A belt conveyor roller fault diagnosis method and system based on multi-scale feature fusion and residual mask convolution attention algorithm

Abstract

The application discloses a kind of based on multi-scale feature fusion and residual mask convolution attention algorithm's belt conveyor roller fault diagnosis method and system, it is related to nondestructive testing field.The application includes: using 8 array element microphone linear array sound signal collector, using adaptive beam forming algorithm, sound data is collected, and the roller sound dataset under actual working condition is made, the sample in data set is carried out short-time Fourier transform to obtain two-dimensional time-frequency domain image;Two-dimensional time-frequency domain image is input into multi-scale feature fusion and residual mask convolution attention algorithm and is trained, to obtain the model after training;The image to be detected is input into the model after training, whether the roller is judged to fail.Wherein multi-scale feature fusion and residual mask convolution attention algorithm includes: multi-scale feature fusion module, residual mask convolution attention module and classifier.The application realizes belt conveyor roller fault detection, and improves the accuracy and reliability of fault detection.

Description

Technical Field

[0001] This invention relates to the fields of signal processing technology and non-destructive testing, and in particular to a method and system for diagnosing faults in belt conveyor idlers based on multi-scale feature fusion and residual mask convolutional attention algorithm. Background Technology

[0002] Belt conveyors are continuous transportation devices widely used in industrial and agricultural production, logistics, and other fields. As a key component supporting the conveyor belt and materials, the health of idlers has a significant impact on the safe and reliable operation of the belt conveyor. In practical applications, due to prolonged high-intensity operation and the impact and wear of materials, idlers frequently experience mechanical failures. These failures can lead to decreased belt conveyor performance, reduced production efficiency, significant economic losses, numerous safety hazards, and even serious safety accidents.

[0003] Currently, diagnostic methods for idler roller faults generally focus on vibration signals and thermal imaging signals, employing signal processing methods for fault analysis. However, traditional signal processing methods suffer from poor accuracy and robustness.

[0004] Sound signals, as an important non-contact detection method, have advantages such as convenient signal acquisition, no interference with the machine, and low cost, thus showing great application potential in fault diagnosis. However, under actual working conditions, when belt conveyors are running, ambient noise can easily drown out the sound of the idlers, making the sound characteristics of the idlers unclear and affecting fault detection performance.

[0005] Therefore, this invention addresses the problem of belt conveyor idler fault detection by proposing a method and system for belt conveyor idler fault diagnosis based on multi-scale feature fusion and residual mask convolutional attention algorithm, aiming to improve the accuracy and real-time performance of belt conveyor idler fault detection. Summary of the Invention

[0006] This invention discloses a fault diagnosis method and system for belt conveyor idlers based on multi-scale feature fusion and residual mask convolutional attention algorithm. It can reconstruct the spatial distribution information of the sound field around the sound source of the belt conveyor idler, and enhance the directionality and denoise the signal in the sound source, thereby improving the signal-to-noise ratio and accuracy of the idler sound signal. It can fully learn the complementary information between features of different scales, enhance the expression of input features, improve the modeling ability of local features and global context information, reduce the computational complexity, and improve the fitting performance of the model, thus providing great help for the fault detection of belt conveyor idlers.

[0007] The first aspect of this invention discloses a fault diagnosis method for belt conveyor idler rollers based on multi-scale feature fusion and residual mask convolutional attention algorithm, comprising: using an 8-element microphone linear array sound signal acquisition device and an adaptive beamforming algorithm to acquire sound data, creating an idler roller sound dataset under actual working conditions; performing short-time Fourier transform on samples in the idler roller sound dataset to obtain a two-dimensional time-frequency domain image I; inputting I into the multi-scale feature fusion and residual mask convolutional attention algorithm for training to obtain a trained model M, wherein the multi-scale feature fusion and residual mask convolutional attention algorithm includes a multi-scale feature fusion module, a residual mask convolutional attention module, and a classifier; inputting the image to be detected into M to obtain the inspection result of whether the belt conveyor idler roller has a fault. Specifically, using an 8-element microphone linear array sound signal acquisition device and an adaptive beamforming noise reduction algorithm to suppress interference signals and noise signals in non-target directions and enhance the sound signal in the target direction, thereby improving the accuracy of fault diagnosis, let the signal received by the microphone array be y = [y1, y2, ..., y n ] T When M at time k n The signal can be defined as

[0008] y n (k)=α n s1[kt-τ n ]+v n (k)=x n (k)+v n (k)

