Belt conveyor roller fault sound detection method and system

Abstract

The present invention provides a method and system for detecting belt conveyor roller faults using sound, relating to the field of nondestructive fault detection. The method comprises: obtaining a time-frequency domain feature map X of a sound signal; extracting features from X using multiple convolutional layers of different sizes to obtain a multi-scale feature map #imgabs0#; sequentially passing #imgabs1# through a first Ghost module, a deep convolutional layer, a spatial-channel collaborative attention module, and a second Ghost module to perform deeper fault feature mapping in spatial and channel dimensions to obtain a high-level semantic feature map #imgabs2#; inputting #imgabs3# into a classifier module to obtain a classification result, and completing model training using a cross-entropy loss function and a reparameterized training method to obtain a fault detection model; and performing a short-time Fourier transform on the sound signal to be detected, inputting the resulting signal into the trained fault detection model to obtain a fault diagnosis result. The present invention improves the accuracy and real-time performance of belt conveyor roller fault detection.

Description

Technical Field

[0001] The present invention relates to the technical field of non-destructive fault detection, and in particular to a method and system for detecting fault sound of a belt conveyor roller. Background Art

[0002] Belt conveyors are widely used in industrial and agricultural production, logistics, and other fields due to their advantages, such as long transport distances, large transport volumes, low freight costs, and stable operation. Rollers are a crucial component of belt conveyors, playing a crucial role in supporting the conveyor belt and materials. However, due to harsh operating environments, rollers often experience damage and failure. Failure to promptly and effectively detect and diagnose these potential failures can lead to degraded equipment performance and reduced production efficiency, at best, or even serious consequences such as economic losses and casualties. Currently, methods for detecting belt conveyor roller failures generally focus on vibration signals and thermal imaging signals, employing signal processing methods for fault analysis. Traditional signal processing methods suffer from poor accuracy and robustness. Acoustic signals, as an important non-contact detection method, offer advantages such as convenient signal acquisition, no interference with the machine, and low cost, thus holding great potential for application in fault diagnosis. However, in real-world operating conditions, ambient noise can easily drown out the operating sound of the rollers, making the bearing sound signature less distinct and compromising fault detection performance. Therefore, improving the accuracy and real-time performance of belt conveyor roller fault detection has become a pressing technical challenge. Summary of the Invention

[0003] The present invention aims to solve at least one of the technical problems existing in the prior art or related art, and discloses a belt conveyor roller fault sound detection method and system, which improves the accuracy and real-time performance of belt conveyor roller fault detection.

[0004] The first aspect of the present invention discloses a method for detecting the sound of a belt conveyor roller fault, comprising: signal acquisition: collecting the sound signal of the belt conveyor roller during operation; time-frequency domain transformation: performing a short-time Fourier transform on the sound signal to obtain a time-frequency domain feature map X of the sound signal; multi-scale feature extraction: extracting features from the time-frequency domain feature map X through multiple convolutional layers of different sizes to obtain feature maps of multiple scales. , the feature map Splicing is performed on the channel dimension to obtain a multi-scale feature map ; High-level semantic extraction: multi-scale feature maps The first Ghost module, deep convolution layer, spatial channel collaborative attention module and the second Ghost module are used to perform deeper fault feature mapping in spatial and channel dimensions to obtain a high-level semantic feature map. ; Feature classification: high-level semantic feature maps Input it into the classifier module to obtain the classification result, and use the cross-entropy loss function and reparameterization training method to complete the model training to obtain the fault detection model; Fault detection: After performing short-time Fourier transform on the sound signal to be detected, input it into the trained fault detection model to obtain the fault diagnosis result.

[0005] According to the belt conveyor roller fault sound detection method disclosed in the present invention, preferably, the time-frequency domain transformation step specifically includes: performing a short-time Fourier transform on the one-dimensional sound signal to obtain a two-dimensional time-frequency domain feature map X, and the time-frequency domain feature map X is input as a fault sample into a subsequent fault detection model for analysis, wherein the short-time Fourier transform calculation process includes:

[0006]

[0007] in, is the window function, is the input signal, express In time and frequency The window function is Hamming window, the window size is set to 2048, and the step size is set to 512.

