Foreign matter identification method for trestle belt conveyor

Abstract

The invention belongs to the technical field of coal mine foreign matter detection, and particularly relates to a trestle belt conveyor foreign matter recognition method. The method comprises the following steps: acquiring an original picture set of a coal mine on a conveyor based on historical data; performing data cleaning on data in the original picture set to obtain an original foreign matter picture set; forming a foreign matter amplification picture set; establishing a foreign matter data set of coal mine foreign matters; an FEM-YOLOv8 model fusing an FC2 module, an EMA mechanism and an MPDIOU loss function is constructed; and deploying the YOLOv8 model, the constructed FEM-YOLOv8 model and the obtained foreign matter data set to an operating system, and detecting the coal mine foreign matter on the conveyor. The existing YOLOv8 is improved, so that the parameter quantity and the calculation quantity of the YOLOv8 model are both reduced by more than 20%, and the calculation burden of the model is reduced. The problems that in the prior art, an identification method is large in calculation amount, low in equipment friendliness and low in detection speed, and missing detection and error detection occur when a small target is faced are solved.

Description

Technical Field

[0001] The invention belongs to the technical field of coal mine foreign body detection, and in particular relates to a foreign body identification method for a trestle belt conveyor. Background Art

[0002] Mineral resources are an important material guarantee for the development of the economy and society, and coal mines occupy a dominant position in them. Coal transportation is an important stage in the coal mining process, and belt conveyors are the mainstream equipment for coal transportation. During the coal transportation process, foreign objects mixed in the coal, such as stones, angle irons and bolts dropped by mechanical vibration, not only have an adverse effect on the product quality of the coal mine, but may even cause the conveyor belt to tear or abnormal shutdown, bringing a series of safety hazards and economic losses, thus delaying the production plan.

[0003] In the existing technology, in order to intelligently identify foreign objects included in coal mines, deep learning foreign object detection methods are mostly used. This method has problems such as large computational complexity, low equipment friendliness, slow detection speed, and missed detection and false detection when facing small targets. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of large amount of calculation, low friendliness and slow detection speed in the prior art, and to provide a method for identifying foreign objects in a trestle belt conveyor with small amount of calculation, high confidence and less prone to missed detection or wrong detection problems.

[0005] The technical solution adopted by the present invention to solve the technical problem is: a method for identifying foreign matter on a trestle belt conveyor, comprising the following contents:

[0006] Obtain original picture sets of coal mines on conveyors based on historical data or on-site collection;

[0007] Clean the data in the original picture set to remove foreign matter or fuzzy data, and obtain the original foreign matter picture set;

[0008] Use an enhancement method to perform data enhancement on the original foreign body image set to form a foreign body amplified image set;

[0009] Establish a foreign body data set of foreign bodies in coal mines; that is, use the Labelme tool to manually annotate the image data in the foreign body amplification image set and generate a label set, divide the image set and the label set into a training set, a validation set, and a test set according to a preset ratio to form a foreign body data set;

[0010] Construct a FEM-YOLOv8 model that integrates the FC2 module, EMA mechanism, and MPDIoU loss function. Use the foreign body dataset as the input of the FEM-YOLOv8 model to train foreign body detection on conveyor belts in coal mines and obtain the detection model weights.

[0011] The YOLOv8 model, the constructed FEM-YOLOv8 model and the obtained foreign body dataset are deployed to the operating system. Based on the detection model weights, the coal mine foreign bodies on the conveyor are detected, and the corresponding data are output according to the detection results.

[0012] Specifically, the FEM-YOLOv8 model is formed based on the YOLOv8 model, and specifically includes the following formation process:

[0013] A1. Replace the C2f module in the YOLOv8 model with the FC2 module;

[0014] A2. Introduce the EMA mechanism into the Neck part of the YOLOv8 model;

[0015] A3. Replace the CIoU loss function of the Head part in the YOLOv8 model with the MPDIoU loss function.

[0016] Specifically, the formation of the FC2 module includes the following: replacing the Bottleneck module in the C2f module of the YOLOv8 model with a partial convolution PConv module in FasterNet to form the FC2 module;

[0017] The calculation formula of the partial convolution PConv is:

[0018]

[0019] Among them, h represents the height of the feature map, w represents the width of the feature map, k represents the convolution kernel size of the feature map, and c p Indicates the number of regular convolution channels.