[0009] Among them, α = [α1, α2, ..., α n ] T Let s1[k] be the target signal, and s1[k] be the attenuation factor caused by propagation effects. The time alignment factor τ = [τ1, τ2, ..., τk] is the time-alignment factor. n ] T Is the target source to M n The time difference between M1 and M2, i.e., the relative time delay, v n (k) is M n The additive noise, assuming it is uncorrelated with the target signal, is denoted by filter h. Then the beamforming of the microphone array is as follows:

[0010]

[0011] Where y a，i (k) is the time alignment M n The goal of the filter is to suppress noise to the maximum extent while keeping the target signal x1[k] unaffected. Therefore, under the constraint that the target signal is unaffected, the filter coefficients are selected... To minimize the output power h, where For y a The correlation matrix of (k) yields the minimum filter weight coefficients, which can be expressed as:

[0012]

[0013] In the formula, α1 is the attenuation factor of the target signal x1(k), which is obtained using the Lagrange multiplier method.

[0014]

[0015] When α1 = 1, the sound field model is a near-field model, and the distance between the sound source and the microphone array is...

[0016] Therefore, the propagation path attenuation is related to distance, and can be obtained as follows:

[0017]

[0018] In the formula, L is the straight-line distance between the target sound source and the reference microphone, and d M Let τ be the spacing between the centers of the array elements. The time alignment factor τ is determined by the distance difference between each array element and the sound source. Let τ1 = 0, then the time delay factor for each array element is calculated as follows:

[0019]

[0020] The multi-scale feature fusion module works as follows: Convolution operations with kernel sizes of 3, 5, 7, and 9 are performed on the input signal X to obtain four feature maps: Feature_C0, Feature_C1, Feature_C2, and Feature_C3. Feature_C0, Feature_C1, Feature_C2, and Feature_C3 are concatenated to obtain a multi-scale feature, Feature_C4. A 1×1 convolution is used to reduce the dimensionality of Feature_C4 to obtain feature map F. Subtracting F from X yields y, which provides complementary information between features at different scales. The calculation formula is as follows:

[0021] y = XF

[0022] The residual masked convolutional attention module is specifically composed of a masked convolutional attention module (MCA) and an optimized ResNet-18 structure. In the optimized ResNet-18 structure, the block layers use the GELU activation function, and each block layer is stacked once. The block layer introduces the masked convolutional attention module (MCA) to enhance the global feature capabilities. The masked convolutional attention module (MCA) consists of a ReLU-activated masked convolutional layer and a channel attention module (SE), and performs predictions within a supervised learning framework. The masked convolution uses four convolutional sub-kernels.

[0023] The second aspect of the present invention discloses a belt conveyor idler fault diagnosis system based on multi-scale feature fusion and residual mask convolutional attention algorithm, comprising: an 8-element microphone linear array sound signal acquisition device for acquiring sound data; a memory for storing program instructions; and a processor for calling the program instructions stored in the memory to implement the belt conveyor idler fault diagnosis method based on multi-scale feature fusion and residual mask convolutional attention algorithm as described in any of the preceding claims.

[0024] The beneficial effects of this invention include at least the following: reconstructing the spatial distribution information of the sound field around the sound source of the belt conveyor idler, enhancing the directionality of the signal in the sound source and denoising it, improving the signal-to-noise ratio and accuracy of the idler sound signal, fully learning the complementary information between features at different scales, enhancing the expression of input features, improving the modeling ability of local features and global context information, while reducing computational complexity and improving the fitting performance of the model. This method can be applied, but is not limited to, publicly available idler or typical bearing sound datasets. Attached Figure Description

[0025] Figure 1 A flowchart illustrating a fault diagnosis method for belt conveyor idler rollers based on multi-scale feature fusion and residual mask convolutional attention algorithm according to an embodiment of the present invention is shown.

[0026] Figure 2 A flowchart illustrating a multi-scale feature fusion and residual mask convolutional attention algorithm according to an embodiment of the present invention is shown.

[0027] Figure 3 A schematic diagram of the spatial noise interference of the target idler roller sound signal is shown according to an embodiment of the present invention.

[0028] Figure 4 A model structure diagram of a belt conveyor idler roller fault diagnosis method based on multi-scale feature fusion and residual mask convolutional attention algorithm according to an embodiment of the present invention is shown.

[0029] Figure 5 A model structure diagram of a mask convolution module according to an embodiment of the present invention is shown.

[0030] Figure 6 A comparison diagram of time-frequency domain results before and after adaptive beamforming noise reduction according to an embodiment of the present invention is shown.

[0031] Figure 7The diagram shows a comparison of the confusion matrix results before and after denoising of the idler roller noise dataset, based on a multi-scale feature fusion and residual mask convolutional attention algorithm according to an embodiment of the present invention. Detailed Implementation

[0032] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] While exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0034] This invention applies adaptive beamforming algorithm and image classification algorithm based on convolutional neural network to fault detection, and proposes a fault diagnosis method for belt conveyor idler rollers based on multi-scale feature fusion and residual mask convolutional attention algorithm, thereby achieving a more accurate and intelligent fault detection scheme.