[0008] According to the belt conveyor roller fault sound detection method disclosed in the present invention, preferably, the multi-scale feature extraction step specifically includes: the time-frequency domain feature map X passes through a 3×3 convolution layer, a 5×5 convolution layer and a 7×7 convolution layer in parallel, and the number of output channels of each convolution layer is 16; the output feature map after the i-th convolution layer is recorded as , then Splicing is performed on the channel dimension to obtain the fused multi-scale feature map .

[0009] According to the belt conveyor roller fault sound detection method disclosed in the present invention, preferably, the step of high-level semantic extraction specifically includes: multi-scale feature map Enter the residual connection branch and the first Ghost module respectively. The first Ghost module performs multi-scale feature maps. Perform channel expansion and preliminary feature extraction to obtain a compact feature map ; Compact feature map Enter the deep convolution layer to further extract local features and obtain enhanced feature maps ; Enhanced feature map Entering the spatial channel collaborative attention module, based on the synergy of spatial attention and channel attention, the network's attention to important features is strengthened, the feature expression ability is improved, and the spatial channel collaborative attention feature map is obtained. ; The second Ghost module coordinates the attention feature map of the spatial channel Compress the number of channels to obtain fusion features ; Multi-scale feature map and fusion features Add together to obtain a high-level semantic feature map .

[0010] According to the belt conveyor roller fault sound detection method disclosed in the present invention, preferably, the calculation process of the spatial channel collaborative attention module specifically includes: the spatial channel collaborative attention is composed of two parts: shared multi-semantic spatial attention and progressive channel attention;

[0011] The calculation process of shared multi-semantic spatial attention includes:

[0012] Enhanced feature map Perform global average pooling along the height and width dimensions to obtain two unidirectional one-dimensional sequences; decompose the two one-dimensional sequences separately to obtain four sub-features of the same size; use four one-dimensional depth-shared convolutions to extract features from each sub-feature to obtain sub-feature spatial structure information; splice the sub-feature spatial structure information along the channel direction, perform group normalization, and calculate the spatial attention weight using the Sigmoid activation function; use the spatial attention weight to enhance the feature map Weighted to obtain a shared multi-semantic spatial attention feature map ;

[0013] The calculation process of progressive channel attention includes:

[0014] Attention feature map for shared multi-semantic space Perform global average pooling and obtain the downsampled feature map after downsampling; perform group normalization on the downsampled feature map, and then use 1×1 depth convolution to generate three sets of corresponding values; input the three sets of values into the channel-based attention to calculate the attention weight, and use global average pooling and Sigmoid activation to generate channel weights; according to the channel weights, the shared multi-semantic space attention feature map is Weighted, we get the spatial channel collaborative attention feature map .

[0015] According to the belt conveyor roller fault sound detection method disclosed in the present invention, preferably, the calculation process of the first Ghost module or the second Ghost module specifically includes: the input feature map is convolved by 1×1 main branch to generate the main branch feature ; Use 3×3 depth convolution to perform the main branch features Perform cheap operations and generate redundant features ; The main branch feature and redundant features Splicing is performed in the channel dimension to obtain the final output of the Ghost module.

[0016] According to the belt conveyor roller fault sound detection method disclosed in the present invention, preferably, the enhanced characteristic map The calculation process includes: compact feature map Input into the depth convolution layer, the calculation process of the depth convolution layer is expressed as:

[0017]

[0018] Among them, K represents the depth convolution kernel size, Represented in the enhanced feature map The value of the Cth channel position (i, j) in, Represents the weight of the depth convolution kernel of the Cth channel at position (m,n), Representing compact feature maps The value of the C-th channel after downsampling with a step size of S.

[0019] According to the belt conveyor roller fault sound detection method disclosed in the present invention, preferably, the signal collection step specifically includes: collecting sound signals using distributed stethoscope optical fibers or microphone arrays.

[0020] The second aspect of the present invention discloses a belt conveyor roller fault sound detection system, comprising: a memory for storing program instructions; a processor for calling the program instructions stored in the memory to implement a belt conveyor roller fault sound detection method as any of the above technical solutions.