[0020] Specifically, the formation of the EMA mechanism includes the following contents: in the Neck part based on the YOLOv8 model, an EMA module is integrated at the input end of each Head.

[0021] Specifically, the calculation formula of the MPDIoU loss function is:

[0022]

[0023] In the formula, Represents the square of the distance between the upper left corner of the predicted bounding box and the true bounding box, Represents the square of the distance between the lower right corner of the predicted bounding box and the true bounding box, Represents the upper left corner of the predicted bounding box, represents the lower right corner of the predicted bounding box, represents the upper left corner of the true bounding box, represents the lower right corner of the true bounding box, A represents the predicted bounding box, B represents the true bounding box, w represents the width of the minimum bounding rectangle of the predicted bounding box and the true bounding box, and h represents the height of the minimum bounding rectangle of the predicted bounding box and the true bounding box.

[0024] Specifically, the establishment of the foreign body data set of coal mine foreign bodies includes the following contents:

[0025] B1. The pictures in the picture collection are manually labeled with codes and corresponding label names by using the Labelme tool. The codes include the stone code set to 0, and its corresponding label name set to stone; the bolt code set to 1, and its corresponding label name set to bolt; the angle iron code set to 2, and its corresponding label name set to angle iron;

[0026] B2. Generate a corresponding .json file for any completed image, and obtain an image set and a label set consisting of a .json file;

[0027] B3. Divide the image set and label set into training set, validation set and test set according to the preset ratio;

[0028] B4. Read the training set, validation set, and test set through the torch.utils.data function of the Pytorch library to confirm that the dataset can be loaded correctly, that is, it is a foreign dataset.

[0029] Specifically, the preset ratio of the training set, validation set and test set in step B3 is 8:1:1.

[0030] Specifically, the data enhancement of the original foreign body image set uses six data enhancement techniques on the images using the OpenCV tool.

[0031] Specifically, the six data enhancement techniques include the following:

[0032] Rotate the data; that is, keep the image size unchanged and rotate the image content 20°~25° clockwise;

[0033] RGB channel adjustment data: that is, the R, G, and B channels of the original image are increased by 20 to 25 values;

[0034] Scaling data; that is, keeping the image size unchanged, enlarging the image content to 1.2 to 1.3 times the original size;

[0035] Brightness value adjustment data; that is, reducing the brightness value of the image to 65% to 70% of the original;

[0036] Shift the data; that is, keep the image size unchanged, shift the image content 80 to 90 pixels to the right, and 20 to 30 pixels upward;

[0037] Add salt and pepper noise data, that is, add 30% to 35% salt and pepper noise to the image.

[0038] The beneficial effects of the foreign body identification method of a trestle belt conveyor of the present invention are:

[0039] The invention is based on historical image data, obtains the original foreign body image set through data cleaning, adopts the enhancement method to perform data enhancement on the original foreign body image set, constructs an accurate coal mine foreign body detection data set, ensures that the trained model can cope with the detection tasks in complex environments and has strong generalization, and improves the existing YOLOv8 to construct a FEM-YOLOv8 model integrating the FC2 module, the EMA mechanism and the MPDIoU loss function, so that the parameter amount and the calculation amount of the YOLOv8 model are reduced by more than 20%, the calculation burden of the model is reduced, and the speed of foreign body detection is improved. The EMA module is integrated in the neck of YOLOv8, the feature extraction ability of the network is enhanced, the missed detection rate and the false detection rate of the model are reduced, and the minimum point distance is taken as a consideration factor, so as to avoid the limitation of the CIoU loss function in the prior art that the effectiveness of the CIoU loss function is reduced when the real box and the predicted box have similar aspect ratios but different height and width dimensions, and simplify the similarity comparison between the real box and the predicted box. The invention solves the problems existing in the recognition method of the prior art, such as large amount of calculation, low device friendliness, slow detection speed, missed detection and wrong detection when facing small targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0041] Figure 1 It is a flow chart of the steps of the foreign body identification method in an embodiment of the present invention.

[0042] Figure 2 2 is a diagram of the formation process of the FEM-YOLOv8 model in an embodiment of the present invention.

[0043] Figure 3 It is a diagram of the FasterNet network structure in an embodiment of the present invention.

[0044] Figure 4 It is a structural diagram of the EMA mechanism in an embodiment of the present invention.

