Helical gear expansion diagnosis method and system based on three-dimensional vibration fusion angular field image

Abstract

The present invention discloses a helical gear expansion diagnosis method and system based on three-way vibration fusion angular field images, which belongs to the field of mechanical fault diagnosis. The Gram angle difference field fused by the three-directional measurement data of the acceleration sensor is used to realize the conversion of the vibration signal into the angular field image. The image after the three-channel data fusion can highlight the operating status characteristics of the helical gear in terms of color and texture. A generative adversarial network model with deep feature fusion of a dual-branch discriminator is designed to improve the expansion quality of small sample data. A stacked block convolutional neural network model with an inverted triangle channel distribution is designed. By matching the number of channels input at each stage, the feature extraction capability of the fault diagnosis model can be improved. The present invention realizes intelligent and accurate fault diagnosis of helical gears, and solves the problem of inaccurate fault diagnosis caused by insufficient sample quantity, unremarkable sample features and insufficient model feature extraction capability in conventional methods.

Description

Technical Field

[0001] The present invention belongs to the field of mechanical fault diagnosis, particularly to the field of helical gear fault diagnosis, and particularly to a helical gear expansion diagnosis method and system based on three-way vibration fusion angular field images. Background Art

[0002] Gearboxes are a crucial component of industrial equipment, boasting advantages such as high transmission efficiency and strong load-bearing capacity, and are widely used in various industrial fields. However, due to the complex operating conditions of gearboxes, the gear failure rate during transmission is high, posing a potential safety hazard to the equipment. Helical gears offer a smoother transmission than spur gears, but their stress characteristics make it unreliable to use unidirectional sensors to represent the operating status. When monitoring the condition of mechanical equipment, it is difficult to obtain large amounts of multi-type fault data in a short period of time, as mechanical equipment spends most of its time in a healthy or normally degraded state. Furthermore, the collection and manual annotation of big data is time-consuming and labor-intensive. Therefore, data-driven gear fault diagnosis technology needs to address the problem of ineffective deep learning with small samples.

[0003] Mechanical equipment is generally very sensitive to vibration signals, and vibration is the primary means of mechanical fault diagnosis. However, vibration signals are often affected by complex working conditions and environmental interference, and exhibit non-stationary characteristics. Images can compensate for the shortcomings of one-dimensional signals such as vibration in expressing the operating characteristics of equipment, and can significantly enhance the expression of the operating characteristics of mechanical equipment, which is conducive to accurate fault diagnosis. In addition, intelligent fault diagnosis based on convolutional neural networks is currently one of the mainstream methods for gear fault diagnosis. However, deep learning models such as convolutional neural networks have data adaptability issues. In other words, intelligent mechanical fault diagnosis models based on deep learning methods are difficult to generalize and need to be trained and verified according to the specific characteristics of different data to meet the actual needs of engineering. Summary of the Invention

[0004] The purpose of the present invention is to provide a helical gear expansion diagnosis method and system based on three-dimensional vibration fusion angular field images, to overcome the problems of sensor direction measurement discrepancies and the difficulty of deep learning with small samples in helical gear fault diagnosis. This method designs a three-dimensional vibration fusion Gram angular difference field image conversion method, and uses a generative adversarial network model with deep feature fusion using a dual-branch discriminator to expand the sample size of small sample data sets. Combined with a stacked block convolutional neural network fault diagnosis model with an inverted triangular channel distribution and strong generalization capability, this method achieves intelligent and accurate fault diagnosis of helical gears with small sample sizes.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The helical gear expansion diagnosis method based on three-dimensional vibration fusion angular field image includes the following steps:

[0007] 1) Constructing a helical gear fault feature dataset

[0008] A three-dimensional accelerometer is used to collect the horizontal radial, vertical radial, and axial vibration signals of the helical gear to obtain a three-dimensional vibration signal. The three-dimensional vibration signal is segmented and approximated according to the specified number of points and normalized to convert it into a GADF image. The pixels of the GADF image corresponding to the three directions are then normalized to between 0 and 255, and their colors are mapped to R, G, and B respectively to create an RGB-GADF image. The RGB-GADF image is downsampled to a preset size to obtain a new image. The new image is manually labeled to establish a helical gear fault feature dataset and divided into a training set and a test set.

[0009] 2) Build a sample expansion model

[0010] A sample expansion model is constructed using a helical gear fault feature dataset. The sample expansion model includes a generator and a discriminator. The generator is composed of multiple layers of deconvolution and adds batch normalization and activation functions. The input is random noise, and the output is a test image of the same size as the RGB-GADF image. The discriminator is a dual-branch deep feature fusion structure. The dual-branch discriminator is composed of multiple layers of convolution and adds batch normalization and activation functions. The deep features extracted by the dual-branch discriminator are divided into adaptive average pooling and adaptive maximum pooling, and the deep overall features and significant features extracted by the two are spliced ​​into a vector and input into a multi-layer perceptron composed of several fully connected layers. Finally, the multi-layer perceptron outputs the result and performs gradient update through loss function feedback. The discriminator inputs the new sample output by the generator and the training sample in the training set, and outputs the sample discrimination probability. The weight parameter of the optimal expansion model is selected and saved using the indicator for evaluating the sample structure similarity. The weight parameter is called for sample expansion. The new training set is composed of a mixture of generated samples and original training samples, namely the helical gear fault feature dataset;

[0011] 3) Build a fault diagnosis model

[0012] A stacked block convolutional neural network fault diagnosis model with an inverted triangular channel distribution is constructed. The stacked block convolutional neural network fault diagnosis model with an inverted triangular channel distribution includes a feature extractor and a classifier. The feature extractor includes four-stage stacked block layers. The stacked block convolutional neural network fault diagnosis model with an inverted triangular channel distribution is trained using a helical gear fault feature dataset. After training, the optimal fault diagnosis model weight parameters are saved to obtain a fault diagnosis model.

[0013] 4) Fault diagnosis

[0014] According to step 1), the helical gear test data is collected and converted into RGB-GADF images as a test set. The fault diagnosis model is used to perform fault diagnosis on the test set data to complete the fault diagnosis of the helical gear.