[0021] The beneficial effects of the present invention include at least the following: the present invention can fully learn the feature information between features of different scales, enhance the input feature expression, improve the modeling ability of local features and global context information, reduce computational complexity, improve the model fitting performance, and achieve accurate and real-time detection of typical belt conveyor roller and bearing faults, and improve the accuracy and robustness of fault diagnosis under noise interference. In addition, the present invention has a small number of parameters and can be deployed in embedded and mobile devices, laying the foundation for intelligent diagnosis of belt conveyor roller faults in practical industrial applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A schematic flow chart of a method for detecting sound of belt conveyor roller faults according to an embodiment of the present invention is shown.

[0023] Figure 2 A schematic diagram of the network structure of a multi-scale convolution module according to an embodiment of the present invention is shown.

[0024] Figure 3 A schematic diagram of the network structure of a Ghost bottleneck module according to an embodiment of the present invention is shown.

[0025] Figure 4 A schematic diagram of the network structure of a classifier module according to an embodiment of the present invention is shown.

[0026] Figure 5 A schematic diagram of the network structure of a Ghost module according to an embodiment of the present invention is shown.

[0027] Figure 6 A schematic diagram of a network structure for sharing multi-semantic spatial attention according to an embodiment of the present invention is shown.

[0028] Figure 7 A schematic diagram of a network structure of progressive channel attention according to an embodiment of the present invention is shown.

[0029] Figure 8 A confusion matrix result diagram of a roller sound dataset according to an embodiment of the present invention is shown.

[0030] Figure 9 A schematic block diagram of a belt conveyor roller fault sound detection system according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0031] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0032] like Figure 1 As shown, according to one embodiment of the present invention, a method for detecting sound of a belt conveyor roller fault is disclosed, comprising:

[0033] Step S1, signal acquisition: using a microphone array to collect the sound signal of the rotating belt conveyor roller when it is running, and capturing the spatial distribution and phase characteristics of the sound signal.

[0034] Step S2, time-frequency domain transformation: Perform a short-time Fourier transform on the one-dimensional sound signal to obtain a two-dimensional time-frequency domain feature map X, which is used as a fault sample input into the fault detection model for analysis. The short-time Fourier transform (STFT) is defined as:

[0035]

[0036] in, is the window function, is the input signal, the output Reflection signal In time and frequency The signal component is , the window function used is Hamming window, the window size is set to 2048, and the step size is set to 512.

[0037] Step S3, multi-scale feature extraction: input the time-frequency domain feature map X into the multi-scale convolution module for feature extraction; the network structure of the multi-scale convolution module is as follows: Figure 2 As shown, C represents splicing:

[0038] First, the time-frequency domain feature map X passes through the 3×3 convolution layer, the 5×5 convolution layer and the 7×7 convolution layer in parallel. The number of output channels of each convolution layer is 16. The output feature map after the i-th convolution layer is recorded as ;

[0039] Then, Splicing is performed on the channel dimension to obtain the fused multi-scale feature map :

[0040]

[0041]

[0042] Among them, B represents the batch size, C represents the number of input channels, H and W represent the height and width of the feature map, Indicates that the convolution kernel size is The convolution operation, Indicates splicing along the channel direction, Represents the Rectified Linear Unit (ReLU) activation function operation.

[0043] Step S4, high-level semantic extraction: multi-scale feature map The input is sent to the Ghost bottleneck module to further extract high-level semantic information. The Ghost bottleneck module consists of two Ghost modules, a deep convolutional layer, and a spatial channel collaborative attention module;

[0044] The network structure of the Ghost bottleneck module is as follows Figure 3 As shown:

[0045] First, multi-scale feature maps Enter the residual connection branch and the first Ghost module respectively. The first Ghost module performs multi-scale feature map Perform channel expansion and preliminary feature extraction to obtain a compact feature map ;

[0046] Then, the compact feature map Enter the deep convolution layer for empty transformation to further extract local features and obtain enhanced feature maps ;

[0047] Next, enhance the feature map Entering the spatial channel collaborative attention module, through the synergy of spatial attention and channel attention, the network's attention to important features is strengthened, the feature expression ability is improved, and the spatial channel collaborative attention feature map is obtained. ;

[0048] Then, the second Ghost module coordinates the attention feature map of the spatial channel Compress the number of channels to obtain fusion features ;

[0049] Finally, the multi-scale feature map and fusion features Add together to obtain a high-level semantic feature map , realize the skip connection of features, enhance feature reuse, and alleviate the gradient disappearance problem.