[0045] Figure 5 It is a structure diagram of MPDIoU in an embodiment of the present invention.

[0046] Figure 6 A structural diagram of the Head portion of the FEM-YOLOv8 model integrating the MPDIoU loss function in an embodiment of the present invention.

[0047] Figure 7 is the original image under normal lighting conditions in the embodiment of the present invention;

[0048] Figure 8 Yes Figure 7 A comparison diagram of existing recognition and recognition in this embodiment.

[0049] Fig. 9 This is the original image in the dust raising environment according to the embodiment of the present invention.

[0050] Fig.10 Yes Fig. 9 A comparison diagram of existing recognition and recognition in this embodiment.

[0051] Fig.11 It is the original image in a weak light environment in the embodiment of the present invention.

[0052] Fig.12 Yes Fig.11 A comparison diagram of existing recognition and recognition in this embodiment. DETAILED DESCRIPTION

[0053] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0054] like Figure 1-Figure 6 The specific embodiment of the method for identifying foreign objects on a trestle belt conveyor of the present invention shown in the figure includes the following contents: obtaining an original picture set of coal mines on a conveyor based on historical data; cleaning the data in the original picture set without foreign objects or fuzzy data to obtain an original foreign object picture set; using an enhancement method to perform data enhancement on the original foreign object picture set to form a foreign object amplified picture set; establishing a foreign object data set of foreign objects in coal mines; that is, using the Labelme tool to manually annotate the picture data in the foreign object amplified picture set and generate a label set, dividing the picture set and the label set into a training set, a verification set and a test set according to a preset ratio to form a foreign object data set; constructing a FEM-YOLOv8 model integrating the FC2 module, the EMA mechanism and the MPDIoU loss function, using the foreign object data set as the input of the FEM-YOLOv8 model, performing coal mine foreign object detection training on a conveyor belt, and obtaining a detection model weight; deploying the YOLOv8 model, the constructed FEM-YOLOv8 model and the obtained foreign object data set to an operating system, detecting coal mine foreign objects on a conveyor based on the detection model weight, and outputting corresponding data according to the detection result.

[0055] The original picture set of coal mines on conveyors is obtained based on historical data. Specifically, pictures of coal mines on conveyor belts with different foreign objects are taken on the spot or a large number of pictures of coal mines on conveyor belts with foreign objects are directly obtained from historical data, which are the original picture set. The original picture set that does not contain foreign objects or blurred data is cleaned, and the picture set with foreign objects is screened out, which is the original foreign object picture set.

[0056] Then, the original foreign body picture set is enhanced by using an enhancement method to form a foreign body amplification picture set. In this embodiment, the data enhancement of the original foreign body picture set uses six data enhancement techniques for the pictures using the OpenCV tool, wherein the six data enhancement techniques include the following: rotation data, RGB channel adjustment data, scaling data, brightness value adjustment data, translation data, and adding salt and pepper noise data. The six parts are enhanced, wherein the rotation data, that is, keeping the picture size unchanged, rotates the picture content clockwise by 20° to 25°. The RGB channel adjustment data, that is, increases the R, G, and B channels of the original image by 20 to 25 values. The scaling data, that is, keeping the picture size unchanged, enlarges the picture content to 1.2 to 1.3 times the original. The brightness value adjustment data, that is, reduces the brightness value of the picture to 65% to 70% of the original. The translation data, that is, keeping the picture size unchanged, translates the picture content to the right by 80 to 90 pixels and translates it upward by 20 to 30 pixels. Add salt and pepper noise data, that is, add 30% to 35% salt and pepper noise to the image. Through six data enhancement methods, a new coal mine foreign body detection dataset on conveyor belts is constructed, making the trained model more generalizable and able to cope with detection tasks in complex environments.

[0057] Then, a foreign body data set of coal mine foreign bodies is established, and the establishment of the foreign body data set of coal mine foreign bodies includes the following contents:

[0058] B1. Use the Labelme tool to manually label the pictures in the picture collection one by one with codes and corresponding label names, wherein the codes include the stone code set to 0, and its corresponding label name set to stone; the bolt code set to 1, and its corresponding label name set to bolt; the angle iron code set to 2, and its corresponding label name set to angle.iron;

[0059] B2. Generate a corresponding .json file for any completed image, and obtain an image set and a label set consisting of a .json file;

[0060] B3. Divide the image set and the label set into a training set, a validation set, and a test set according to a preset ratio; wherein the preset ratio of the training set, the validation set, and the test set is 8:1:1;

[0061] B4. Read the training set, validation set, and test set through the torch.utils.data function of the Pytorch library to confirm that the dataset can be loaded correctly, that is, it is a foreign dataset.