[0035] Reference Figure 1 , Figure 2 As shown, the present invention discloses a method for fault diagnosis of belt conveyor idler rollers based on multi-scale feature fusion and residual mask convolutional attention algorithm, comprising the following steps:

[0036] S1. Using an 8-element microphone linear array sound signal acquisition device, an adaptive beamforming algorithm is used to acquire sound data and create a roller sound dataset under actual working conditions. Short-time Fourier transform is performed on the samples in the roller sound dataset to obtain a two-dimensional time-frequency domain image I.

[0037] S2. Input I into the multi-scale feature fusion and residual mask convolutional attention algorithm for training to obtain the trained model M. The multi-scale feature fusion and residual mask convolutional attention algorithm includes a multi-scale feature fusion module, a residual mask convolutional attention module, and a classifier.

[0038] S3. Input the image to be inspected into M to obtain the inspection results of whether the belt conveyor rollers are faulty.

[0039] In step S1 above, an adaptive beamforming algorithm is used; in this embodiment of the invention, see [link to relevant documentation]. Figure 3As shown, the spatial noise interference encountered by the microphone linear array sound acquisition device in collecting the sound signal of the target idler roller under actual working conditions is simulated. The adaptive beamforming algorithm is used to reconstruct the spatial distribution information of the sound field around the sound source of the belt conveyor idler roller, and the signal in the sound source is enhanced directionally and denoised to improve the signal-to-noise ratio of the sound signal.

[0040] In step S2 above, the image I obtained after the short-time Fourier transform is input into a multi-scale feature fusion and residual mask convolutional attention algorithm for training, resulting in the trained model M; in this embodiment of the invention, see... Figure 4 As shown, the multi-scale feature fusion and residual mask convolutional attention algorithm includes: 1) a multi-scale feature fusion module; 2) a residual mask convolutional attention module; and 3) a classifier.

[0041] ① Multi-scale feature fusion module

[0042] The main purpose of the multi-scale feature fusion module is to capture fault features at different scales using convolutions at four different scales, and to learn complementary information between features at different scales by using a feature fusion method of residual subtraction, thereby enhancing the expressive power of the input features.

[0043] See Figure 4 As shown in (b), the specific process of the multi-scale feature fusion module is as follows:

[0044] S1. Perform convolution operations on the input signal X with kernel sizes of 3, 5, 7 and 9 respectively to obtain four feature maps: Feature_C0, Feature_C1, Feature_C2 and Feature_C3.

[0045] S2. Concatenate Feature_C0, Feature_C1, Feature_C2, and Feature_C3 to obtain the multi-scale feature Feature_C4. Use a 1×1 convolution to reduce the dimensionality of Feature_C4 to obtain the feature map F.

[0046] S3. Subtract F from X to obtain y, thus acquiring complementary information between features at different scales. The calculation formula is:

[0047] y = XF

[0048] ② Residual Mask Convolutional Attention Module

[0049] In previous fault detection tasks, CNN-based methods can extract highly localized features through kernels to achieve a certain fault identification accuracy. However, they have weak global contextual modeling capabilities, making it difficult to fuse feature information from different distances and establish high-level semantic relationships. In this embodiment of the invention, the residual mask convolutional attention module is composed of a mask convolutional attention module (MCA) and an optimized ResNet-18 structure, which improves the modeling capabilities of local features and global contextual information.

[0050] See Figure 4 As shown in (c), the optimized ResNet-18 structure specifically refers to:

[0051] B i This is an optimized ResNet-18 architecture block layer. The ReLU activation function in the original ResNet-8 block layer is replaced with GELU to increase the model's fitting ability and improve fault diagnosis accuracy. Each block layer is stacked once to reduce the number of model parameters and improve running speed. Finally, a masked convolutional attention module (MCA) is introduced into the block layer to enhance global feature capabilities. i For an optimized ResNet-18 architecture, a Block layer, the masked convolution operation in the masked attention module enables the model to learn the intrinsic representation of the corresponding position, enhances the modeling of global content, and the masking mechanism improves the selective attention to unmasked salient features, enhancing the discriminativeness of features and the generalization of feature representation.

[0052] See Figure 4 As shown in (d), the Masked Convolutional Attention Module (MCA) specifically refers to:

[0053] The masked convolutional attention module consists of a ReLU-activated masked convolutional layer and a channel attention module (SE). It makes predictions within a supervised learning framework, enabling the model to learn the intrinsic representation of the corresponding position, enhancing the modeling of global content. The masking mechanism improves the selective attention to unmasked salient features, enhancing the discriminativeness of features and the generalization of feature representation.