[0050] Step S5, feature classification: high-level semantic feature maps The input is sent to the classifier module for feature classification and the model training is completed using the cross entropy loss function and reparameterization training method to obtain a fault detection model.

[0051] The network structure of the classifier module is as follows Figure 4 As shown: First, the high-level semantic feature map The input is sent to the adaptive average pooling layer, and then the number of channels is converted through a 1×1 convolution operation. Finally, the fully connected layer maps the feature vector to the number of output categories:

[0052]

[0053] in, represents a full connection operation, represents the global average pooling operation, and Cls represents the output of the classifier module.

[0054] Optimize model training using the cross-entropy loss function: The loss function is a tool for measuring the difference between the probability distribution of the model's predicted labels and the probability distribution of the true labels. In the core process of network training, the essence of model optimization is to continuously reduce the value of the loss function using the backpropagation algorithm. This embodiment uses the cross-entropy loss function as a means of optimizing network parameters. The calculation formula is:

[0055]

[0056] in, represents the true fault category of the i-th element, represents the probability that the model predicts x belongs to the i-th fault category.

[0057] The reparameterized training method is used to train the optimized model: linear parallel branches are introduced in depthwise convolution and 1×1 convolution, which does not bring additional cost during inference; since convolution and BN normalization operations are linear during inference, they can be folded into a single convolution layer, and the weight matrix of each convolution layer is , the bias is , Indicates the number of output channels, Represents the number of input channels, k represents the convolution kernel size; the weight matrix and bias are reparameterized to the sum of the weights and biases of all branches; during inference, duplicate branches are deleted through the inverse reparameterization process; this process involves reparameterizing the folded weights and biases in all branches, thereby achieving the purpose of reducing the amount of computation during inference; specifically, the inverse reparameterization of weights and biases is shown in the following formula:

[0058]

[0059]

[0060] where i is the index of the repeated branch, is the reparameterized weight, is the reparameterized bias.

[0061] Step S6, fault detection: After performing short-time Fourier transform on the sound signal to be detected, it is input into the trained fault detection model to obtain the fault diagnosis result. The fault detection model is implemented based on the above steps S3, S4 and S5. The time-frequency domain feature map X of the input model is extracted through the multi-scale convolution module to obtain a multi-scale feature map , and then multiple GBNM modules are used to extract high-level semantics to obtain high-level semantic feature maps , and finally obtain the classification results through the classifier module.

[0062] like Figures 3 to 6 As shown, according to the above embodiment, further, step S4 specifically includes:

[0063] S4.1, through the first Ghost module in the Ghost bottleneck module (the first Ghost module) multi-scale feature map Perform channel expansion and preliminary feature extraction to obtain a compact feature map The specific steps are as follows:

[0064] like Figure 5 As shown, first, the input feature map is convolved with the 1×1 main branch to generate a part of the feature map, and the main branch output is recorded as , the number of channels is ; Next, the Ghost module uses 3×3 depth convolution to Perform cheap operations to generate redundant feature maps ;Finally, the main branch features and cheap branching features Splicing is performed on the channel dimension to obtain the final compact feature map output by the Ghost module :

[0065]

[0066]

[0067]

[0068] in, Indicates the number of output channels, Represents a convolution operation with a convolution kernel size of 1×1, Represents a depthwise convolution operation with a kernel size of 3×3.

[0069] S4.2, through the depth convolution layer (DW Conv) to the compact feature map Perform empty transformation to further extract local features and obtain enhanced feature maps The specific steps are as follows:

[0070] The compact feature map processed by the first Ghost module Input to the depth convolution layer, C represents the number of channels, H and W represent the height and width of the feature map; the output feature map is Y, where S represents the stride. When the stride is Stride>1, the depth convolution downsamples the multi-scale feature map, and K represents the depth convolution kernel size:

[0071]

[0072] in, Represented in the enhanced feature map The value of the Cth channel position (i, j) in, Represents the weight of the depth convolution kernel of the Cth channel at (m,n), Represents the input compact feature map The value of the C-th channel after downsampling with a step size of S.