[0062] like Figure 2 As shown, the FEM-YOLOv8 model described in this embodiment is formed based on the YOLOv8 model, and specifically includes the following formation process:

[0063] A1. Replace the C2f module in the YOLOv8 model with the FC2 module;

[0064] A2. Introduce the EMA mechanism into the Neck part of the YOLOv8 model;

[0065] A3. Replace the CIoU loss function of the Head part in the YOLOv8 model with the MPDIoU loss function.

[0066] As an implementation method, the formation of the FC2 module in this embodiment includes the following: replacing the Bottleneck module in the C2f module of the YOLOv8 model with a partial convolution PConv module in FasterNet to form the FC2 module. Figure 3 This is the FasterNet network structure diagram in this embodiment.

[0067] The calculation formula of the partial convolution PConv is:

[0068]

[0069] Among them, h represents the height of the feature map, w represents the width of the feature map, k represents the convolution kernel size of the feature map, and c p Indicates the number of regular convolution channels.

[0070] like Figure 4As shown in Figure 1, the formation of the EMA mechanism includes the following: Based on the Neck part of the YOLOv8 model, an EMA module is integrated at the input end of each Head. Specifically, the input feature map is divided into multiple groups and then transmitted to three parallel sub-networks to extract the attention weight descriptor. The two branches use 1×1 convolution kernels for global average pooling to extract global information in different spatial directions, and then the output feature vectors are spliced ​​into one branch. Finally, two independent attention weight maps are obtained through the Sigmoid function to complete the feature recalibration. In the third branch, each feature group only needs a 3×3 convolution kernel to capture multi-scale feature representation. Cross-spatial information fusion is the second step of EMA. It extracts features from all output information from the first step through normalization and average pooling operations, and then performs a dot multiplication operation on the feature map obtained in the first step to form the final spatial attention map. Finally, the Sigmoid function is used to further adjust the fused attention map and output the result. The EMA module is integrated in the YOLOv8 neck (Neck), which enhances the feature extraction capability of the network and reduces the missed detection rate and false detection rate of the model.

[0071] See also Figure 5 and Figure 6 Finally, the MPDIoU loss function is used to replace the CIoU loss function of the Head part in the original YOLOv8 model to construct a new Detect, where the calculation formula of the MPDIoU loss function is:

[0072]

[0073] In the formula, Represents the square of the distance between the upper left corner of the predicted bounding box and the true bounding box, Represents the square of the distance between the lower right corner of the predicted bounding box and the true bounding box, Represents the upper left corner of the predicted bounding box, represents the lower right corner of the predicted bounding box, represents the upper left corner of the true bounding box, represents the lower right corner of the true bounding box, A represents the predicted bounding box, B represents the true bounding box, w represents the width of the minimum bounding rectangle of the predicted bounding box and the true bounding box, and h represents the height of the minimum bounding rectangle of the predicted bounding box and the true bounding box.

[0074] The MPDIoU loss function is used to replace the original CIoU loss function, and a new Detect is constructed, taking the minimum point distance as a consideration. This solves the limitation of the CIoU loss function that the effectiveness of the CIoU loss function is reduced when the true box and the predicted box have similar aspect ratios but different height and width dimensions, and simplifies the similarity comparison between the true box and the predicted box.

[0075] The invention is based on historical image data, obtains the original foreign body image set through data cleaning, adopts the enhancement method to perform data enhancement on the original foreign body image set, constructs an accurate coal mine foreign body detection data set, ensures that the trained model can cope with the detection tasks in complex environments and has strong generalization, and improves the existing YOLOv8 to construct a FEM-YOLOv8 model integrating the FC2 module, the EMA mechanism and the MPDIoU loss function, so that the parameter amount and the calculation amount of the YOLOv8 model are reduced by 25.2% and 21.4% respectively, reducing the calculation burden of the model, and improving the speed of foreign body detection. The EMA module is integrated in the neck of YOLOv8, which enhances the feature extraction ability of the network, reduces the missed detection rate and the false detection rate of the model, takes the minimum point distance as a consideration factor, avoids the limitation of the CIoU loss function in the prior art that the effectiveness of the CIoU loss function is reduced when the real box and the predicted box have similar aspect ratios but different height and width dimensions, and simplifies the similarity comparison between the real box and the predicted box. The invention solves the problems existing in the recognition method of the prior art, such as large amount of calculation, low device friendliness, slow detection speed, missed detection and wrong detection when facing small targets.

