Conveying belt defect detection method based on GLSA-YOLOv11

Abstract

The invention provides a conveying belt defect detection method based on a GLSA-YOLOv11 model, and relates to the technical field of target defect identification, and the method comprises the following steps: S1, constructing a target defect data set; s2, preprocessing the data set; s3, constructing a GLSA-YOLOv11-based transport belt tearing and fracture defect detection model comprising a FasterNet backbone network, a GLSA weight adjustment module, a dual-convolution ELAN neck network and a prediction layer, and inputting the preprocessed image into the model for pre-training to obtain a transport belt defect detection model; s4, acquiring a video image of the conveying belt in a monitoring scene, performing the preprocessing operation in the S2, and inputting a processing result into the pre-trained conveying belt defect detection model; and S5, outputting a target defect judgment result of the target transportation equipment, and performing alarm early warning if defects exist, so as to solve the technical problems of low image quality and difficulty in real-time online high-precision detection and identification of belt defects caused by interference in a complex scene of a coal mine in the prior art.

Description

Technical Field

[0001] The present invention relates to the field of target defect recognition technology, and more specifically, to a conveyor belt defect detection method based on GLSA-YOLOv11. Background Art

[0002] Object detection is one of the core technologies in the field of computer vision and is widely used in scenarios such as security monitoring, autonomous driving, and industrial inspection. Among them, the YOLOv series of models are widely favored for their remarkable balance between speed and accuracy. Today, YOLOv11 has further significantly improved its performance by optimizing the network structure and training strategy, and has been well applied in many fields. In the field of intelligent security, the model supports face recognition, behavior analysis and dangerous goods detection in real-time video streams, greatly improving the response speed of the monitoring system; in autonomous driving, YOLOv11 can accurately identify vehicles, pedestrians and traffic signs to ensure driving safety; in industrial scenarios, it is used for product defect detection, and improves quality inspection efficiency by locating the defect position with high precision.

[0003] Although YOLOv11 has shown significant potential in the field of industrial defect detection with its autonomous feature learning capabilities, it still faces multiple challenges in some complex industrial environments (such as coal mines): First, the harsh environment underground causes serious degradation of image quality, low light conditions weaken feature visibility, and dynamically diffused coal dust and water mist form continuous interference noise, significantly reducing the signal-to-noise ratio of the target area; second, because the belt is subjected to high loads and complex working conditions (such as coal block impact, mechanical wear, humid environment, etc.) for a long time, it is easy to cause defects such as surface tearing and longitudinal fracture. These defects are usually special, and the tearing cracks are mostly fine. The long, discontinuous linear structure and the morphology of the fracture area change randomly with the direction of force. The defective target occupies a small proportion in the image and has blurred edges, which can easily be drowned out by background noise such as belt texture and coal powder adhesion. In addition, underground safety monitoring has extremely high real-time requirements. If it is not detected in time, it may cause belt breakage, material leakage or even fire accidents, causing significant economic losses and casualties. Therefore, millisecond-level detection response is required to quickly block the accident chain. However, existing deep learning models often find it difficult to meet the requirements of lightweight deployment while ensuring the accuracy of small target detection. As a result, the speed-accuracy balance has become a key bottleneck restricting the implementation of the technology. Summary of the Invention

[0004] In response to the defects of the existing technology or the need for improvement, the present invention provides a conveyor belt defect detection method based on GLSA-YOLOv11, which aims to solve the technical problem that the existing technology has low light, high dynamic coal dust, and water mist interference in complex scenes of coal mines, resulting in poor image quality and difficulty in real-time online high-precision detection and identification of belt defects.

[0005] According to an embodiment of the present invention, a method for detecting conveyor belt defects based on GLSA-YOLOv11 includes the following steps:

[0006] S1: Build target defect dataset D;

[0007] S2: Preprocess the dataset D;

[0008] S3: Build a conveyor belt tear and break defect detection model based on GLSA-YOLOv11, including the FasterNet backbone network, the GLSA weight adjustment module, the dual convolutional ELAN neck network, and the prediction layer. Input the preprocessed images into the model for pre-training to obtain the conveyor belt defect detection model.

[0009] S4: Acquire a video image of the conveyor belt in the monitoring scene, perform the preprocessing operation in S2, and input the processing result into the pre-trained conveyor belt defect detection model;

[0010] S5: Output the target defect identification results of the target transportation equipment. If there is a defect, an alarm warning will be issued and the defect type will be transmitted to the console staff for timely manual intervention.

