Gearbox fault diagnosis method based on multi-channel feature fusion

Abstract

The invention discloses a gearbox fault diagnosis method based on multi-channel feature fusion. The gearbox fault diagnosis method comprises the steps that vibration signals of a gearbox to be diagnosed are collected through sensors installed at different positions in a multi-channel mode; for each vibration signal, obtaining a two-dimensional time-frequency image of the vibration signal by using a sliding window function, and respectively extracting features of the vibration signal and the two-dimensional time-frequency image by using a multi-layer gated convolutional network to obtain a one-dimensional feature and a two-dimensional feature of the vibration signal; fusing the one-dimensional features and the two-dimensional features of the vibration signals collected by the channels by using a convolution attention module to obtain fusion features of the gearbox to be diagnosed; and based on the fusion features of the to-be-diagnosed gearbox, performing fault diagnosis on the to-be-diagnosed gearbox by using a gearbox fault diagnosis model. According to the method, the comprehensiveness of feature expression is improved, and the channel correlation and the spatial correlation of the features are better captured, so that the modeling capability of the model for complex features is improved.

Description

Technical Field

[0001] The present application belongs to the technical field of gearbox fault diagnosis, and in particular relates to a gearbox fault diagnosis method based on multi-channel feature fusion. Background Art

[0002] A gearbox, a complex system consisting of gears, bearings, and drive shafts, plays a critical role in motion conversion and power transmission. In modern industry, the proper functioning of a gearbox is crucial for ensuring safe and continuous operation of machinery. When gearboxes operate at high speeds, under heavy loads, or in other challenging operating conditions for extended periods, they can fail, resulting in significant economic losses and even loss of life. Therefore, developing intelligent and accurate fault diagnosis methods to detect potential gearbox failures early is crucial.

[0003] The gearbox fault diagnosis process can be summarized as: analyzing the gearbox vibration signal, extracting fault features, performing feature analysis, and finally determining the fault type. Methods for analyzing and processing vibration signals have been extensively studied. For example, methods such as analyzing peak value, mean, variance, root mean square value, kurtosis, margin, wavelet transform, short-time Fourier transform, local mean decomposition, empirical mode decomposition, and variational mode decomposition can be used to reduce noise, extract features, and select features from vibration signals. However, these methods rely on manual extraction and screening of fault features, requiring prior knowledge and significant time and effort. With the continuous advancement of technology, deep learning has attracted considerable attention due to its powerful adaptive feature extraction and fault classification capabilities. Deep learning can autonomously learn and extract deep and rich feature information from large amounts of raw data, eliminating the need for manual feature extraction. This allows for end-to-end fault diagnosis and promotes the further development of intelligent gearbox fault diagnosis technology.

[0004] In the existing technology, only the raw vibration signal collected by a single sensor is usually processed and used for fault diagnosis. However, as a gearbox is a complex rotating machine, the vibration signal collected by a single sensor can usually only provide local and one-sided information. This can easily lead to incorrect decisions in the subsequent feature extraction and selection process, thereby affecting the fault diagnosis results. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a gearbox fault diagnosis method based on multi-channel feature fusion. By extracting one-dimensional and two-dimensional features from the vibration signals collected from multiple channels respectively, the comprehensiveness of feature expression can be improved. By fusing the features through the convolutional attention module, the channel correlation and spatial correlation of the features can be better captured, thereby improving the model's modeling ability for complex features.

[0006] This application provides a gearbox fault diagnosis method based on multi-channel feature fusion, including:

[0007] The vibration signals of the gearbox to be diagnosed are collected through multiple channels of sensors installed at different positions;

[0008] For each vibration signal, a sliding window function is used to obtain a two-dimensional time-frequency image of the vibration signal, and a multi-layer gated convolutional network is used to extract the features of the vibration signal and the two-dimensional time-frequency image respectively to obtain the one-dimensional and two-dimensional features of the vibration signal;

[0009] The one-dimensional features and two-dimensional features of the vibration signals collected by each channel are fused using a convolutional attention module to obtain the fused features of the gearbox to be diagnosed;

[0010] Based on the fusion characteristics of the gearbox to be diagnosed, a gearbox fault diagnosis model is used to perform fault diagnosis on the gearbox to be diagnosed.

[0011] Furthermore, the method of obtaining a two-dimensional time-frequency image of the vibration signal using a sliding window function includes:

[0012] The vibration signal is divided by a sliding window function to obtain vibration signal segments corresponding to different time periods; wherein the window function is a Hamming window or a Hanning window;

[0013] Perform Fourier segmentation on each vibration signal segment to obtain the spectrum corresponding to the time period;

[0014] The frequency spectra of different time periods are combined to obtain a two-dimensional time-frequency image of the vibration signal.

