defect feature extraction method provided in this application;
[0024] Figure 2 A schematic diagram of the stacked image method process provided for this application;
[0025] Figure 3 The Koopman denoising autoencoder network framework diagram provided for this application;
[0026] Figure 4 The multi-scale convolution kernel network framework diagram provided for this application;
[0027] Figure 5 L-Transformer network framework diagram provided for this application;
[0028] Figure 6 The random dynamic residual network framework diagram provided for this application;
[0029] Figure 7 This is a network flow chart of the method for extracting features of motor stator and rotor eccentricity defects provided in this application. DETAILED DESCRIPTION
[0030] 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.
[0031] like Figure 1 As shown, Figure 1 The present invention provides a flow chart of a method for extracting features of motor stator and rotor eccentricity defects, the method comprising:
[0032] S11: The collected motor stator current time domain raw data is converted into grayscale image data using the stacked image method, and then organized into a current grayscale image data set.
[0033] It should be noted that the stacked image method is introduced in this application to convert the collected motor stator current time domain raw data into grayscale image data. This is because the stacked image method can convert the one-dimensional current time domain signal into a two-dimensional grayscale image without pre-setting any parameters, which can eliminate the influence of expert experience as much as possible, and the two-dimensional grayscale image contains more characteristics of the motor stator and rotor eccentricity defects.
[0034] like Figure 2 As shown, the process of obtaining the current grayscale image data set using the stacked image method includes five steps.
[0035] Step 11: Install the current sensor in the power distribution cabinet of the running motor to collect the stator three-phase current signal with a sampling frequency of N kHz;
[0036] Step 12: Sequentially extract segments of length K from the sampled signal 2 The segmentation samples;
[0037] Step 13: Fill each segmented sample with pixels of a K×K image in order;
[0038] Step 14: Normalize the pixels of the image from 0 to 255, and finally obtain a grayscale image converted from the original data in the current time domain. The normalization calculation formula (1) is shown as follows:
[0039]
[0040] In the above formula represents the value of the split sample, Represents the pixel intensity of the image;
[0041] Step 15: All the obtained grayscale images are sorted according to whether they are motor stator and rotor eccentricity defects to form a noise-free current grayscale image data set.
[0042] Based on the above discussion, in an optional embodiment of the present application, the process of obtaining the current grayscale image data set by using the stacked image method includes:
[0043] In an optional embodiment of S11, the sampling frequency of the current sensor is 10 kHz, and the size of the converted grayscale image is 128×128.
[0044] S12: Use the Koopman denoising autoencoder to reconstruct and denoise the current grayscale image dataset samples and remove outliers to obtain a noise-free current grayscale image dataset.
[0045] It should be noted that the present application introduces the Koopman denoising autoencoder to reconstruct, denoise and remove outliers on samples of the current grayscale image dataset because the encoding and decoding process of the encoder is dynamic and linear, and can still perform well under various input conditions.
[0046] exist Figure 3 middle represents the current grayscale image, represents the current grayscale image after adding random Gaussian noise, Represents the encoding result of the encoder, Applying linear operators , the potential state Propagate forward one step to , represents the input of the decoder, Representing the output of the decoder, the Koopman denoising autoencoder is used to obtain a noise-free current grayscale image dataset, which specifically includes the following four steps.
[0047] Step 21: Define Linear Operator , which can be represented as a complex diagonal matrix with eigenvalues Arranged along its main diagonal, for convenience, the eigenvalues Change to polar coordinates It can be effectively parameterized , and then the corresponding eigenvalue modulo and from Parameters Stacked to their natural rotation matrices In this way, the matrix The real block diagonal form is shown in formula (2):
[0048]
[0049] Step 22: Define the total loss function during training, as shown in formula (3):
[0050]
[0051] The first term on the right side of the above equation is called the reconstruction error, which represents the current grayscale image containing Gaussian random noise After encoding and reconstruction, the original image The second term represents the underlying linear dynamics condition of the Koopman denoising autoencoder, the operator The distance that satisfies the equation; the third loss term is the prediction error; the last loss term is the regularization term of the encoder / decoder network parameters; the calculation parameters Used to adjust the impact of each error term on the overall loss function; Represents the code, Decoder
[0052] Step 23: Convert the Current Grayscale Image Each pixel is added with Gaussian random noise with mean M and variance S to obtain , follow the above steps to define the Koopman autoencoder and , and finally Input the encoder training;
[0053] Step 24: After the training is completed, the current grayscale image dataset samples are input into the Koopman denoising autoencoder again for reconstruction, denoising and outlier removal to obtain pure current grayscale image data, which are sorted according to whether the fault is the stator and rotor eccentricity of the motor to form a noise-free current grayscale image dataset.