[0011] Preferably, step S1 includes the following sub-steps:

[0012] S11: Acquire multiple sets of historical defect detection record image data of target transportation equipment, and identify different defect types, thereby constructing a data set D = {d1, d2, ..., d t};

[0013] S12: extracting the first image feature information from the plurality of sets of image feature information with defect type identification, wherein the first image feature information corresponds to the first defect type d1, and so on, the tth image feature information corresponds to the tth defect type d t .

[0014] Preferably, step S2 includes the following sub-steps:

[0015] S21: Using the dataset D as the input image f, grayscale erosion, grayscale expansion, and grayscale opening and closing operations are performed;

[0016] S22: performing de-darkening processing on the image f to obtain a dark channel of the input image f;

[0017] S23: Obtain unknown prior conditions that need to be estimated in the model based on the dark channel

[0018] S24: Based on the unknown prior conditions obtained And the current ambient light value A, get the darkened image J(x);

[0019] S25: Optimize the speed and accuracy of image defect edge extraction, and use a linear transformation preprocessing method to process the image J(x) after the darkening operation.

[0020] Preferably, step S3 includes the following sub-steps:

[0021] S31: Build a conveyor belt tear and break defect detection model based on GLSA-YOLOv11, including the FasterNet backbone network, GLSA weight adjustment module, dual convolutional ELAN neck network and prediction layer;

[0022] S32: Input the preprocessed image into the FasterNet backbone network and output the feature map;

[0023] S33: Input the feature map formed by the FasterNet backbone network into the GLSA weight adjustment module, determine the weight vector of the defect feature through the GLSA weight adjustment module, and perform weight distribution to obtain a new feature map;

[0024] S34: The weighted feature map is input into the dual convolutional ELAN neck network. In the GLBiFPN of the dual convolutional ELAN neck network, the feature map is propagated along the upper and lower structures of the network, and feature fusion is performed at each level to obtain meta-features.

[0025] S35: Use the multi-scale decoupling head Detect of the prediction layer to output the detection results;

[0026] S36: Design multi-task classification training loss function F loss .

[0027] Preferably, the FasterNet backbone network in step S32 is responsible for extracting global image defect feature information to form a feature map, which includes four layers, and each layer is provided with an embedding layer or a merging layer at the front end for spatial downsampling and channel expansion. FasterNet is constructed based on PConv, and the global pooling in the Fasternet backbone network is replaced by strip pooling. The strip pooling structure consists of two branches, which respectively use H×1 and 1×W pooling kernels, focusing on encoding long-range context in the horizontal and vertical spatial dimensions to achieve the expansion of the backbone network receptive field, and defining a certain position of the feature map as y∈R C×H×W ,in but:

[0028]

[0029] Finally, the output feature map calculation formula is:

[0030] z=S(x,σ(f(y))),

[0031] Where x is the input feature map, S(·) is the numerical multiplication, σ(·) is the sigmoid activation function, and f(·) is the 1×1 convolutional layer.

[0032] Preferably, the GLSA weight adjustment module in step S33 divides the channels of the input feature map Fi into two groups Fi1 and Fi2, and sends them to the global spatial attention module and the local spatial attention module for processing respectively, and then fuses the outputs of the two modules through a 1×1 convolution layer to generate an enhanced feature map. This process expression is:

[0033]

[0034] In the GSA module, the global spatial attention mechanism complements the local spatial attention by capturing the long-distance dependencies between pixels. The calculation formula is:

[0035]

[0036]

[0037] Among them, AttG(·) represents the global attention operation, C (1×1) represents 1×1 convolution, Represents matrix multiplication. MLP consists of two fully connected layers, which are connected by ReLU activation function and normalization layer. The first layer expands the input to high-dimensional space with an expansion ratio of 2, and the second layer compresses its dimension to be consistent with the original input. In the LSA module, the local spatial attention mechanism further enhances the feature expression ability by extracting local features of the region of interest in the spatial dimension. The calculation formula is:

[0038]

[0039] Among them, F c (·) denotes a cascade structure consisting of three 1×1 convolutional layers and one 3×3 depthwise convolutional layer, AttL(·) denotes the local attention operation, σ(·) is the Sigmoid function, and ⊙ denotes point-wise multiplication.