[0073] S4.3, enhance the feature map through the spatial channel collaborative attention module (SCSA) Based on the synergy of spatial attention and channel attention, the network's attention to important features is strengthened, the feature expression ability is improved, and the spatial channel collaborative attention feature map is obtained. Among them, spatial channel collaborative attention consists of two parts: shared multi-semantic spatial attention and progressive channel attention. The specific implementation steps are as follows:

[0074] like Figure 6 As shown in Figure 2, the calculation process of shared multi-semantic spatial attention includes:

[0075] Enhanced feature map After inputting into the spatial channel collaborative attention, first the multi-semantic spatial attention is shared to enhance the feature map Perform global average pooling along the height and width dimensions to obtain two unidirectional one-dimensional sequences and Next, shared multi-semantic spatial attention is used to reduce the number of parameters and computational costs, and to meet the purpose of learning different spatial distributions and contextual relationships. The two one-dimensional sequences are decomposed and divided into K independent sub-features of the same size. and ; The number of channels of each sub-feature is C / K, K=4:

[0076]

[0077]

[0078]

[0079]

[0080] Among them, B represents the batch size, C represents the number of input channels, H and W represent the height and width of the feature map, Indicates a global average pooling operation along the horizontal direction, Indicates a global average pooling operation along the vertical direction, Represents the first sub-feature, which is composed of local features in the height dimension and global features at three different scales in the height dimension , , composition, Represents the second sub-feature, which is composed of local features on the width dimension Global features of three different scales in the width dimension , , composition;

[0081] Then use four convolution kernel sizes The one-dimensional depth shared convolution of 3, 5, 7, and 9 for each sub-feature and Perform feature extraction and implicitly establish a dependency relationship between the two dimensions by learning consistent features in the horizontal and vertical dimensions:

[0082]

[0083]

[0084] in, and Represents the spatial structure information of the i-th sub-feature obtained after the depth-shared convolution operation; Indicates that the convolution kernel size is One-dimensional depth-shared convolution operation;

[0085] The extracted sub-feature spatial structure information is spliced along the channel direction and processed using group normalization GN; the spatial attention weight is calculated using the Sigmoid activation function and ; Finally, the spatial attention weight is used to enhance the feature map Perform weighting to enhance the features of key areas:

[0086]

[0087]

[0088] X s =Attn H ×Attn W ×F E

[0089] in, Represents the Sigmoid activation function operation, and denotes K groups of normalized GN operations along the horizontal and vertical directions, respectively. Represents the final output of the shared multi-semantic spatial attention module.

[0090] like Figure 7 As shown in the figure, the calculation process of progressive channel attention includes: first, the feature map processed by shared multi-semantic space attention is Perform global average pooling operation and obtain the downsampled feature map after downsampling , in order to reduce the computational complexity while aggregating local features and improving the receptive field:

[0091]

[0092] Among them, B represents the batch size, C represents the number of input channels, H and W represent feature maps The height and width, and Representation feature map The height and width, represents a pooling operation with a kernel size of 7×7;

[0093] Then the downsampled feature map Perform group normalization GN processing, then use 1×1 depth convolution to generate three sets of corresponding values, and then input the three sets of values into the channel-based attention to calculate the attention weights , use global average pooling and Sigmoid activation to generate channel weights; finally, the channel weights are combined with Weighted, improve the network's perception ability, and obtain the spatial channel collaborative attention feature map :

[0094]

[0095] Among them, B represents the batch size, C represents the number of input channels, H and W represent the height and width of the feature map, represents the global average pooling operation, It represents the final output of the progressive channel attention and also the final output of the collaborative attention of the entire spatial channel.

[0096] S4.4, through the second Ghost module to coordinate attention feature map of spatial channels Compress the number of channels to obtain fusion features , the multi-scale feature map With the fusion feature Add together to obtain a high-level semantic feature map .