[0054] Based on the above discussion, in an optional embodiment of the present application, the process of reconstructing and denoising the current grayscale image dataset samples and removing outliers using the Koopman denoising autoencoder to obtain a noise-free current grayscale image dataset includes:
[0055] In an optional embodiment of S12, the input dimension of the Koopman denoising autoencoder is 128*128, the encoding result size is 16*16, and the linear operator It contains 64 eigenvalues, the input epoch is 256, the Adam optimization algorithm is used, and the learning rate is 0.001.
[0056] S13: A multi-scale kernel convolutional network is used to extract multi-scale features of the noise-free current grayscale image dataset.
[0057] It should be noted that this application introduces a multi-scale kernel convolution network to extract multi-scale features of the motor noise-free current grayscale image dataset because the convolution layer of the multi-scale kernel convolution network contains 6 convolution kernels of different sizes. These four convolution kernels can extract multi-scale features of the training set samples under 6 different receptive fields. Compared with the features extracted by single-scale convolution, the multi-scale features contain richer motor stator and rotor eccentricity fault features.
[0058] The multi-scale kernel one-dimensional convolutional network structure is as follows Figure 4 As shown, the network framework consists of convolution layer, splicing layer and pooling layer in sequence, wherein the convolution layer contains convolution kernels of 6 scales, namely N1×N1, N2×N2, N3×N3, N4×N4, N5×N5 and N6×N6, and the pooling layer is the maximum pooling layer. The network extracts multi-scale features of the noise-free current grayscale image dataset, specifically including the following 3 steps.
[0059] Step 31: Input the noise-free current grayscale image data set samples into the network for multi-scale feature extraction. The samples enter the convolution layer of the network. The convolution layer has 6 convolution kernels, namely N1×N1, N2×N2, N3×N3, N4×N4, N5×N5, and N6×N6, which perform 4 convolutions on the samples in parallel to obtain sample features under 6 receptive fields.
[0060] Step 32: The concatenation layer flattens the six sample features output by the convolution layer and concatenates them end to end in sequence;
[0061] Step 33: The pooling layer performs maximum pooling on the output of the concatenation layer with a step size of 1×M1 to obtain multi-scale features of the noise-free current grayscale image dataset samples.
[0062] Based on the above discussion, in an optional embodiment of the present application, the multi-scale feature setting process for extracting the noise-free current grayscale image data set using the multi-scale kernel convolution network may include:
[0063] In an optional embodiment of S13, the sizes of the six convolution kernels N1×N1, N2×N2, N3×N3, N4×N4, N5×N5, and N6×N6 in the network convolution layer are 1×1, 3×3, 5×5, 7×7, 9×9, and 11×11 respectively, the step size of the pooling layer is 1×5, and the input epoch is 256.
[0064] S14: The L-Transformer network is used to perform feature adaptation on multi-scale features to obtain adaptive fusion features.
[0065] It should be noted that the present application introduces the L-Transformer network to perform feature adaptation on multi-scale features because the self-attention layer of the L-Transformer network can explore the correlations between multi-scale features and enhance the features related to the eccentricity defects of the motor stator and rotor. The self-attention layer of the L-Transformer network uses queries and keys to calculate attention. The self-attention layer and feedforward layer of the L-Transformer run in parallel. Therefore, compared with other self-attention networks, the L-Transformer network has higher computational efficiency under conditions of similar accuracy.