[0015] Furthermore, in step 1), a helical gear fault feature dataset is constructed, specifically:

[0016] 1.1) A three-axis accelerometer is used to collect the horizontal radial, vertical radial, and axial vibration signals of the helical gear. The three-axis vibration signals are subjected to PAA at a specified number of points, normalized, and then converted into a GADF image. PAA segments the vibration signal using a pre-set local window, calculates the average value of each segmented vibration signal, and reconstructs the time series data using the average value of all segments. The GADF image conversion process is as follows:

[0017] ① For time series data X={x1,x2,…,x n} to normalize and transform the amplitude to [-1,1]. The calculation formula is as follows:

[0018]

[0019] Where: x i is the i-th value of the time series data; is the i-th value of the time series data after normalization; max(X) is the maximum value of the time series data; min(X) is the minimum value of the time series data;

[0020] ② Convert the normalized value into the cosine of the angle through calculation, use the timestamp as the radius, and convert the transformed time series data from the Cartesian coordinate system to the polar coordinate system. The calculation formulas for the polar angle and polar diameter in polar coordinates are as follows:

[0021]

[0022] Where: is the i-th value of the normalized time series data; for The corresponding polar angle; t i is the timestamp corresponding to the i-th value of the time series data; N is the constant factor for adjusting the span of the polar coordinate system; r i is the polar radius corresponding to the i-th value of the time series data;

[0023] ③ Use trigonometric function difference to represent the correlation between each time point of the GADF image. During the conversion, the GADF image is established by calculating the angle difference of the sine function. The GADF image calculation is as follows:

[0024]

[0025] Where GADF is the GADF image; sin is the sine function; φ1 is the polar angle corresponding to the first data point; φ n The polar angle corresponding to the nth data point;

[0026] 1.2) Normalize the GADF image pixels corresponding to the three directions to between 0 and 255, and map their colors to R, G, and B respectively. Input the image information corresponding to the three directions into the three channels of the RGB image to construct the RGB-GADF image;

[0027] 1.3) Scale the converted RGB-GADF image to 256×256 pixels;

[0028] 1.4) All the obtained RGB-GADF images are labeled according to different helical gear states and divided into training and test sets according to preset requirements.

[0029] Furthermore, the designated number of points is set according to the sampling frequency and the rotational speed of the shaft on which the helical gear is located, and each signal segment of the designated number of points contains at least 1 / 4 cycle information.

[0030] Furthermore, in step 1.1), the local window size preset by PAA is 2.

[0031] Furthermore, in step 1.4), the training set is preset to have x RGB-GADF images per category, and the test set is preset to have y RGB-GADF images per category, where x ≥ y.

[0032] Furthermore, in step 2), a sample expansion model is constructed, specifically:

[0033] 2.1) Using the helical gear fault feature dataset, a sample expansion model is constructed. The sample expansion model includes a generator and a discriminator. The generator is composed of multiple layers of deconvolution and incorporates batch normalization and activation functions. The input is random noise, and the output is a new image of the same size as the RGB-GADF image.

[0034] 2.2) The discriminator is a two-branch deep feature fusion structure. The two-branch discriminator is composed of multiple layers of convolution and adds batch normalization and activation functions. The deep features extracted by the two-branch discriminator are shunted into adaptive average pooling and adaptive maximum pooling. The deep overall features and significant features extracted by the two are spliced ​​into a vector and input into a multi-layer perceptron composed of several fully connected layers. The multi-layer perceptron finally outputs the result and performs gradient update through loss function feedback. The discriminator input is the new sample output by the generator and the training samples in the training set, and the output is the sample discrimination probability;

[0035] 2.3) Using the index for evaluating sample structure similarity, the weight parameters of the optimal expansion model are selected and saved. The weight parameters are used to expand the sample. The new training set is composed of a mixture of generated samples and original training samples.

[0036] Furthermore, in step 3), a fault diagnosis model is constructed, specifically:

[0037] 3.1) Constructing a stacked block convolutional neural network fault diagnosis model with an inverted triangle channel distribution, wherein the stacked block convolutional neural network fault diagnosis model with an inverted triangle channel distribution includes a feature extractor and a classifier, wherein the feature extractor includes four-stage stacked block layers;

[0038] 3.2) Use the new training set expanded in step 2) to train the fault diagnosis model. Update the weight parameters of the fault diagnosis model based on its performance on the training set. Use the classification accuracy of the fault diagnosis model on the test set as the evaluation indicator. The classification accuracy calculation formula is:

[0039]

[0040] Where: A represents the classification accuracy of the test set, P represents the number of correctly classified samples in the test set, and Q represents the total number of test set samples;

[0041] 3.3) In the model testing phase, the test set in 3.2) is substituted into the fault diagnosis model, and the classification accuracy of the fault diagnosis model on the test set is used as the evaluation indicator. When the classification accuracy is greater than or equal to 90%, the optimal fault diagnosis model weight parameters are saved.

[0042] Furthermore, the stacking ratio of the number of layers of the stacked block layers in the four stages is 3:3:9:3, and the number of input channels decreases successively, namely 128, 64, 32 and 16.

[0043] Furthermore, the fault diagnosis in step 4) is specifically as follows:

[0044] 4.1) Using the operation in step 1), collect the vibration signal of the helical gear, convert the vibration signal into an RGB-GADF image and use it as the test set data;

[0045] 4.2) Use the fault diagnosis model trained with the new training set expanded in step 3), import the fault diagnosis weights, and input the test set data to complete the fault diagnosis of the helical gear.

[0046] The helical gear expansion diagnosis system based on three-way vibration fusion angular field images includes:

[0047] Helical gear fault feature dataset construction module: It is used to use a three-dimensional acceleration sensor to collect the horizontal radial, vertical radial and axial vibration signals of the helical gear to obtain the three-dimensional vibration signals. The three-dimensional vibration signals are segmented and approximated according to the specified number of points and normalized before being converted into GADF images. The pixels of the GADF images corresponding to the three directions are then normalized to between 0 and 255, and their colors are mapped to R, G, and B respectively to establish an RGB-GADF image. The RGB-GADF image is downsampled to a preset size to obtain a new image. After manually labeling the new image, the helical gear fault feature dataset is established and divided into a training set and a test set.