[0040] Preferably, the dual convolution ELAN neck network in step S34 adopts a CSP architecture, which divides the feature map into two parts for independent processing. One part of the feature map is directly transmitted, and the other part stacks multiple dual convolution DualConv modules to gradually fuse multi-scale features. The first half directly retains the original information and is spliced ​​with the enhanced features. The GLBiFPN structure adopts a bidirectional cross-scale connection. This connection method can not only perform bottom-up feature fusion, but also top-down fusion. Each fusion node in the GLBiFPN structure adopts weighted summation to fuse different features, including the same-scale features of the previous layer, the higher-scale features obtained by upsampling, and the lower-scale features obtained by downsampling. GLBiFPN adopts a fast normalized fusion weighting method to add corresponding weights to each input feature. The calculation formula is as follows:

[0041]

[0042] Among them, for each w i Add activation function ReLU to ensure w i >0 to avoid numerical instability.

[0043] Preferably, the loss function F in step S36 loss The calculation formula is as follows:

[0044]

[0045] Among them, i is the index number of the anchor point of the i-th candidate region in the image, p i Represents the target prediction probability of the anchor point; represents the candidate region label, N loc With N cls Respectively represent the classification loss function in the positioning loss L loc With similarity metric loss L cls Normalization coefficient after functionalization, λ represents N loc With N cls The balance weight between loc represents the positioning loss function, L cls Represents the similarity measurement loss function, the positioning loss function L loc The calculation formula is as follows:

[0046] L loc =αL reg +βL CIoU ,

[0047] Among them L reg Represents the position regression loss function, L CIoU is the generalized intersection-over-union loss function, and α and β are hyperparameters used to adjust the functions of each part;

[0048] The similarity metric loss function L cls The calculation formula is as follows:

[0049]

[0050] Among them, x represents the true sample label, and x′ represents the detection output after the activation function.

[0051] Preferably, step S4 includes the following sub-steps:

[0052] S41: Reasonably determine multiple monitoring points and obtain monitoring stream videos T of different transportation angles;

[0053] S42: Extract frames from video T to obtain a monitoring image of the transport equipment belt, and set the current frame image as T t , then the previous frame image is T t-1 ;

[0054] S43: Perform image scale normalization and image standardization on each frame, and perform image normalization on the current frame image T t Perform the preprocessing operation in S2 and then input it into the pre-trained conveyor belt defect detection model to detect the defect type.

[0055] Preferably, step S5 includes the following sub-steps:

[0056] S51: Obtaining a target defect discrimination result of the target transport equipment belt through a transport belt defect detection model;

[0057] S52: If there is a defect, an alarm is issued and the defect type is transmitted to the control console staff;

[0058] S53: Determine whether manual intervention is required for the belt defect based on the target defect identification result at the current moment, and perform maintenance and repair in a timely manner.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] 1. The method of the present invention is based on the single-stage detection framework of YOLOv11. It introduces deep separable convolution through the FasterNet backbone network to compress the computational complexity, optimize the dynamic label allocation strategy, enhance the discrimination between positive and negative samples, and significantly improve the detection accuracy while maintaining real-time inference performance, thus achieving the coordinated optimization of speed and accuracy.

[0061] 2. The method of the present invention innovatively designs the GLSA mechanism, uses global context modeling to capture the continuity feature association of long-distance tear cracks, and focuses on the pixel-level differences in subtle fracture areas through local spatial attention, effectively resisting environmental interference such as coal dust adhesion and water mist occlusion, enhancing the ability to extract significant features of small target defects, and significantly reducing the false detection rate in complex scenarios.

[0062] 3. Aiming at the low computing power deployment requirements underground, the method of the present invention constructs multi-scale feature fusion technology through a dual-convolutional ELAN neck network to enhance the model's generalization ability for complex morphological defects. Combined with lightweight strategies such as network layer pruning and quantization compression, it greatly reduces the model complexity and computing resource requirements, ensuring the real-time operation of the algorithm on embedded devices and meeting the deployment constraints of underground industrial scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a flow chart of a method for detecting conveyor belt defects according to an embodiment of the present invention.

[0064] Figure 2 It is a structural diagram of a conveyor belt defect detection model according to an embodiment of the present invention.

[0065] Figure 3 4 is a diagram of a strip pooling structure in a Fasternet backbone network according to an embodiment of the present invention.

[0066] Figure 4 4 is a structural diagram of a GLSA weight adjustment module according to an embodiment of the present invention.

[0067] Figure 5 2 is a diagram of the dual convolutional ELAN neck network structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0068] The technical solutions of the present invention are further described below with reference to the accompanying drawings and embodiments.