[0066] The L-Transformer network structure is as follows Figure 5 As shown, the network has N feature encoding blocks, and each block contains a self-attention layer and a feedforward layer running in parallel, wherein the self-attention layer uses Query and Value to calculate the attention between multi-scale features, and the feedforward layer contains two fully connected layers and a regularization layer for removing redundant information in multi-scale features. The network performs feature adaptation on multi-scale features to obtain adaptive fusion features, which specifically includes three steps.
[0067] Step 41: Input the multi-scale features into the network for feature adaptation. The multi-scale features enter the feature encoding block of the network. The self-attention layer in the feature encoding block uses Query and Value to calculate the attention between the multi-scale features to obtain the attention encoding of the multi-scale features. The calculation is shown in the following formula:
[0068]
[0069] In the formula, is a multi-scale feature, are the query and key parameters, respectively. is the attention encoding of multi-scale features, Is a mask matrix. The value range of the matrix is ;
[0070] Step 42: The multi-scale features enter the feedforward layer, and the multi-scale features are processed by two fully connected layers and one regularization layer in the feedforward layer to output a feedforward vector;
[0071] Step 43: The attention code and the feedforward vector are concatenated head to tail as the input of the next feature coding block. The multi-scale features are processed in sequence by N feature coding blocks to obtain adaptive fusion features.
[0072] Based on the above discussion, in an optional embodiment of the present application, the process of performing feature adaptation on multi-scale features using the L-Transformer network may include:
[0073] In an optional embodiment of S14, the number of feature encoding blocks of the L-Transformer network is 6, the input and output of the attention layer are 512*1 dimensions, the random inactivation rate of the fully connected layer in the feedforward layer is 0.2, and the input epoch is 256.
[0074] S15: Input the adaptive fusion features into the random dynamic residual network to extract the defect features of the motor stator and rotor eccentricity.
[0075] It should be noted that the present application introduces a random dynamic residual network to extract the defect characteristics of motor stator and rotor eccentricity because the random dynamic residual network adopts a dynamic adjustment mechanism and random connection technology during the training process. The network can dynamically adjust its parameters and structure according to the adaptive input, so that compared with ordinary feature extraction networks, the network converges faster, is more robust and has higher accuracy.
[0076] The random dynamic residual network structure is as follows Figure 6 As shown, the network has N residual blocks and N gating layers, each residual block is preceded by a gating layer, each residual block contains two fully connected layers and a regularization layer, wherein the residual blocks are randomly dynamically jumped connected, and the network extracts the motor stator and rotor eccentricity defect features in the adaptive fusion features, specifically including 4 steps.
[0077] Step 51: Input the adaptive fusion features into the random dynamic residual network, and the adaptive fusion features enter the gating layer. The gating layer calculates the dynamic weight W through forward propagation, and the gating layer outputs the weighted features. The calculation formula is shown in the following formula:
[0078]
[0079]
[0080] In the above formula, and is the weight matrix of the gating layer, and is the bias vector of the gating layer, Activation functions ensure dynamic weights Between (0,1), is the weighted adaptive fusion feature, represents element-wise product;
[0081] Step 52: Weighted adaptive fusion features Input residual blocks and define skip probability for each residual block , during the training process, with probability Skip this residual block and generate a random number for each residual block during each forward propagation process ,if , then skip the residual block and pass the input directly to the output, otherwise, pass the residual block normally;
[0082] Step 53: When the input enters the residual block normally, it enters the two fully connected layers and one regularization layer in the residual block in sequence and then outputs the residual features;
[0083] Step 54: The residual features continue to be input into the next gated layer and residual block, and the adaptive fusion features are extracted in sequence through N gated layers and residual blocks, and finally the defect features of the stator and rotor eccentricity of the motor are obtained.
[0084] Based on the above discussion, in an optional embodiment of the present application, the process of inputting the adaptive fusion feature into the random dynamic residual network to extract the defect feature of the stator and rotor eccentricity of the motor includes:
[0085] In an optional embodiment of S15, the number of residual blocks of the random dynamic residual network is 16, and the dynamic weight of the gated layer is The function is normalized, the random skip rate 𝑝 between residual blocks is 0.5, and the input epoch is 256.