[0048] Sample expansion model construction module: used to construct a sample expansion model using the helical gear fault feature dataset. The sample expansion model includes a generator and a discriminator. The generator is composed of multiple layers of deconvolution and adds batch normalization and activation function. The input is random noise and the output is a test image with the same size as the RGB-GADF image. The discriminator is a dual-branch deep feature fusion structure. The dual-branch discriminator is composed of multiple layers of convolution and adds batch normalization and activation function. The deep features extracted by the dual-branch discriminator are shunted to the adaptive average pooling and adaptive maximum pooling, and the The deep overall features and significant features extracted by the two are spliced ​​into a vector and input into a multi-layer perceptron composed of several fully connected layers. The multi-layer perceptron finally outputs the result and performs gradient update through loss function feedback. The discriminator inputs the new samples output by the generator and the training samples in the training set, and outputs the sample discrimination probability. The weight parameters of the optimal expansion model are selected and saved using the indicator for evaluating the similarity of the sample structure. The weight parameters are called to expand the sample. The new training set is composed of a mixture of the generated samples and the original training samples, namely the helical gear fault feature dataset.

[0049] A fault diagnosis model construction module is used to construct a stacked block convolutional neural network fault diagnosis model with an inverted triangle channel distribution. The stacked block convolutional neural network fault diagnosis model with an inverted triangle channel distribution includes a feature extractor and a classifier. The feature extractor includes four-stage stacked block layers. The stacked block convolutional neural network fault diagnosis model with an inverted triangle channel distribution is trained using a helical gear fault feature dataset. After training, the optimal fault diagnosis model weight parameters are saved to obtain a fault diagnosis model.

[0050] Fault diagnosis module: It is used to collect the helical gear test data according to the method of building a module of the helical gear fault feature dataset, and convert it into RGB-GADF images as a test set. The fault diagnosis model is used to perform fault diagnosis on the test set data to complete the fault diagnosis of the helical gear.

[0051] Compared with the prior art, the present invention has the following beneficial technical effects:

[0052] 1) To address the problem that vibration signals collected by single-directional sensors cannot fully characterize the operating state of helical gears, a sample conversion method is proposed to characterize helical gear fault characteristics using RGB-GADF images fused from three-directional vibration sensor data. The RGB-GADF images highlight the operating characteristics of different helical gear states through color and texture, avoiding the influence of the number of sensors installed and the measurement direction, making them suitable for helical gear fault diagnosis.

[0053] 2) In order to solve the problem that small samples are difficult to learn in depth in helical gear fault diagnosis, an intelligent fault diagnosis model combining a generative adversarial network with a stacked block convolutional neural network is proposed. A generative adversarial network model with deep feature fusion of a dual-branch discriminator is designed. By fusing the deep overall information and salient information of the discriminator, the sample expansion quality is improved compared with the traditional generative adversarial network. The matching design of the number of input channels of the stacked block convolutional neural network fault diagnosis model improves the generalization ability and diagnostic accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The drawings in the specification are used to provide further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0055] Figure 1 This is a flow chart of the helical gear expansion diagnosis method based on three-way vibration fusion angular field images;

[0056] Figure 2 RGB-GADF images generated from the three-dimensional vibration signals of helical gears in different states, where (a) is the normal state, (b) is a local tooth breakage, (c) is tooth surface pitting, (d) is tooth surface wear, and (e) is a tooth root crack.

[0057] Figure 3 Schematic diagram of the sample expansion model, where (a) is the generator structure diagram and (b) is the discriminator structure diagram;

[0058] Figure 4 Schematic diagram of generated samples, where (a) is the normal state of dataset A after expansion, (b) is the local tooth breakage after expansion of dataset A, (c) is the tooth surface pitting after expansion of dataset A, (d) is the tooth surface wear after expansion of dataset A, (e) is the tooth root crack after expansion of dataset A, (f) is the normal state of dataset B after expansion, (g) is the local tooth breakage after expansion of dataset B, (h) is the tooth surface pitting after expansion of dataset B, (i) is the tooth surface wear after expansion of dataset B, and (j) is the tooth root crack after expansion of dataset B;

[0059] Figure 5 To construct a schematic diagram of the fault diagnosis model;

[0060] Figure 6 The corresponding diagnosis results before and after expansion for different data sets, where (a) is the diagnosis result of data set A before expansion, (b) is the diagnosis result of data set A after expansion, (c) is the diagnosis result of data set B before expansion, and (d) is the diagnosis result of data set B after expansion;

[0061] Figure 7 Schematic diagram comparing the diagnosis results of the fault diagnosis model of the present invention and the ConvNeXt-Tiny model, where (a) is the diagnosis result of the ConvNeXt-Tiny model and (b) is the diagnosis result of the present invention. DETAILED DESCRIPTION

[0062] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0063] It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments of the invention described herein can be implemented in sequences other than those illustrated or described herein. In addition, the terms "comprises" and "comprising" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus that includes a series of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, product, or apparatus.

[0064] Example 1

[0065] The helical gear expansion diagnosis method based on three-dimensional vibration fusion angular field image includes the following steps:

[0066] 1) Constructing a helical gear fault feature dataset: A three-dimensional accelerometer is used to collect the horizontal radial, vertical radial, and axial vibration signals of the helical gear to obtain three-dimensional vibration signals. The three-dimensional vibration signals are piecewise aggregated and approximated (PAA) according to the specified number of points and normalized before being converted into Gramianangula difference field (GADF) images. The pixels of the GADF images corresponding to the three directions are then normalized to between 0 and 255 and their colors are mapped to red (R), green (G), and blue (B) respectively to establish RGB-GADF images. The RGB-GADF images are then downsampled to the image input size required by the sample expansion model and the fault diagnosis model to obtain new images. The new images are manually labeled to establish the helical gear fault feature dataset and divided into a training set and a test set.