[0069] Example 1

[0070] like Figure 1 As shown, the embodiment of the present invention proposes a conveyor belt defect detection method based on GLSA-YOLOv11, which includes the following steps:

[0071] S1: Build target defect dataset D;

[0072] S2: Preprocess the dataset D;

[0073] S3: Build a conveyor belt tear and break defect detection model based on GLSA-YOLOv11, including the FasterNet backbone network, the GLSA weight adjustment module, the dual convolutional ELAN neck network, and the prediction layer. Input the preprocessed images into the model for pre-training to obtain the conveyor belt defect detection model.

[0074] S4: Acquire a video image of the conveyor belt in the monitoring scene, perform the preprocessing operation in S2, and input the processing result into the pre-trained conveyor belt defect detection model;

[0075] S5: Output the target defect identification results of the target transportation equipment. If there is a defect, an alarm warning will be issued and the defect type will be transmitted to the console staff for timely manual intervention.

[0076] Preferably, step S1 includes the following sub-steps:

[0077] S11: Acquire multiple sets of historical defect detection record image data of target transportation equipment, and identify different defect types, thereby constructing a data set D = {d1, d2, ..., d t};

[0078] S12: extracting the first image feature information from the plurality of sets of image feature information with defect type identification, wherein the first image feature information corresponds to the first defect type d1, and so on, the tth image feature information corresponds to the tth defect type d t .

[0079] Preferably, step S2 includes the following sub-steps:

[0080] S21: The dataset D is used as the input image f to perform grayscale corrosion, grayscale expansion, and grayscale opening and closing operations to eliminate interference information.

[0081] The function of grayscale corrosion is to process points with large grayscale values ​​on the edge of the image and reduce their grayscale values. Its calculation method can be expressed as:

[0082] (fΘg)(s,t)=min{f(s+x,t+y)-g(x,y)|s+x,t+y∈D f ,x+y∈D g},

[0083] Where fΘg represents the grayscale corrosion of the input image f(x,y) using the structural element g(x,y), D f 、D g Represent the domains of the input image f and the structural element g respectively;

[0084] Grayscale expansion is used to process points with smaller grayscale values ​​at the edge of an image, increase their grayscale values ​​and blend them with the image edge. Its calculation method can be expressed as:

[0085]

[0086] in Indicates that the grayscale expansion of the input image f(x,y) is performed using the structural element g(x,y), D f 、D g Represent the domains of the input image f and the structural element g respectively;

[0087] Use the structural element g(x,y) to perform grayscale erosion operation on the input image f(x,y), and then use the structural element g(x,y) to perform grayscale dilation operation. Then the image f(x,y) completes a grayscale opening operation. The calculation method of the grayscale opening operation can be expressed as:

[0088]

[0089] in Represents the grayscale opening operation. The grayscale opening operation removes the bright spots in the input image f(x,y) whose area is smaller than the structuring element g(x,y), keeps the larger bright areas unchanged, and keeps the overall grayscale value of the image unchanged.

[0090] Use the structuring element g(x,y) to perform grayscale dilation on the input image f(x,y). Then use the structuring element g(x,y) to perform grayscale erosion on the dilation result. Then the image f(x,y) completes a grayscale closing operation. The calculation method of the grayscale closing operation can be expressed as:

[0091]

[0092] Where f·g represents the grayscale closing operation. Grayscale closing removes dark spots in the input image f(x,y) that are smaller than the structuring element g(x,y), while preserving the bright areas of the image and the overall grayscale value of the image.

[0093] S22: performing a de-darkening process on the image f to obtain a dark channel of the input image f, thereby further eliminating interference information caused by the dim environment.

[0094] Get the dark channel map of the input image, find the top 0.1% pixels with the highest brightness, and form a pixel set. For the input image f, the mathematical description of the dark channel is as follows:

[0095] f dark (x) = min y∈Ω(x) [min c∈{r,g,b} f c (y)],

[0096] Among them, f dark is the dark channel, f c (y) represents the three color channels of the input image f, Ω(x) is the area window centered on pixel x, and according to the dark channel principle, f dark The intensity is extremely low, approaching 0.

[0097] S23: Obtain unknown prior conditions that need to be estimated in the model based on the dark channel

[0098] Assume that the transmittance of the pixel point is the same constant in the local neighborhood window Ω(x), which is recorded as Then take the minimum value in each neighborhood window, and on this basis take the minimum value of the color channel. According to the dark channel prior theory, there is a dark channel f of the light image f dark The value approaches 0, and the transmittance is obtained. for:

[0099]

[0100] Among them, ω is the parameter for adjusting the degree of darkening, f c (y) are the three color channels of the input captured image f.