[0076] The specific identification process of the identification method in the above embodiment is used to conduct a comparative experiment with the existing method. It should be understood that although only one identification process is listed here, the screened ones are universal. The details are as follows:

[0077] An original picture set is established by taking pictures of coal mines on conveyor belts with foreign objects on site. The original resolution of the pictures is 416*416. Data cleaning in the first embodiment is to manually check the original picture set, delete the pictures that do not contain foreign objects or are blurred for data cleaning, and finally obtain 3368 pictures.

[0078] Data enhancement was performed on 3368 images for two purposes: first, to expand the size of the data set to avoid overfitting when training the model; second, a complex data set can enhance the generalization and robustness of the model. The so-called data enhancement specifically uses the OpenCV tool to apply six data enhancement techniques to the images: rotate the data, keep the image size unchanged, and rotate the image content clockwise by 20°; adjust the RGB channel data, increase the R, G, and B channels of the original image by 20 values, where R represents the red channel, G represents the green channel, and B represents the blue channel, and the value range of the three channels is 0 to 255; scale the data, keep the image size unchanged, and enlarge the image content to 1.3 times the original; adjust the brightness value data, reduce the brightness value of the image to 70% of the original; translate the data, keep the image size unchanged, translate the image content 80 pixels to the right and 20 pixels upward; add salt and pepper noise data, and add 30% salt and pepper noise to the image. It is further explained that, in addition to determining the color itself, the RGB channel adjustment data in this embodiment, the values ​​of the R, G, and B channels also affect the saturation and brightness of the color. When the values ​​of the three channels are high, the color will appear bright and saturated; when the values ​​of one or several channels are low, the color will appear dim or unsaturated. Other adjustments can be made according to specific needs, which will not be elaborated here.

[0079] Use the Labelme tool to manually label each image in the image collection, as follows: Figure 2 As shown in the figure, the stone code is 0 and the label name is stone; the bolt code is 1 and the label name is bolt; the angle iron code is 2 and the label name is angle iron. Each time a picture is annotated, a corresponding .json file will be generated. After all pictures are annotated, a picture set and a label set consisting of .json files are finally obtained. The picture set and the label set are imported into the data set partitioning program, and the ratio of the training set, the validation set and the test set is set to 8:1:1. After running the program, the final coal mine foreign body data set on the conveyor belt is generated. The foreign body data set is read through the torch.utils.data function of the Pytorch library to confirm that the data set can be loaded correctly.

[0080] Use Fasternet's PConv technology to optimize YOLOv8's C2f module. FasterNet is an inverted residual structure with a 3×3 partial convolution (PConv) in the first layer, and two layers of 1×1 convolution kernels of point-by-point convolution (PWConv) are attached after PConv. Then, batch normalization (BN layer) and activation layer (ReLu) are added between the two point-by-point convolution layers (PWConv) to maintain feature diversity and reduce latency. Finally, feature conversion and classification are achieved through a 1×1 point-by-point convolution layer (PWConv), a global average pooling layer, and a fully connected layer. The calculation formula for partial convolution PConv is:

[0081]

[0082] Among them, the calculation amount of PConv is h represents the height of the feature map, w represents the width of the feature map, k represents the convolution kernel size of the feature map, and c p Represents the number of channels of a regular convolution. The memory access of PConv is about 1 / 4 of that of a regular convolution, so its computational effort is about 1 / 16 of that of a regular convolution. When taking 1 / 4 of the total number of input channels, the amount of calculation is reduced to 1 / 16 of the standard convolution.