[0067] The specific steps for constructing the helical gear fault feature dataset in step 1) are as follows:

[0068] 1.1) A three-axis accelerometer is used to collect the horizontal radial, vertical radial, and axial vibration signals of the helical gear. The three-axis vibration signals are subjected to PAA at a specified number of points, normalized, and then converted into a GADF image. PAA segments the vibration signal using a pre-set local window, calculates the average value of each segmented vibration signal, and reconstructs the time series data using the average value of all segments. The GADF image conversion process is as follows:

[0069] ① For time series data X={x1,x2,…,x n} to normalize and transform the amplitude to [-1,1]. The calculation formula is as follows:

[0070]

[0071] Where: x i is the i-th value of the time series data; is the i-th value of the time series data after normalization; max(X) is the maximum value of the time series data; min(X) is the minimum value of the time series data.

[0072] ② Convert the normalized value into the cosine of the angle through calculation, use the timestamp as the radius, and convert the transformed time series data from the Cartesian coordinate system to the polar coordinate system. The calculation formulas for the polar angle and polar diameter in polar coordinates are as follows:

[0073]

[0074] Where: is the i-th value of the normalized time series data; for The corresponding polar angle; t i is the timestamp corresponding to the i-th value of the time series data; N is the constant factor for adjusting the span of the polar coordinate system; r i is the polar radius corresponding to the i-th value of the time series data.

[0075] ③ Use trigonometric function difference to represent the correlation between each time point of the GADF image. During the conversion, the GADF image is established by calculating the angle difference of the sine function. The GADF image calculation is as follows:

[0076]

[0077] Where GADF is the GADF image; sin is the sine function; φ1 is the polar angle corresponding to the first data point; φ n The polar angle corresponding to the nth data point.

[0078] 1.2) Normalize the GADF image pixels corresponding to the three directions to between 0 and 255, and map their colors to red, green, and blue, respectively. Input the image information corresponding to the three directions into the three channels of the RGB image to construct the RGB-GADF image;

[0079] 1.3) Scale the converted RGB-GADF image to 256×256 pixels;

[0080] 1.4) All obtained RGB-GADF images are labeled according to different helical gear states and divided into training and test sets according to preset requirements. The training set is preset to have x RGB-GADF images per class, and the test set is preset to have y RGB-GADF images per class, where x ≥ y, "class" refers to helical gears with different fault types, and x and y refer to the number of training and test sets per class, respectively.

[0081] 2) Constructing a sample expansion model: Constructing a helical gear fault feature dataset through the operation in step 1), and constructing a sample expansion model using the helical gear fault feature dataset. The sample expansion model includes a generator and a discriminator. The generator is composed of multiple layers of deconvolution and adds batch normalization and activation function. The input is random noise, and the output is a test image with the same size as the RGB-GADF image. The discriminator is a dual-branch deep feature fusion structure. The dual-branch discriminator is composed of multiple layers of convolution and adds batch normalization and activation function. The deep features extracted by the dual-branch discriminator are diverted to the adaptive average pooling and Adaptive maximum pooling is performed, and the deep overall features and significant features extracted by the two are spliced ​​into a vector, which is input into a multi-layer perceptron composed of several fully connected layers. The multi-layer perceptron finally outputs the result and performs gradient update through loss function feedback. The discriminator inputs the new samples output by the generator and the training samples in the training set, and outputs the sample discrimination probability. The weight parameters of the optimal expansion model are selected and saved using the indicator for evaluating the similarity of the sample structure. The weight parameters are called to expand the sample. The new training set is composed of a mixture of the generated samples and the original training samples, namely the helical gear fault feature dataset;

[0082] The specific steps for building the sample expansion model in step 2) are as follows:

[0083] 2.1) Constructing a helical gear fault feature dataset through the operations in step 1), and constructing a sample expansion model using the helical gear fault feature dataset. The sample expansion model includes a generator and a discriminator. The generator is composed of multiple layers of deconvolution and incorporates batch normalization and an activation function. The input of the generator is random noise, and the output is a new image of the same size as the RGB-GADF image.

[0084] 2.2) The discriminator is a dual-branch deep feature fusion structure. The dual-branch discriminator is composed of multiple layers of convolution and adds batch normalization and activation functions. The deep features extracted by the dual-branch discriminator are shunted into adaptive average pooling and adaptive maximum pooling. The deep overall features and significant features extracted by the two are spliced ​​into a vector input into a multi-layer perceptron composed of several fully connected layers. The multi-layer perceptron finally outputs the result and performs gradient update through loss function feedback. The discriminator input is the new sample output by the generator and the training samples in the training set, and the output is the sample discrimination probability;

[0085] 2.3) Using the index for evaluating sample structure similarity, the weight parameters of the optimal expansion model are selected and saved. The weight parameters are used to expand the sample. The new training set is composed of a mixture of generated samples and original training samples.

[0086] 3) Construct a fault diagnosis model: Use the helical gear fault feature dataset constructed in step 2) to construct a stacked block convolutional neural network fault diagnosis model with an inverted triangle channel distribution. The stacked block convolutional neural network fault diagnosis model with an inverted triangle channel distribution includes a feature extractor and a classifier. The feature extractor includes four-stage stacked block layers. The number of stacking times of the four-stage stacked block layers is 3:3:9:3, and the number of input channels decreases successively, namely 128, 64, 32 and 16. The stacked block convolutional neural network fault diagnosis model with an inverted triangle channel distribution is trained using the helical gear fault feature dataset. After training, the optimal fault diagnosis model weight parameters are saved to obtain a fault diagnosis model.