[0101] S24: Based on the unknown prior conditions obtained And the current ambient light value A, the darkened image J(x) is obtained.

[0102] The calculation formula of the darkened image J(x) is as follows:

[0103]

[0104] Among them, J(x) represents the obtained de-darkened image, f(x) is the collected dark image, t(x) is the prior estimated transmittance, and the transmittance t(x) and the global ambient light value A are used to avoid restoration anomalies caused by a transmittance of 0. The algorithm introduces a constant parameter t0 as a lower limit. The calculation formula for the deep de-darkening network loss is as follows:

[0105]

[0106] in, represents the normal light image, J(x i ) is the de-darkened image output by the network, i represents the channel index, and N is the total number of pixels;

[0107] S25: Optimize the speed and accuracy of image defect edge extraction, and use a linear transformation preprocessing method to process the image J(x) after the darkening operation.

[0108] Assuming that the grayscale value of the image to be processed J(x) is in the interval [m,M], the grayscale value of the image D(x) after linear transformation will be expanded to [n,N]. The linear transformation process can be expressed as:

[0109]

[0110] Transformed into:

[0111] D(x)=αJ(x)+β,

[0112] in,

[0113] Preferably, Figure 2-5 As shown, step S3 includes the following sub-steps:

[0114] S31: Build a conveyor belt tear and break defect detection model based on GLSA-YOLOv11, including the FasterNet backbone network, GLSA weight adjustment module, dual convolutional ELAN neck network and prediction layer;

[0115] S32: Input the preprocessed image into the FasterNet backbone network and output the feature map;

[0116] S33: Input the feature map formed by the FasterNet backbone network into the GLSA weight adjustment module, determine the weight vector of the defect feature through the GLSA weight adjustment module, and perform weight distribution to obtain a new feature map;

[0117] S34: The weighted feature map is input into the dual convolutional ELAN neck network. In the GLBiFPN (global to local bidirectional feature fusion structure) of the dual convolutional ELAN neck network, the feature map is propagated along the upper and lower structures of the network, and feature fusion is performed at each level to obtain meta-features.

[0118] S35: Use the multi-scale decoupling head Detect of the prediction layer to output the detection results;

[0119] S36: Design multi-task classification training loss function F loss .

[0120] Preferably, the FasterNet backbone network in step S32 is responsible for extracting global image defect feature information to form a feature map, that is, achieving efficient feature extraction and aggregation. It contains four levels, and each level is provided with an embedding layer (batch size is 4, 4×4 convolution) or a merging layer (batch size is 2, 2×2 convolution) at the front end for spatial downsampling and channel expansion.

[0121] FasterNet is built based on PConv. The global pool in the Fasternet backbone network is replaced by strip pooling, which solves the problem of feature contamination caused by the traditional square pooling kernel. Figure 3 As shown, the strip pooling structure consists of two branches, using H×1 and 1×W pooling kernels respectively, focusing on encoding long-distance context in the horizontal and vertical spatial dimensions to achieve the expansion of the backbone network receptive field, defining a certain position of the feature map as y∈R C×H×W ,in but:

[0122]

[0123] Finally, the output feature map calculation formula is:

[0124] z=S(x,σ(f(y))),

[0125] Where x is the input feature map, S(·) is the numerical multiplication, σ(·) is the sigmoid activation function, and f(·) is the 1×1 convolutional layer.

[0126] Preferably, the structure of the GLSA in step S33 is as follows: Figure 4 As shown in Figure 1, the GLSA module combines global context information and local spatial information, and enhances feature expression capabilities through a global attention mechanism and local convolution branches. Specifically, the GLSA weight adjustment module divides the channels of the input feature map Fi into two groups Fi1 and Fi2, which are respectively sent to the global spatial attention (GSA) and local spatial attention (LSA) modules for processing. The outputs of the two modules are then fused through a 1×1 convolution layer to generate an enhanced feature map. This process is expressed as:

[0127]

[0128] In the GSA module, the global spatial attention mechanism complements the local spatial attention by capturing the long-distance dependencies between pixels. The calculation formula is:

[0129]