[0054] See Figure 5 As shown, the masked convolutional layer specifically refers to:

[0055] The masked convolutional layer consists of four convolutional kernels of sizes K1×K1, K2×K2, K3×K3, and K4×K4, where d is the masking distance between the receptive field center M and the convolutional kernels. Let X∈R h×w×c The input features of the mask convolutional layer are given by c, where c represents the number of channels, and h and w represent the height and width, respectively. If K i , If both d and d are 1, then the masked convolution kernel K equals 5, where K idenoted as the size of the receptive field space of the convolutional kernel.

[0056] Furthermore, the neural network training parameters were set as follows: batch size 32, number of iterations 1200, initial learning rate 0.01, cross-entropy loss function and stochastic gradient descent (SGD) algorithm as the network optimizer, learning rate with fixed epoch decay, halving every 300 epochs, momentum coefficient 0.9, and weight decay coefficient 5e-4.

[0057] Furthermore, fault sound signals with typical bearing defects were selected, and the time-frequency domain comparison before and after adaptive beamforming is shown in the figure. Figure 6 As shown, Figure 6 The left side is the time-frequency diagram before noise reduction. Figure 6 The right side is the time-frequency graph after adaptive beamforming noise reduction, for comparison. Figure 6 As can be seen from the left and right parts, the sound signal before beamforming contains a lot of strong noise. The signal after adaptive beamforming greatly reduces the strong noise, and the sound signal that was submerged by noise becomes visible. The sound characteristics of the faulty idler are more obvious, which is more conducive to fault detection and classification.

[0058] Furthermore, in this embodiment of the invention, to intuitively explain the various fault diagnosis results, a confusion matrix is ​​plotted, with the horizontal axis representing the classification result and the vertical axis representing the actual fault category, such as... Figure 7 The left side shows the result without noise reduction. Figure 7 The right side shows the denoised result. The dataset consists of 5400 time-frequency images, including 17 typical idler roller faults and one intact idler roller as a reference. The dataset is divided into training and test sets in a 7:3 ratio. (Comparison) Figure 7 As can be seen from the left and right parts, the method of the present invention realizes the fault detection of belt conveyor idler rollers, and improves the accuracy and reliability of fault detection.

[0059] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for fault diagnosis of belt conveyor idler rollers based on multi-scale feature fusion and residual mask convolutional attention algorithm, characterized in that, Includes the following steps: S1. Using an 8-element microphone linear array sound signal acquisition device, an adaptive beamforming algorithm is used to acquire sound data and create a roller sound dataset under actual working conditions. Short-time Fourier transform is performed on the samples in the roller sound dataset to obtain a two-dimensional time-frequency domain image I. S2. Input I into the multi-scale feature fusion and residual mask convolutional attention algorithm for training to obtain the trained model M. The multi-scale feature fusion and residual mask convolutional attention algorithm includes a multi-scale feature fusion module, a residual mask convolutional attention module, and a classifier. The multi-scale feature fusion module specifically includes the following steps: a. Perform convolution operations with kernel sizes of 3, 5, 7 and 9 on the input signal X to obtain four feature maps: Feature_C0, Feature_C1, Feature_C2 and Feature_C3. b. Concatenate Feature_C0, Feature_C1, Feature_C2, and Feature_C3 to obtain the multi-scale feature Feature_C4. Use a 1×1 convolution to reduce the dimensionality of Feature_C4 to obtain the feature map F. c. Subtracting X from F yields y, which provides complementary information between features at different scales. The formula is y = XF. The residual masked convolutional attention module consists of a masked convolutional attention module (MCA) and an optimized ResNet-18 structure. In the optimized ResNet-18 structure, the Block layer uses the GELU activation function, and each Block layer is stacked once. The Block layer introduces the masked convolutional attention module (MCA) to enhance the global feature capability. The Masked Convolutional Attention Module (MCA) consists of a ReLU-activated masked convolutional layer and a channel attention module (SE). It performs predictions within a supervised learning framework, and the masked convolution uses four convolutional sub-kernels. S3. Input the image to be inspected into M to obtain the inspection results of whether the belt conveyor rollers are faulty.

2. A fault diagnosis system for belt conveyor idlers based on multi-scale feature fusion and residual mask convolutional attention algorithm, characterized in that, include: An 8-element microphone linear array sound signal acquisition device is used to acquire sound data; Memory, used to store program instructions; The processor is used to call program instructions stored in the memory to implement the belt conveyor idler roller fault diagnosis method based on multi-scale feature fusion and residual mask convolutional attention algorithm as described in claim 1.

Images