[0087] The specific steps for building a fault diagnosis model in step 3 are as follows:

[0088] 3.1) constructing a stacked block convolutional neural network fault diagnosis model with an inverted triangular channel distribution, wherein the stacked block convolutional neural network fault diagnosis model with an inverted triangular channel distribution includes a feature extractor and a classifier, wherein the feature extractor includes four stages of stacked block layers, wherein the number of stacking times of the four stages of stacked block layers is 3:3:9:3, and the number of input channels decreases successively to 128, 64, 32, and 16, respectively;

[0089] 3.2) Use the newly expanded training set from step 2) to train the fault diagnosis model. Update the weight parameters of the fault diagnosis model based on its performance on the training set. Use the classification accuracy of the fault diagnosis model on the test set as the evaluation metric. According to the fault diagnosis algorithm evaluation requirements in the "National Standard of the People's Republic of China (GB / T 43555-2023)" "Evaluation Method for Intelligent Service Predictive Maintenance Algorithms," the qualified accuracy criterion for machine learning algorithms is greater than 70%. The classification accuracy calculation formula is:

[0090]

[0091] Where: A represents the classification accuracy of the test set, P represents the number of correctly classified samples in the test set, and Q represents the total number of test set samples.

[0092] 3.3) In the model testing phase, the test set in 3.2) is substituted into the fault diagnosis model, and the classification accuracy of the fault diagnosis model on the test set is used as the evaluation indicator. When the classification accuracy is greater than or equal to 90%, the optimal fault diagnosis model weight parameters are saved.

[0093] 4) Fault diagnosis: According to step 1), the helical gear test data is collected and converted into RGB-GADF images as a test set. The fault diagnosis model is used to perform fault diagnosis on the test set data to complete the fault diagnosis of the helical gear.

[0094] The specific steps for fault diagnosis in step 4) are as follows:

[0095] 4.1) Using the operation in step 1), collect the vibration signal of the helical gear, convert the vibration signal into an RGB-GADF image and use it as the test set data;

[0096] 4.2) Use the fault diagnosis model trained with the new training set expanded in step 3), import the fault diagnosis weights, and input the test set data to complete the fault diagnosis of the helical gear.

[0097] Example 2

[0098] The present invention is further described in detail below with reference to specific examples. The contents described are all explanations of the present invention, but these descriptions cannot be understood as limiting the scope of the present invention. The scope of protection of the present invention is defined by the appended claims, and any changes based on the claims of the present invention fall within the scope of protection of the present invention.

[0099] See attached Figure 1 The helical gear expansion diagnosis method based on three-dimensional vibration fusion angular field images of the present invention mainly includes the following steps:

[0100] 1) A helical gear fault simulation test bench was used to conduct a helical gear fault simulation experiment. The motor was set up in five different states: normal, partially broken, pitting, worn, and cracked. The motor speed was set at 700 r / min and the load was 25 N·m. A three-axis vibration accelerometer was installed above the helical gearbox to collect vibration signals.

[0101] 1.1) The vibration signal of the helical gear is collected to obtain raw data. The raw data is divided into multiple groups of data with 2048 sampling points as a group without duplication. The PAA method with a local window of 2 is used to approximately aggregate each group of data into 1024 points. The GADF image corresponding to each direction is constructed separately. The GADF image conversion process is as follows:

[0102] ① For time series data X={x1,x2,…,x n} to normalize and transform the amplitude to [-1,1]. The calculation formula is as follows:

[0103]

[0104] Where: x i is the i-th value of the time series data; is the i-th value of the time series data after normalization; max(X) is the maximum value of the time series data; min(X) is the minimum value of the time series data.

[0105] ② Convert the normalized value into the cosine of the angle through calculation, use the timestamp as the radius, and convert the transformed time series data from the Cartesian coordinate system to the polar coordinate system. The calculation formulas for the polar angle and polar diameter in polar coordinates are as follows:

[0106]

[0107] Where: is the i-th value of the normalized time series data; for The corresponding polar angle; t i is the timestamp corresponding to the i-th value of the time series data; N is the constant factor for adjusting the span of the polar coordinate system; r i is the polar radius corresponding to the i-th value of the time series data;

[0108] ③ Use trigonometric function difference to represent the correlation between each time point of the GADF image. During the conversion, the GADF image is established by calculating the angle difference of the sine function. The GADF image calculation is as follows:

[0109]

[0110] Where GADF is the GADF image; sin is the sine function; φ1 is the polar angle corresponding to the first data point; φ n The polar angle corresponding to the nth data point.

[0111] 1.2) Normalize the GADF image pixels corresponding to the three directions to between 0 and 255, and map their colors to red, green, and blue, respectively. Input the image information corresponding to the three directions into the three channels of the RGB image to construct the RGB-GADF image;

[0112] 1.3) Scale the converted RGB-GADF image to 256×256 pixels;

[0113] The RGB-GADF image samples of various states generated after processing are shown in the attached figure. Figure 2 shown.

[0114] 1.4) All obtained RGB-GADF images are labeled according to different helical gear states. The preset training sets are 50 and 70 RGB-GADF images per category, respectively, and the test set is 50 RGB-GADF images per category.

[0115] 2) Construct a sample expansion model, which includes the following steps:

[0116] 2.1) Constructing a helical gear fault feature dataset through the operations in step 1), and constructing a sample expansion model using the helical gear fault feature dataset. The sample expansion model includes a generator and a discriminator. The generator is composed of multiple layers of deconvolution and incorporates batch normalization and an activation function. The input of the generator is random noise, and the output is a new image of the same size as the RGB-GADF image.

[0117] 2.2) The discriminator is a dual-branch deep feature fusion structure. The dual-branch discriminator is composed of multiple layers of convolution and adds batch normalization and activation functions. The deep features extracted by the dual-branch discriminator are shunted into adaptive average pooling and adaptive maximum pooling. The deep overall features and significant features extracted by the two are spliced ​​into a vector input into a multi-layer perceptron composed of several fully connected layers. The multi-layer perceptron finally outputs the result and performs gradient update through loss function feedback. The discriminator input is the new sample output by the generator and the training samples in the training set, and the output is the sample discrimination probability;

[0118] The network structure diagram of the generator and discriminator is shown in the attached Figure 3 As shown, the specific parameter settings of each layer of the model are shown in Table 1 and Table 2;

[0119] Table 1 Generator layer parameter settings

[0120]

[0121] Table 2 Parameter settings of each layer of the discriminator

[0122]

[0123] 2.3) Use FID (Frechet inception distance, FID), which can measure the similarity and diversity of generated samples, to evaluate the quality of sample expansion. Select and save the weight parameters of the optimal expansion model. Then call the weight parameters and set different expansion ratios to expand the sample. The new training set is composed of a mixture of generated samples and original training samples.