[0130] Among them, AttG(·) represents the global attention operation, C (1×1) represents 1×1 convolution, Represents matrix multiplication. MLP (Multi-layer Perceptron) consists of two fully connected layers, which are connected by ReLU activation function and normalization layer. The first layer expands the input to high-dimensional space with an expansion ratio of 2, and the second layer compresses its dimension to be consistent with the original input. In the LSA module, the local spatial attention mechanism further enhances the feature expression ability by extracting local features of the region of interest in the spatial dimension. The calculation formula is:

[0131]

[0132] Among them, F c (·) represents a cascade structure consisting of three 1×1 convolutional layers and one 3×3 depthwise convolutional layer, with the number of channels adjusted to 32. AttL(·) represents the local attention operation, σ(·) is the Sigmoid function, and ⊙ represents point-by-point multiplication. In this way, local spatial information can be effectively aggregated with fewer parameters while maintaining high computational efficiency.

[0133] Preferably, the dual convolutional ELAN neck network in step S34 is optimized by the original C3k2 module of YOLOv11, and the dual convolutional ELAN neck network structure is as follows: Figure 5 As shown in the figure, it adopts the CSP architecture, which divides the feature map into two parts for independent processing. One part of the feature map is directly passed, and the other part stacks multiple dual convolution modules to gradually fuse multi-scale features. The first half directly retains the original information and splices it with the enhanced features to avoid information loss. This not only effectively integrates the feature information of different layers, but also reduces redundant calculations, further alleviating the computational burden of the network.

[0134] The GLBiFPN structure adopts bidirectional cross-scale connections, which can not only perform bottom-up feature fusion (i.e., from low resolution to high resolution), but also top-down fusion (i.e., from high resolution to low resolution). This bidirectional information flow can enhance the communication between features of different scales, so that high-level semantic information and low-level detail information can be effectively integrated.

[0135] Each fusion node in the GLBiFPN structure uses weighted summation to fuse different features, including the same-scale features of the previous layer, the higher-scale features obtained by upsampling, and the lower-scale features obtained by downsampling. This weighted summation method allows the model to adaptively adjust the importance of features from different sources based on the training data. GLBiFPN uses a fast normalized fusion weighting method to add corresponding weights to each input feature. The calculation formula is as follows:

[0136]

[0137] Among them, for each w i Add activation function ReLU to ensure w i >0 to avoid numerical instability.

[0138] Preferably, the loss function F in step S36 loss The calculation formula is as follows:

[0139]

[0140] Among them, i is the index number of the anchor point of the i-th candidate region in the image, p i Represents the target prediction probability of the anchor point; represents the candidate region label, N loc With N cls Respectively represent the classification loss function in the positioning loss L loc With similarity metric loss L cls Normalization coefficient after functionalization, λ represents N loc With N cls The balance weight between loc represents the positioning loss function, L cls Represents the similarity measurement loss function, the positioning loss function L loc The calculation formula is as follows:

[0141] L loc =αL reg +βL CIoU ,

[0142] Among them L reg Represents the position regression loss function, L CIoU is the generalized intersection-over-union loss function, and α and β are hyperparameters used to adjust the functions of each part;

[0143] The similarity metric loss function L cls The calculation formula is as follows:

[0144]

[0145] Among them, x represents the true sample label, and x′ represents the detection output after the activation function.

[0146] Preferably, step S4 includes the following sub-steps:

[0147] S41: Reasonably determine multiple monitoring points and obtain monitoring stream videos T of different transportation angles;

[0148] S42: Extract frames from video T to obtain a monitoring image of the transport equipment belt, and set the current frame image as T t , then the previous frame image is T t-1 ;

[0149] S43: Perform image scale normalization and image standardization on each frame, and perform image normalization on the current frame image T t Perform the preprocessing operation in S2 and then input it into the pre-trained conveyor belt defect detection model to detect the defect type.

[0150] Preferably, step S5 includes the following sub-steps:

[0151] S51: Obtaining a target defect discrimination result of the target transport equipment belt through a transport belt defect detection model;

[0152] S52: If there is a defect, an alarm is issued and the defect type is transmitted to the control console staff;

[0153] S53: Determine whether manual intervention is required for the belt defect based on the target defect identification result at the current moment, and perform maintenance and repair in a timely manner.