[0097] According to the above embodiment, further, the neural network training parameters are set: batch size 32, number of iterations 1200, initial learning rate 0.01, cross entropy loss function is used and stochastic gradient descent algorithm SGD is adopted as the network optimizer, the learning rate adopts fixed epoch decay, the learning rate is halved every 300 epochs, the momentum coefficient is 0.9, and the weight decay coefficient is 5e-4.

[0098] like Figure 8As shown in the figure, in order to intuitively explain the results of various fault diagnosis, a confusion matrix is drawn. The horizontal axis represents the classification result, and the vertical axis represents the actual fault category. The data set size is 5400 time-frequency graphs, with a total of 17 types of typical roller faults and 1 intact roller as a reference, which are divided into training set and test set in a ratio of 7:3. Figure 8 It can be seen that the above embodiments of the present invention achieve belt conveyor roller fault detection, thereby improving the accuracy and reliability of fault detection.

[0099] like Figure 9 As shown, according to another embodiment of the present invention, a belt conveyor roller fault sound detection system 900 is also disclosed, including: a memory 901 for storing program instructions; a processor 902 for calling the program instructions stored in the memory to implement the belt conveyor roller fault sound detection method as described in the above embodiment.

[0100] In summary, the present invention uses distributed auscultation fibers or microphone arrays to collect sound signals from belt conveyor rollers during operation, capturing the spatial distribution and phase characteristics of the sound signals. A short-time Fourier transform (SFT) is used to obtain a time-frequency domain feature map of the sound signal, improving the ability to express fault characteristics. The time-frequency domain feature map is input into a multi-scale convolution module composed of a multi-scale convolution parallel structure to obtain a multi-scale feature map containing multi-scale information, while also improving the model's ability to capture features at different scales. The multi-scale feature map is input into a Ghost bottleneck module composed of a Ghost module, a deep convolution layer, and a spatial channel collaborative attention module to perform deeper fault feature mapping in the spatial and channel dimensions, obtaining a high-level semantic feature map and improving the accuracy of the model's fault identification. The high-level semantic feature map is input into a classifier module to output the number of categories, and the model is trained using a cross-entropy loss function and a reparameterized training method. The sound signal to be detected is subjected to a SFT and then input into the trained model to determine whether the roller has failed. This achieves a more accurate and intelligent fault detection solution.

[0101] All or part of the steps in the various methods of the above embodiments can be completed by controlling related hardware through a program. The program can be stored in a readable storage medium, and the storage medium includes read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electronically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, magnetic disk storage, magnetic tape storage, or any other readable medium that can be used to carry or store data.

[0102] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A belt conveyor roller fault sound detection method, characterized in that: include: Signal collection: collect the sound signals of belt conveyor rollers when they are running; Time-frequency domain transformation: performing short-time Fourier transform on the sound signal to obtain a time-frequency domain feature map X of the sound signal; Multi-scale feature extraction: extract features from the time-frequency domain feature map X through multiple convolutional layers of different sizes to obtain feature maps f of multiple scales. i , the feature map f i Splicing is performed on the channel dimension to obtain a multi-scale feature map F C The multi-scale feature extraction step specifically includes: the time-frequency domain feature map X passes through a 3×3 convolution layer, a 5×5 convolution layer, and a 7×7 convolution layer in parallel, and the number of output channels of each convolution layer is 16; the output feature map after the i-th convolution layer is denoted as f i , then f i Splicing is performed on the channel dimension to obtain the fused multi-scale feature map F C ; High-level semantic extraction: The multi-scale feature map F C The first Ghost module, deep convolution layer, spatial channel collaborative attention module and the second Ghost module are used to perform deeper fault feature mapping in spatial and channel dimensions to obtain a high-level semantic feature map F H ; Feature classification: the high-level semantic feature map F H Input into the classifier module to obtain the classification results, and use the cross entropy loss function and reparameterization training method to complete the model training to obtain the fault detection model; Fault detection: After performing short-time Fourier transform on the sound signal to be detected, it is input into the trained fault detection model to obtain the fault diagnosis result.