[0124] The generated samples are as follows Figure 4 To demonstrate the superiority of the method of the present invention, the method of the present invention is compared with two generative adversarial network models, DCGAN and WGAN-GP. The comparison results are shown in Table 3. In the table, dataset A represents 50 training samples of each category, and dataset B represents 70 training samples of each category.

[0125] Table 3 Minimum and average FID values ​​for different datasets and sample expansion models at 700 r / min and 25 N·m

[0126]

[0127]

[0128] 3) Construct a fault diagnosis model, which specifically includes the following steps:

[0129] 3.1) constructing a stacked block convolutional neural network fault diagnosis model with an inverted triangular channel distribution, wherein the stacked block convolutional neural network fault diagnosis model with an inverted triangular channel distribution includes a feature extractor and a classifier, wherein the feature extractor includes four stages of stacked block layers, wherein the number of stacking times of the four stages of stacked block layers is 3:3:9:3, and the number of input channels decreases successively to 128, 64, 32, and 16, respectively;

[0130] The structure of the stacked block convolutional neural network fault diagnosis model with inverted triangle channel distribution is shown in the attached figure. Figure 5 As shown, the comparison of the parameters of each layer of the fault diagnosis model of the present invention and the ConvNeXt-Tiny model is shown in Table 4, where s represents the convolution kernel size and dim represents the number of input channels.

[0131] Table 4 ConvNeXt Block layer parameters of ConvNeXt-Tiny and the four stages of the present invention

[0132]

[0133] 3.2) Use the newly expanded training set from step 2) to train the fault diagnosis model. Update the weight parameters of the fault diagnosis model based on its performance on the training set. Use the classification accuracy of the fault diagnosis model on the test set as the evaluation metric. According to the fault diagnosis algorithm evaluation requirements in the "National Standard of the People's Republic of China (GB / T 43555-2023)" "Evaluation Method for Intelligent Service Predictive Maintenance Algorithms," the qualified accuracy criterion for machine learning algorithms is greater than 70%. The classification accuracy calculation formula is:

[0134]

[0135] Where: A represents the classification accuracy of the test set, P represents the number of correctly classified samples in the test set, and Q represents the total number of test set samples.

[0136] 3.3) In the model testing phase, the test set in 3.2) is substituted into the fault diagnosis model, and the classification accuracy of the fault diagnosis model on the test set is used as the evaluation indicator. When the classification accuracy is greater than or equal to 90%, the optimal fault diagnosis model weight parameters are saved.

[0137] 4) Fault diagnosis, including the following steps:

[0138] 4.1) Using the operation in step 1), collect the vibration signal of the helical gear, convert the vibration signal into an RGB-GADF image and use it as the test set data;

[0139] 4.2) Use the fault diagnosis model trained with the new training set expanded in step 3), import the fault diagnosis weights, and input the test set data to complete the fault diagnosis of the helical gear.

[0140] Table 5 shows the diagnostic accuracy of the two training datasets before and after expansion at 700 r / min and a load of 25 N·m. When there are 50 images per class, the accuracy reaches 93.6% after expansion, and when there are 70 images per class, the accuracy reaches 98.4% after expansion. The corresponding diagnostic results before and after expansion for different datasets are shown in the attached figure. Figure 6 shown.

[0141] Table 5 Diagnostic accuracy of the two training datasets before and after expansion at 700 r / min and load 25 N·m

[0142]

[0143] To demonstrate the superiority of the method of the present invention, the present invention is compared with the ConvNeXt-Tiny model. The training set has 50 samples per category and the test set has 50 samples per category. The results are shown in Table 6. Figure 7 Classification of two fault diagnosis models.

[0144] Table 6 Training parameters of two fault diagnosis models

[0145]

[0146] Example 3

[0147] The helical gear expansion diagnosis system based on three-way vibration fusion angular field images includes:

[0148] Helical gear fault feature dataset construction module: It is used to use a three-dimensional acceleration sensor to collect the horizontal radial, vertical radial and axial vibration signals of the helical gear to obtain the three-dimensional vibration signals. The three-dimensional vibration signals are segmented and approximated according to the specified number of points and normalized before being converted into GADF images. The pixels of the GADF images corresponding to the three directions are then normalized to between 0 and 255, and their colors are mapped to R, G, and B respectively to establish an RGB-GADF image. The RGB-GADF image is downsampled to a preset size to obtain a new image. After manually labeling the new image, the helical gear fault feature dataset is established and divided into a training set and a test set.

[0149] Sample expansion model construction module: used to construct a sample expansion model using the helical gear fault feature dataset. The sample expansion model includes a generator and a discriminator. The generator is composed of multiple layers of deconvolution and adds batch normalization and activation function. The input is random noise and the output is a test image with the same size as the RGB-GADF image. The discriminator is a dual-branch deep feature fusion structure. The dual-branch discriminator is composed of multiple layers of convolution and adds batch normalization and activation function. The deep features extracted by the dual-branch discriminator are shunted to the adaptive average pooling and adaptive maximum pooling, and the The deep overall features and significant features extracted by the two are spliced ​​into a vector and input into a multi-layer perceptron composed of several fully connected layers. The multi-layer perceptron finally outputs the result and performs gradient update through loss function feedback. The discriminator inputs the new samples output by the generator and the training samples in the training set, and outputs the sample discrimination probability. The weight parameters of the optimal expansion model are selected and saved using the indicator for evaluating the similarity of the sample structure. The weight parameters are called to expand the sample. The new training set is composed of a mixture of the generated samples and the original training samples, namely the helical gear fault feature dataset.