[0154] The method of the present invention is based on the single-stage detection framework of YOLOv11. It introduces deep separable convolution through the FasterNet backbone network to compress the computational complexity and optimizes the dynamic label allocation strategy to enhance the discrimination between positive and negative samples. It significantly improves the detection accuracy while maintaining real-time inference performance, and realizes the coordinated optimization of speed and accuracy. Through the innovative design of the GLSA mechanism, global context modeling is used to capture the continuity feature association of long-distance tear cracks, and local spatial attention is used to focus on the pixel-level differences in subtle fracture areas, effectively resisting environmental interference such as coal dust adhesion and water mist occlusion, enhancing the ability to extract significant features of small target defects, and significantly reducing the false detection rate in complex scenarios. In response to the low computing power deployment requirements of underground mines, multi-scale feature fusion technology is constructed through the dual convolutional ELAN neck network to enhance the model's generalization ability for complex morphological defects. Combined with lightweight strategies such as network layer pruning and quantization compression, the model complexity and computing resource requirements are greatly reduced, ensuring the real-time operation of the algorithm on embedded devices and meeting the deployment constraints of underground industrial scenarios.

[0155] It can be understood that the contents of the above method embodiments are all applicable to the present storage medium embodiment, the functions specifically implemented by the present storage medium embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[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 limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A conveyor belt defect detection method based on the GLSA-YOLOv11 model, characterized in that: The steps include: S1: Build target defect dataset D; S2: Preprocess the dataset D; S3: Build a conveyor belt tear and break defect detection model based on GLSA-YOLOv11, including the FasterNet backbone network, the GLSA weight adjustment module, the dual convolutional ELAN neck network, and the prediction layer. Input the preprocessed images into the model for pre-training to obtain the conveyor belt defect detection model. S4: Acquire a video image of the conveyor belt in the monitoring scene, perform the preprocessing operation in S2, and input the processing result into the pre-trained conveyor belt defect detection model; S5: Output the target defect identification results of the target transportation equipment. If there is a defect, an alarm warning will be issued and the defect type will be transmitted to the console staff for timely manual intervention.

2. The method for detecting conveyor belt defects based on the GLSA-YOLOv11 model according to claim 1, wherein step S1 comprises the following sub-steps: S11: Acquire multiple sets of historical defect detection record image data of target transportation equipment, and identify different defect types, thereby constructing a data set D = {d1, d2, ..., d t }; S12: extracting the first image feature information from the plurality of sets of image feature information with defect type identification, wherein the first image feature information corresponds to the first defect type d1, and so on, the tth image feature information corresponds to the tth defect type d t .

3. The method for detecting conveyor belt defects based on GLSA-YOLOv11 according to claim 1, wherein step S2 comprises the following sub-steps: S21: Using the dataset D as the input image f, grayscale erosion, grayscale expansion, and grayscale opening and closing operations are performed; S22: performing de-darkening processing on the image f to obtain a dark channel of the input image f; S23: Obtain unknown prior conditions that need to be estimated in the model based on the dark channel S24: Based on the unknown prior conditions obtained And the current ambient light value A, get the darkened image J(x); S25: Optimize the speed and accuracy of image defect edge extraction, and use a linear transformation preprocessing method to process the image J(x) after the darkening operation.

4. The method for detecting conveyor belt defects based on GLSA-YOLOv11 according to claim 1, wherein step S3 comprises the following sub-steps: S31: Build a conveyor belt tear and break defect detection model based on GLSA-YOLOv11, including the FasterNet backbone network, GLSA weight adjustment module, dual convolutional ELAN neck network and prediction layer; S32: Input the preprocessed image into the FasterNet backbone network and output the feature map; S33: Input the feature map formed by the FasterNet backbone network into the GLSA weight adjustment module, determine the weight vector of the defect feature through the GLSA weight adjustment module, and perform weight distribution to obtain a new feature map; S34: The weighted feature map is input into the dual convolutional ELAN neck network. In the GLBiFPN of the dual convolutional ELAN neck network, the feature map is propagated along the upper and lower structures of the network, and feature fusion is performed at each level to obtain meta-features. S35: Use the multi-scale decoupling head Detect of the prediction layer to output the detection results; S36: Design multi-task classification training loss function F loss .

5. A method for detecting conveyor belt defects based on GLSA-YOLOv11 according to claim 4, wherein the FasterNet backbone network in step S32 is responsible for extracting global image defect feature information to form a feature map, which comprises four layers, each layer having an embedding layer or a merging layer at the front end for spatial downsampling and channel expansion, FasterNet is constructed based on PConv, and the global pooling in the Fasternet backbone network is replaced by strip pooling. The strip pooling structure consists of two branches, using H×1 and 1×W pooling kernels respectively, focusing on encoding long-range context in the horizontal and vertical spatial dimensions to achieve the expansion of the backbone network receptive field, and defining a certain position of the feature map as y∈R C×H×W ,in but: Finally, the output feature map calculation formula is: z=S(x,σ(f(y))), in, x is the input feature map, S(·) is the numerical multiplication, σ(·) is the sigmoid activation function, and f(·) is the 1×1 convolutional layer.