[0015] Furthermore, the extracting features of the vibration signal and the two-dimensional time-frequency image respectively using a multi-layer gated convolutional network includes:

[0016] Performing a common convolution operation on the vibration signal to obtain a common convolution output, which serves as a one-dimensional feature of the vibration signal;

[0017] For this two-dimensional time-frequency image, a gated weight matrix is used to control the importance of each element in the image, and the importance of each element is mapped to a range of 0-1 through an activation function;

[0018] The output of the ordinary convolution is multiplied by the mapped gated weight matrix one by one to obtain the two-dimensional features of the vibration signal.

[0019] Furthermore, the convolutional attention module includes a channel attention mechanism and a spatial attention mechanism;

[0020] The convolutional attention module is used to fuse the one-dimensional features and two-dimensional features of the vibration signals collected by each channel, including:

[0021] For each vibration signal, the spatial attention mechanism is used to fuse the one-dimensional and two-dimensional features of the vibration signal to obtain the fused features of the vibration signal;

[0022] The channel attention mechanism is used to fuse the fusion features of the vibration signals collected by each channel to obtain the fusion features of the gearbox to be diagnosed.

[0023] Furthermore, the gearbox fault diagnosis model is obtained by the following method:

[0024] Obtain vibration signals of the test gearbox under various fault types collected through multi-channel acquisition;

[0025] The vibration signals collected from multiple channels at the same time are grouped together to obtain multiple sets of fusion features of the experimental gearbox and the corresponding fault types;

[0026] The capsule network is trained using multiple sets of fused features of the test gearbox and corresponding fault types to obtain the gearbox fault diagnosis model.

[0027] Furthermore, the fault types include: bearing fault and gear fault; the bearing fault includes: rolling element fault, inner ring fault, outer ring fault and composite fault; the gear fault includes: gear notch, gear tooth breakage, tooth root crack and tooth surface wear.

[0028] The gearbox fault diagnosis method based on multi-channel feature fusion provided in this application can improve the comprehensiveness of feature expression by extracting one-dimensional and two-dimensional features from vibration signals collected through multiple channels respectively. By fusing the features through the convolutional attention module, the channel correlation and spatial correlation of the features can be better captured, thereby improving the model's ability to model complex features. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A flow chart of a gearbox fault diagnosis method based on multi-channel feature fusion provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solution and advantages of this technical solution more clear, the following technical solution is further described in detail in conjunction with specific implementation methods. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of this technical solution.

[0031] Please refer to Figure 1 The flowchart of the gearbox fault diagnosis method based on multi-channel feature fusion is shown in FIG. Figure 1As shown, the method includes:

[0032] S101. Collect vibration signals of the gearbox to be diagnosed through multiple channels of sensors installed at different positions.

[0033] In this step, the vibration signals of the gearbox to be diagnosed collected through multiple channels include: motor vibration signal, vibration signals of the planetary gearbox in three directions (X, Y, Z), motor torque and vibration signals of the parallel gearbox in three directions (X, Y, Z).

[0034] S102. For each vibration signal, a sliding window function is used to obtain a two-dimensional time-frequency image of the vibration signal, and features of the vibration signal and the two-dimensional time-frequency image are respectively extracted using a multi-layer gated convolutional network to obtain one-dimensional features and two-dimensional features of the vibration signal.

[0035] In specific implementation, the two-dimensional time-frequency image of the vibration signal can be obtained by using a sliding window function in the following manner:

[0036] Step 1021: Divide the vibration signal by a sliding window function to obtain vibration signal segments corresponding to different time periods.

[0037] The window function is a Hamming window or a Hanning window.

[0038] Step 1022: Perform Fourier transformation on each vibration signal segment to obtain a frequency spectrum corresponding to the time period.

[0039] Step 1023: Combine the frequency spectra of different time periods to obtain a two-dimensional time-frequency image of the vibration signal.

[0040] The method of converting the vibration signal into a two-dimensional time-frequency image shown in steps 1021-1023 can be expressed as formula (1):

[0041]

[0042] Where x(t) is the vibration signal; w(t-τ) is the sliding window function, τ is the center position; e -j2πft is the Fourier transform kernel function, X(τ,f) is the result of short-time Fourier transform, which represents the spectrum value at time τ and frequency f.