[0150] A fault diagnosis model construction module is used to construct a stacked block convolutional neural network fault diagnosis model with an inverted triangle channel distribution. The stacked block convolutional neural network fault diagnosis model with an inverted triangle channel distribution includes a feature extractor and a classifier. The feature extractor includes four-stage stacked block layers. The stacked block convolutional neural network fault diagnosis model with an inverted triangle channel distribution is trained using a helical gear fault feature dataset. After training, the optimal fault diagnosis model weight parameters are saved to obtain a fault diagnosis model.

[0151] Fault diagnosis module: It is used to collect the helical gear test data according to the method of building a module of the helical gear fault feature dataset, and convert it into RGB-GADF images as a test set. The fault diagnosis model is used to perform fault diagnosis on the test set data to complete the fault diagnosis of the helical gear.

[0152] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0153] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0154] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0155] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0156] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that after reading the present invention, those skilled in the art may still make various changes, modifications or equivalent substitutions to the specific implementation methods of the invention, but these changes, modifications or equivalent substitutions are all within the scope of protection of the pending claims of the invention.

Claims

1. A helical gear expansion diagnosis method based on three-way vibration fusion angular field images is characterized by: The following steps are involved: 1) Constructing a helical gear fault feature dataset A three-dimensional accelerometer is used to collect the horizontal radial, vertical radial, and axial vibration signals of the helical gear to obtain a three-dimensional vibration signal. The three-dimensional vibration signal is segmented and approximated according to the specified number of points and normalized to convert it into a GADF image. The pixels of the GADF image corresponding to the three directions are then normalized to between 0 and 255, and their colors are mapped to R, G, and B respectively to create an RGB-GADF image. The RGB-GADF image is downsampled to a preset size to obtain a new image. The new image is manually labeled to establish a helical gear fault feature dataset and divided into a training set and a test set. 2) Build a sample expansion model A sample expansion model is constructed using a helical gear fault feature dataset. The sample expansion model includes a generator and a discriminator. The generator is composed of multiple layers of deconvolution and adds batch normalization and activation functions. The input is random noise, and the output is a test image of the same size as the RGB-GADF image. The discriminator is a dual-branch deep feature fusion structure. The dual-branch discriminator is composed of multiple layers of convolution and adds batch normalization and activation functions. The deep features extracted by the dual-branch discriminator are divided into adaptive average pooling and adaptive maximum pooling, and the deep overall features and significant features extracted by the two are spliced ​​into a vector and input into a multi-layer perceptron composed of several fully connected layers. Finally, the multi-layer perceptron outputs the result and performs gradient update through loss function feedback. The discriminator inputs the new sample output by the generator and the training sample in the training set, and outputs the sample discrimination probability. The weight parameter of the optimal expansion model is selected and saved using the indicator for evaluating the sample structure similarity. The weight parameter is called for sample expansion. The new training set is composed of a mixture of generated samples and original training samples, namely the helical gear fault feature dataset; 3) Build a fault diagnosis model A stacked block convolutional neural network fault diagnosis model with an inverted triangular channel distribution is constructed. The stacked block convolutional neural network fault diagnosis model with an inverted triangular channel distribution includes a feature extractor and a classifier. The feature extractor includes four-stage stacked block layers. The stacked block convolutional neural network fault diagnosis model with an inverted triangular channel distribution is trained using a helical gear fault feature dataset. After training, the optimal fault diagnosis model weight parameters are saved to obtain a fault diagnosis model. 4) Fault diagnosis According to step 1), the helical gear test data is collected and converted into RGB-GADF images as a test set. The fault diagnosis model is used to perform fault diagnosis on the test set data to complete the fault diagnosis of the helical gear.

2. The helical gear expansion diagnosis method based on three-dimensional vibration fusion angular field image according to claim 1 is characterized in that: In step 1), a helical gear fault feature dataset is constructed, specifically: 1.1) A three-axis accelerometer is used to collect the horizontal radial, vertical radial, and axial vibration signals of the helical gear. The three-axis vibration signals are subjected to PAA at a specified number of points, normalized, and then converted into a GADF image. PAA segments the vibration signal using a pre-set local window, calculates the average value of each segmented vibration signal, and reconstructs the time series data using the average value of all segments. The GADF image conversion process is as follows: ① For time series data X={x1,x2,…,x n } to normalize and transform the amplitude to [-1,1]. The calculation formula is as follows: Where: x i is the i-th value of the time series data; is the i-th value of the time series data after normalization; max(X) is the maximum value of the time series data; min(X) is the minimum value of the time series data; ② Convert the normalized value into the cosine of the angle through calculation, use the timestamp as the radius, and convert the transformed time series data from the Cartesian coordinate system to the polar coordinate system. The calculation formulas for the polar angle and polar diameter in polar coordinates are as follows: Where: is the i-th value of the normalized time series data; for The corresponding polar angle; t i is the timestamp corresponding to the i-th value of the time series data; N is the constant factor for adjusting the span of the polar coordinate system; r i is the polar radius corresponding to the i-th value of the time series data; ③ Use trigonometric function difference to represent the correlation between each time point of the GADF image. During the conversion, the GADF image is established by calculating the angle difference of the sine function. The GADF image calculation is as follows: Where GADF is the GADF image; sin is the sine function; φ1 is the polar angle corresponding to the first data point; φ n The polar angle corresponding to the nth data point; 1.2) Normalize the GADF image pixels corresponding to the three directions to between 0 and 255, and map their colors to R, G, and B respectively. Input the image information corresponding to the three directions into the three channels of the RGB image to construct the RGB-GADF image; 1.3) Scale the converted RGB-GADF image to 256×256 pixels; 1.4) All the obtained RGB-GADF images are labeled according to different helical gear states and divided into training and test sets according to preset requirements.

3. The helical gear expansion diagnosis method based on three-dimensional vibration fusion angular field image according to claim 2 is characterized in that: The designated number of points is set according to the sampling frequency and the rotational speed of the shaft where the helical gear is located, and each signal segment of the designated number of points contains at least 1 / 4 cycle information.

4. The helical gear expansion diagnosis method based on three-dimensional vibration fusion angular field image according to claim 2 is characterized in that: The local window size preset by PAA in step 1.1) is 2.