[0086] Figure 7 The network flow chart of the motor stator and rotor eccentricity defect feature extraction method provided in this application, the application first uses the stacked image method to convert the collected motor stator current time domain raw data into grayscale image data, and organizes it to form a current grayscale image data set; then, the Koopman denoising autoencoder is used to reconstruct the current grayscale image data set samples for denoising and remove outliers to obtain a noise-free current grayscale image data set; then, a multi-scale kernel convolution network is used to extract the multi-scale features of the noise-free current grayscale image data set; the multi-scale features are adaptively processed through the L-Transformer network to obtain adaptive fusion features; finally, the adaptive fusion features are input into the random dynamic residual network to extract the motor stator and rotor eccentricity defect features. The feature extraction method proposed in this application has high accuracy and strong robustness, and can extract defect features that are distinguished from other motor defects, providing an effective method for feature extraction of motor stator and rotor eccentricity.
[0087] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Specific examples are used herein to illustrate the principles and implementation methods of the present invention, and the description of the above embodiments is only used to help understand the method of the present invention and its core idea. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made to 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 method for extracting motor stator and rotor eccentricity defects based on a multivariate deep network. It is characterized in that include: The original stator current time domain data of the motor is collected when it is running. The stacked image method is used for the original current time domain data. K segments of length are sequentially intercepted from the original time domain data. 2 The sample segments are normalized from 0 to 255, and the time domain signals are sequentially filled with pixels of the image to obtain a grayscale image with K×K pixels. The converted images are organized to form a current grayscale image dataset; The samples of the current grayscale image dataset are added with random Gaussian noise and then input into the Koopman denoising autoencoder for training. The Koopman denoising autoencoder uses the reconstructed grayscale image as output and learns a mapping model from noisy data to noise-free data. The samples of the current grayscale image dataset are input into the trained Koopman denoising autoencoder again for reconstruction, denoising and outlier removal to obtain a noise-free current grayscale image dataset. The noise-free current grayscale image dataset samples are input into the multi-scale kernel convolutional network. The multi-scale feature extraction unit of the multi-scale kernel convolutional network contains N 1 ×N 1 、N 2 ×N 2 、N 3 ×N 3 、N 4 ×N 4 、N 5 ×N 5 、N 6 ×N 6 There are 6 scales of convolution kernels in total. The multi-scale kernel convolution network extracts multi-scale features of the noise-free current grayscale image dataset under 6 different scales of convolution kernels; The L-Transformer network is used to perform feature adaptation on multi-scale features to obtain adaptive fusion features; The adaptive fusion features are input into the random dynamic residual network to extract the defect features of the motor stator and rotor eccentricity.
2. The method for extracting the characteristics of the stator and rotor eccentricity defects of a motor as claimed in claim 1, It is characterized in that An L-Transformer network is used to perform feature adaptation on multi-scale features to obtain adaptive fusion features, including: inputting multi-scale features into the L-Transformer network for feature adaptation, the self-attention layer in the L-Transformer network calculating the attention distribution of the multi-scale features, the feedforward layer in the L-Transformer network removing redundant information of the multi-scale features, the self-attention layer and the feedforward layer of the L-Transformer network running in parallel, and the L-Transformer network finally outputting the adaptive fusion features.
3. The method for extracting the characteristics of the stator and rotor eccentricity defects of a motor as claimed in claim 1, It is characterized in that The adaptive fusion features are input into a random dynamic residual network to extract the defect features of the motor stator and rotor eccentricity, including: the adaptive fusion features are input into the random dynamic residual network, the random dynamic residual network includes N residual blocks and N gating layers, the gating layer in the random dynamic residual network weights the adaptive fusion features and then inputs the residual blocks, the residual blocks are connected by probabilistic jumps, and the random dynamic residual network finally outputs the defect features of the motor stator and rotor eccentricity.
Citation Information
Patent Citations
Digital printing fabric defect detection method based on generative adversarial network
CN117237274A
Failure diagnosis method for power transformer winding based on gsmallat-NIN-CNN network
US20210382120A1
Bearing defect identification method based on SDAE and improved GWO-svm
WO2021128510A1