6. The method for detecting conveyor belt defects based on GLSA-YOLOv11 according to claim 4, wherein the GLSA weight adjustment module in step S33 divides the channels of the input feature map Fi into two groups Fi1 and Fi2, and respectively sends them to the global spatial attention module and the local spatial attention module for processing, and then fuses the outputs of the two modules through a 1×1 convolutional layer to generate an enhanced feature map. This process is expressed as: In the GSA module, the global spatial attention mechanism complements the local spatial attention by capturing the long-distance dependencies between pixels. The calculation formula is: in, AttG(·) represents the global attention operation, C (1×1) represents 1×1 convolution, Represents matrix multiplication. MLP consists of two fully connected layers, which are connected by ReLU activation function and normalization layer. The first layer expands the input to high-dimensional space with an expansion ratio of 2, and the second layer compresses its dimension to be consistent with the original input. In the LSA module, the local spatial attention mechanism further enhances the feature expression ability by extracting local features of the region of interest in the spatial dimension. The calculation formula is: Among them, F c (·) denotes a cascade structure consisting of three 1×1 convolutional layers and one 3×3 depthwise convolutional layer, AttL(·) denotes the local attention operation, σ(·) is the Sigmoid function, and ⊙ denotes point-wise multiplication.

7. A method for detecting conveyor belt defects based on GLSA-YOLOv11 as described in claim 4, wherein the dual convolution ELAN neck network in step S34 adopts a CSP architecture, which divides the feature map into two parts for independent processing. One part of the feature map is directly transmitted, and the other part stacks multiple dual convolution Dual Conv modules to gradually fuse multi-scale features. The first half directly retains the original information and is spliced ​​with the enhanced features. The GLBiFPN structure adopts a bidirectional cross-scale connection. This connection method can not only perform bottom-up feature fusion, but also top-down fusion. Each fusion node in the GLBiFPN structure adopts weighted summation to fuse different features, including the same-scale features of the previous layer, higher-scale features obtained by upsampling, and lower-scale features obtained by downsampling. GLBiFPN adopts a fast normalized fusion weighting method to add corresponding weights to each input feature. The calculation formula is as follows: in, For each w i Add activation function ReLU to ensure w i >0 to avoid numerical instability.

8. A method for detecting conveyor belt defects based on GLSA-YOLOv11 according to claim 4, wherein the loss function F in step S36 is loss The calculation formula is as follows: in, i is the index number of the anchor point of the i-th candidate region in the image, p i Represents the target prediction probability of the anchor point; represents the candidate region label, N loc With N cls Respectively represent the classification loss function in the positioning loss L loc With similarity metric loss L cls Normalization coefficient after functionalization, λ represents N loc With N cls The balance weight between loc represents the positioning loss function, L cls Represents the similarity measurement loss function, the positioning loss function L loc The calculation formula is as follows: L loc =αL reg +βL CIoU , Among them L reg Represents the position regression loss function, L CIoU is the generalized intersection-over-union loss function, and α and β are hyperparameters used to adjust the functions of each part; The similarity metric loss function L cls The calculation formula is as follows: Among them, x represents the true sample label, and x′ represents the detection output after the activation function.

9. The method for detecting conveyor belt defects based on GLSA-YOLOv11 according to claim 1, wherein step S4 comprises the following sub-steps: S41: Reasonably determine multiple monitoring points and obtain monitoring stream videos T of different transportation angles; S42: Extract frames from video T to obtain a monitoring image of the transport equipment belt, and set the current frame image as T t , then the previous frame image is T t-1 ; S43: Perform image scale normalization and image standardization on each frame, and perform image normalization on the current frame image T t Perform the preprocessing operation in S2 and then input it into the pre-trained conveyor belt defect detection model to detect the defect type.

10. The method for detecting conveyor belt defects based on GLSA-YOLOv11 according to claim 1, wherein step S5 comprises the following sub-steps: S51: Obtaining a target defect discrimination result of the target transport equipment belt through a transport belt defect detection model; S52: If there is a defect, an alarm is issued and the defect type is transmitted to the control console staff; S53: Determine whether manual intervention is required for the belt defect based on the target defect identification result at the current moment, and perform maintenance and repair in a timely manner.

Images