[0043] Furthermore, a multi-layer gated convolutional network is used to extract features of the vibration signal and the two-dimensional time-frequency image, including:

[0044] Step 1024: Perform a normal convolution operation on the vibration signal to obtain a normal convolution output, which serves as a one-dimensional feature of the vibration signal.

[0045] Step 1025: For the two-dimensional time-frequency image, a gated weight matrix is used to control the importance of each element in the image, and the importance of each element is mapped to a value between 0 and 1 through an activation function.

[0046] Step 1026: Multiply the output of the ordinary convolution by the mapped gated weight matrix one by one to obtain the two-dimensional features of the vibration signal.

[0047] The feature extraction method of the multi-layer gated convolutional network shown in steps 1024-1026 above can be expressed as formula (2):

[0048] Y=(W x *X)⊙σ(W g *X); (2)

[0049] Where X is the input of the multi-layer gated convolutional network, Y is the output of the multi-layer gated convolutional network, ⊙ represents element-by-element multiplication, * represents the ordinary convolution operation, and W x 、W g They represent the convolution kernel for extracting features and the convolution kernel for generating the gated weight matrix, and σ is the Sigmoid activation function.

[0050] S103. Using a convolutional attention module, the one-dimensional features and the two-dimensional features of the vibration signals collected by each channel are fused to obtain fused features of the gearbox to be diagnosed.

[0051] The convolutional attention module includes a channel attention mechanism and a spatial attention mechanism;

[0052] In specific implementation, the one-dimensional features and two-dimensional features of the vibration signals collected by each channel can be fused in the following ways:

[0053] Step 1031: For each vibration signal, use the spatial attention mechanism to fuse the one-dimensional features and the two-dimensional features of the vibration signal to obtain the fused features of the vibration signal.

[0054] In this step, the spatial attention mechanism focuses on the importance of the feature map in the spatial dimension. Different weights are assigned according to the feature maps at different positions, thereby adjusting the importance of the feature map in the spatial position. Its operating formula is as follows (3):

[0055] M s =σ(f([AvgPool(F);MaxPool(F)])); (3)

[0056] Where F represents the feature map, AvgPool represents the average pooling operation, MaxPool represents the maximum pooling operation, and f represents the convolution operation.

[0057] Step 1032: Utilize the channel attention mechanism to fuse the fusion features of the vibration signals collected by each channel to obtain the fusion features of the gearbox to be diagnosed.

[0058] In this step, the channel attention mechanism focuses on the influence of the feature map in the channel dimension. By weighting each feature channel, it adjusts the contribution of different channels to improve the feature expression ability. Its operation formula is as follows (4):

[0059] M c =σ(MLP(AvgPool(F))+MLP(MaxPool(F))); (4)

[0060] Where MLP is a multi-layer perceptron and σ is a Sigmoid activation function.

[0061] S104 : Based on the fusion features of the gearbox to be diagnosed, perform fault diagnosis on the gearbox to be diagnosed using a gearbox fault diagnosis model.

[0062] Among them, the fault types include: bearing fault and gear fault; the bearing fault includes: rolling element fault, inner ring fault, outer ring fault and combined fault; the gear fault includes: gear gap, gear tooth breakage, tooth root crack and tooth surface wear.

[0063] In specific implementation, the gearbox fault diagnosis model can be obtained by the following methods:

[0064] Step 201: Obtain vibration signals of a test gearbox under various fault types collected through multiple channels.

[0065] In this step, the publicly available gearbox dataset from Southeast University in China can be directly accessed to obtain multi-channel vibration signals of the gearbox under various fault types. This dataset was collected from a drivetrain dynamic simulator (DDS). The test bench consists of a motor, motor controller, planetary gearbox, parallel gearbox, brake, and brake controller. Specifically, the dataset contains two sub-datasets: a bearing dataset and a gearbox dataset. Each sub-dataset contains two operating conditions: 20Hz speed, 0V load, and 30Hz speed, 2V load. Bearings have four fault types and five states: rolling element fault, inner race fault, outer race fault, combined fault, and healthy state. Gears also have four fault types and five states: gear notch, gear tooth breakage, tooth root crack, tooth surface wear, and healthy state.

[0066] Step 202: The vibration signals collected from multiple channels at the same time are grouped together to obtain multiple sets of fusion features of the experimental gearbox and corresponding fault types.

[0067] In this step, multiple groups of fusion features are obtained using the above-mentioned methods of S102-S103, and the fault type corresponding to each group of fusion features is marked.