2. The belt conveyor roller fault sound detection method according to claim 1, characterized in that: The time-frequency domain transformation step specifically includes: Perform a short-time Fourier transform on the one-dimensional sound signal to obtain a two-dimensional time-frequency domain feature map X. The time-frequency domain feature map X is input as a fault sample into the subsequent fault detection model for analysis. The calculation process of the short-time Fourier transform includes: Where g(t) is the window function, f(t) is the input signal, STFT(ω,τ) represents the signal component of f(t) at time τ and frequency ω, the window function is a Hamming window, the window size is set to 2048, and the step size is set to 512.

3. The belt conveyor roller fault sound detection method according to claim 1, characterized in that: The step of extracting high-level semantics specifically includes: The multi-scale feature map F C Enter the residual connection branch and the first Ghost module respectively, the first Ghost module performs the multi-scale feature map F C Perform channel expansion and preliminary feature extraction to obtain a compact feature map F M ; The compact feature map F M Enter the deep convolution layer to further extract local features and obtain the enhanced feature map F E ; The enhanced feature map F E Entering the spatial channel collaborative attention module, based on the synergy of spatial attention and channel attention, the network's attention to important features is strengthened, the feature expression ability is improved, and the spatial channel collaborative attention feature map X is obtained. Co ; The second Ghost module coordinates the attention feature map X of the spatial channel Co Compress the number of channels to obtain the fusion feature X F ; The multi-scale feature map F C With the fusion feature X F Add together to obtain the high-level semantic feature map F H .

4. The belt conveyor roller fault sound detection method according to claim 3, characterized in that: The calculation process of the spatial channel collaborative attention module specifically includes: spatial channel collaborative attention consists of two parts: shared multi-semantic spatial attention and progressive channel attention; The calculation process of the shared multi-semantic spatial attention includes: For the enhanced feature map F E Perform global average pooling along the height and width dimensions to obtain two unidirectional one-dimensional sequences; Decompose the two one-dimensional sequences separately to obtain four sub-features of the same size; Use four one-dimensional depth-shared convolutions to extract features from each sub-feature and obtain the sub-feature spatial structure information; The sub-feature spatial structure information is spliced along the channel direction, group normalization is performed, and the spatial attention weight is calculated using the Sigmoid activation function; The enhanced feature map F is processed using spatial attention weights E Perform weighting to obtain the shared multi-semantic space attention feature map X S ; The calculation process of the progressive channel attention includes: For the shared multi-semantic spatial attention feature map X S Perform global average pooling operation and obtain the downsampled feature map after downsampling; Perform group normalization on the downsampled feature map, and then generate three sets of corresponding values using 1×1 depthwise convolution; Input the three sets of values into the channel-wise attention to calculate the attention weights, and use global average pooling and Sigmoid activation to generate channel weights; According to the channel weight, the shared multi-semantic spatial attention feature map X S Weighted, we get the spatial channel collaborative attention feature map X Co .

5. The belt conveyor roller fault sound detection method according to claim 3, characterized in that: The calculation process of the first Ghost module or the second Ghost module specifically includes: The input feature map is convolved with a 1×1 main branch to generate the main branch feature Y1; Using 3×3 depth convolution to perform cheap operations on the main branch feature Y1 to generate redundant features Y2; The main branch feature Y1 and the redundant feature Y2 are concatenated in the channel dimension to obtain the final output of the Ghost module.

6. The belt conveyor roller fault sound detection method according to claim 3, characterized in that: Enhanced feature map F E The calculation process includes: The compact feature map F M Input into the depth convolution layer, the calculation process of the depth convolution layer is expressed as: Among them, K represents the depth convolution kernel size, Y C (i,j) represents the enhanced feature map F E The value of the Cth channel position (i, j), ω C (m,n) represents the weight of the depth convolution kernel of the C channel at position (m,n), X C (S·i+m,S·j+n) represents the compact feature map F M The value of the C-th channel after downsampling with a step size of S.

7. The belt conveyor roller fault sound detection method according to any one of claims 1 to 6, characterized in that: The signal acquisition step specifically includes: The sound signal is collected using a distributed auscultation optical fiber or a microphone array.

8. A belt conveyor roller fault sound detection system, characterized in that: include: a memory for storing program instructions; A processor is used to call the program instructions stored in the memory to implement the belt conveyor roller fault sound detection method according to any one of claims 1 to 7.

Images