5. The helical gear expansion diagnosis method based on three-dimensional vibration fusion angular field image according to claim 2 is characterized in that: In step 1.4), the training set is preset to have x RGB-GADF images per class, and the test set is preset to have y RGB-GADF images per class, where x ≥ y.

6. The helical gear expansion diagnosis method based on three-dimensional vibration fusion angular field image according to claim 1 is characterized in that: The sample expansion model is constructed in step 2), specifically: 2.1) Using the helical gear fault feature dataset, a sample expansion model is constructed. The sample expansion model includes a generator and a discriminator. The generator is composed of multiple layers of deconvolution and incorporates batch normalization and activation functions. The input is random noise, and the output is a new image of the same size as the RGB-GADF image. 2.2) The discriminator is a two-branch deep feature fusion structure. The two-branch discriminator is composed of multiple layers of convolution and adds batch normalization and activation functions. The deep features extracted by the two-branch discriminator are shunted into adaptive average pooling and adaptive maximum pooling. The deep overall features and significant features extracted by the two are spliced ​​into a vector and input into a multi-layer perceptron composed of several fully connected layers. The multi-layer perceptron finally outputs the result and performs gradient update through loss function feedback. The discriminator input is the new sample output by the generator and the training samples in the training set, and the output is the sample discrimination probability; 2.3) Using the index for evaluating sample structure similarity, the weight parameters of the optimal expansion model are selected and saved. The weight parameters are used to expand the sample. The new training set is composed of a mixture of generated samples and original training samples.

7. The helical gear expansion diagnosis method based on three-dimensional vibration fusion angular field images according to claim 1 is characterized in that: The fault diagnosis model is constructed in step 3), specifically: 3.1) Constructing a stacked block convolutional neural network fault diagnosis model with an inverted triangle channel distribution, wherein the stacked block convolutional neural network fault diagnosis model with an inverted triangle channel distribution includes a feature extractor and a classifier, wherein the feature extractor includes four-stage stacked block layers; 3.2) Use the new training set expanded in step 2) to train the fault diagnosis model. Update the weight parameters of the fault diagnosis model based on its performance on the training set. Use the classification accuracy of the fault diagnosis model on the test set as the evaluation indicator. The classification accuracy calculation formula is: Where: A represents the classification accuracy of the test set, P represents the number of correctly classified samples in the test set, and Q represents the total number of test set samples; 3.3) In the model testing phase, the test set in 3.2) is substituted into the fault diagnosis model, and the classification accuracy of the fault diagnosis model on the test set is used as the evaluation indicator. When the classification accuracy is greater than or equal to 90%, the optimal fault diagnosis model weight parameters are saved.

8. The helical gear expansion diagnosis method based on three-dimensional vibration fusion angular field images according to claim 7 is characterized in that: The stacking ratio of the stacked block layers in the four stages is 3:3:9:3, and the number of input channels decreases successively to 128, 64, 32 and 16 respectively.

9. The helical gear expansion diagnosis method based on three-dimensional vibration fusion angular field images according to claim 1 is characterized in that: The fault diagnosis in step 4) is specifically as follows: 4.1) Using the operation in step 1), collect the vibration signal of the helical gear, convert the vibration signal into an RGB-GADF image and use it as the test set data; 4.2) Use the fault diagnosis model trained with the new training set expanded in step 3), import the fault diagnosis weights, and input the test set data to complete the fault diagnosis of the helical gear.

10. Helical gear expansion diagnosis system based on three-way vibration fusion angular field image, characterized by: include: Helical gear fault feature dataset construction module: It is used to use a three-dimensional acceleration sensor to collect the horizontal radial, vertical radial and axial vibration signals of the helical gear to obtain the three-dimensional vibration signals. The three-dimensional vibration signals are segmented and approximated according to the specified number of points and normalized before being converted into GADF images. The pixels of the GADF images corresponding to the three directions are then normalized to between 0 and 255, and their colors are mapped to R, G, and B respectively to establish an RGB-GADF image. The RGB-GADF image is downsampled to a preset size to obtain a new image. After manually labeling the new image, the helical gear fault feature dataset is established and divided into a training set and a test set. Sample expansion model construction module: used to construct a sample expansion model using the helical gear fault feature dataset. The sample expansion model includes a generator and a discriminator. The generator is composed of multiple layers of deconvolution and adds batch normalization and activation function. The input is random noise and the output is a test image with the same size as the RGB-GADF image. The discriminator is a dual-branch deep feature fusion structure. The dual-branch discriminator is composed of multiple layers of convolution and adds batch normalization and activation function. The deep features extracted by the dual-branch discriminator are shunted to the adaptive average pooling and adaptive maximum pooling, and the The deep overall features and significant features extracted by the two are spliced ​​into a vector and input into a multi-layer perceptron composed of several fully connected layers. The multi-layer perceptron finally outputs the result and performs gradient update through loss function feedback. The discriminator inputs the new samples output by the generator and the training samples in the training set, and outputs the sample discrimination probability. The weight parameters of the optimal expansion model are selected and saved using the indicator for evaluating the similarity of the sample structure. The weight parameters are called to expand the sample. The new training set is composed of a mixture of the generated samples and the original training samples, namely the helical gear fault feature dataset. A fault diagnosis model construction module is used to construct a stacked block convolutional neural network fault diagnosis model with an inverted triangle channel distribution. The stacked block convolutional neural network fault diagnosis model with an inverted triangle channel distribution includes a feature extractor and a classifier. The feature extractor includes four-stage stacked block layers. The stacked block convolutional neural network fault diagnosis model with an inverted triangle channel distribution is trained using a helical gear fault feature dataset. After training, the optimal fault diagnosis model weight parameters are saved to obtain a fault diagnosis model. Fault diagnosis module: It is used to collect the helical gear test data according to the method of building a module of the helical gear fault feature dataset, and convert it into RGB-GADF images as a test set. The fault diagnosis model is used to perform fault diagnosis on the test set data to complete the fault diagnosis of the helical gear.

Images