[0068] Step 203: Use the multiple sets of fusion features of the test gearbox and the corresponding fault types to train a capsule network to obtain the gearbox fault diagnosis model.

[0069] In this step, the capsule network consists of a convolutional layer, a main capsule layer, and a digital capsule layer. Specifically, the convolutional layer extracts the fused features again to prepare for the subsequent capsule layer; the main capsule layer converts the "scalar" in the traditional CNN into a "vector". In addition to representing features, vectors can also represent directions and states. Using vectors as the input and output of the network can reduce the loss of feature information and mine deeper feature information in the data, thereby improving the model's feature extraction capabilities; the digital capsule layer uses a dynamic routing algorithm to aggregate the information of the main capsule into the corresponding classification capsule. The dynamic routing algorithm is an iterative process that continuously adjusts the parameters of the main capsule layer and the digital capsule layer to improve the fault diagnosis performance. By continuously iterating and optimizing the connection weights, it ensures that the capsule network can effectively capture the hierarchical relationship between features. The embodiment of the present application uses multiple sets of fused features and corresponding fault types of the test gearbox to train the capsule network and ultimately obtain a gearbox fault diagnosis model.

[0070] The above content is only a preferred embodiment of the present invention. For ordinary technicians in this field, many changes can be made in the specific implementation methods and application scopes based on the ideas of the present technical content. As long as these changes do not deviate from the concept of the present invention, they all fall within the scope of protection of the present invention.

Claims

1. A gearbox fault diagnosis method based on multi-channel feature fusion, characterized in that: The method comprises: The vibration signals of the gearbox to be diagnosed are collected through multiple channels of sensors installed at different positions; For each vibration signal, a sliding window function is used to obtain a two-dimensional time-frequency image of the vibration signal, and a multi-layer gated convolutional network is used to extract the features of the vibration signal and the two-dimensional time-frequency image respectively to obtain the one-dimensional and two-dimensional features of the vibration signal; The one-dimensional features and two-dimensional features of the vibration signals collected by each channel are fused using a convolutional attention module to obtain the fused features of the gearbox to be diagnosed; Based on the fusion characteristics of the gearbox to be diagnosed, a gearbox fault diagnosis model is used to perform fault diagnosis on the gearbox to be diagnosed.

2. The method according to claim 1, wherein The method of obtaining a two-dimensional time-frequency image of the vibration signal by using a sliding window function includes: The vibration signal is divided by a sliding window function to obtain vibration signal segments corresponding to different time periods; wherein the window function is a Hamming window or a Hanning window; Perform Fourier segmentation on each vibration signal segment to obtain the spectrum corresponding to the time period; The frequency spectra of different time periods are combined to obtain a two-dimensional time-frequency image of the vibration signal.

3. The method according to claim 1, wherein The method of extracting features of the vibration signal and the two-dimensional time-frequency image respectively by using a multi-layer gated convolutional network includes: Performing a common convolution operation on the vibration signal to obtain a common convolution output, which serves as a one-dimensional feature of the vibration signal; For this two-dimensional time-frequency image, a gated weight matrix is used to control the importance of each element in the image, and an activation function is used to map the importance of each element to a value between 0 and 1. The output of the ordinary convolution is multiplied by the mapped gated weight matrix one by one to obtain the two-dimensional features of the vibration signal.

4. The method according to claim 1, wherein The convolutional attention module includes a channel attention mechanism and a spatial attention mechanism; The convolutional attention module is used to fuse the one-dimensional features and two-dimensional features of the vibration signals collected by each channel, including: For each vibration signal, the spatial attention mechanism is used to fuse the one-dimensional and two-dimensional features of the vibration signal to obtain the fused features of the vibration signal; The channel attention mechanism is used to fuse the fusion features of the vibration signals collected by each channel to obtain the fusion features of the gearbox to be diagnosed.

5. The method according to claim 1, wherein The gearbox fault diagnosis model is obtained by the following method: Obtain vibration signals of the test gearbox under various fault types collected through multi-channel acquisition; The vibration signals collected from multiple channels at the same time are grouped together to obtain multiple sets of fusion features of the experimental gearbox and the corresponding fault types; The capsule network is trained using multiple sets of fused features of the test gearbox and corresponding fault types to obtain the gearbox fault diagnosis model.

6. The method according to claim 1, wherein The fault types include: Bearing failure and gear failure; the bearing failure includes: rolling element failure, inner ring failure, outer ring failure and combined failure; the gear failure includes: gear notch, gear tooth breakage, tooth root crack and tooth surface wear.

Images