[0083] In the Neck part of YOLOv8, an EMA module is integrated at the input of each Head. The EMA module first divides the input feature map into multiple groups, and then transmits it to three parallel sub-networks to extract the attention weight descriptor. The two branches use 1×1 convolution kernels for global average pooling to extract global information in different spatial directions, and then splice the output feature vectors into one branch. Finally, two independent attention weight maps are obtained through the Sigmoid function to complete the feature recalibration. In the third branch, each feature group only needs a 3×3 convolution kernel to capture multi-scale feature representation. Cross-spatial information fusion is the second step of EMA. It extracts features from all the output information from the first step through normalization and average pooling operations, and then performs a dot multiplication operation on the feature map obtained in the first step to form the final spatial attention map. Finally, the Sigmoid function is used to further adjust the fused attention map and output the result. The CIoU loss function in the YOLOv8 Head part is replaced by the MPDIoU loss function. The MPDIoU loss function calculates IoU by minimizing the key point distance between the predicted box and the real box, which solves the problem that the original loss function cannot be effectively optimized when the predicted bounding box and the actual bounding box have the same aspect ratio but different height values, and simplifies the calculation process. The calculation formula is:

[0084]

[0085] In the formula, Represents the square of the distance between the upper left corner of the predicted bounding box and the true bounding box, Represents the square of the distance between the lower right corner of the predicted bounding box and the true bounding box, Represents the upper left corner of the predicted bounding box, represents the lower right corner of the predicted bounding box, represents the upper left corner of the true bounding box, represents the lower right corner of the true bounding box, A represents the predicted bounding box, B represents the true bounding box, w represents the width of the minimum bounding rectangle of the predicted bounding box and the true bounding box, and h represents the height of the minimum bounding rectangle of the predicted bounding box and the true bounding box.

[0086] The YOLOv8 model, the improved YOLOv8 model and the constructed foreign body data set are deployed to the operating system Ubuntu18.04. The CPU model in this specific implementation process is Intel Xeon(R) Silver4210R, and the GPU model is GeForceRTX4090. On a computer with 128G memory, the foreign body data set is read through the torch.uti l s.data function of the Pytorch library, and the training command is enabled to start the training of foreign body detection. After the training is completed, the optimal weights of the YOLOv8 model and the improved YOLOv8 model are saved to the specified folder, and the foreign body detection code is written. The optimal weights and the foreign body images to be detected are imported through the model = YOLO() command and the img = cv2.imread() command, respectively. The program run button is started to output the detection results. The FEM-YOLOv8 model is now compared with the existing YOLOv8 model, and the four indicators of parameter quantity, calculation amount, mAP@0.5, and FPS used in the experiment are evaluated. The specific results are shown in Table 1 below.

[0087] Table 1:

[0088] Model Params / M FLOPs / G mAP@0.5 / % FPS(GPU) YOLOv8 11.1 28.7 88.9 91 FEM-YOLOv8 8.4 22.7 90.6 119

[0089] According to the comparison results, the FEM-YOLOv8 model has significantly reduced the number of parameters and the amount of calculation compared with the YOLOv8 model, which is reduced by 25.2% and 21.4% respectively, which helps to reduce the deployment requirements of the model and speed up the reasoning speed. In terms of detection accuracy, the FEM-YOLOv8 model is 1.7% higher than the YOLOv8 model. This shows that the foreign body identification method of the trestle belt conveyor of the present invention has significantly improved in reducing the amount of calculation.

[0090] See also Figures 7 to 12, which are visualization diagrams corresponding to the above-mentioned multiple detection results. It can be clearly seen from the three comparison diagrams that the recognition method of the FED-YOLOv8 module of the present invention shows a higher confidence score. The recognition method of the present invention is more accurate, and the existing YOLOv8 module recognition method is prone to false detection, which is specifically manifested as follows: Figure 8 The existing comparison in the yellow circle identifies that there is a foreign object, but there is actually no foreign object there. However, the recognition method of the present invention does not have the problem of false detection. Fig.10 It can be clearly seen from the values ​​in the red circle that the recognition method of the present invention exhibits excellent positioning accuracy. Fig.12 ,The images identified by the recognition method of the present invention show higher detection performance, that is, it has better robustness and effective detection performance than the prior art.

[0091] It should be understood that the specific embodiments described above are only used to explain the present invention, and are not used to limit the present invention. Obvious changes or modifications derived from the spirit of the present invention are still within the protection scope of the present invention.

Claims

1. A method for identifying foreign matter on a trestle belt conveyor, characterized in that: Includes the following: Get the original picture set of coal mine on conveyor based on historical data or field collection; Clean the data in the original picture set to remove foreign matter or fuzzy data, and obtain the original foreign matter picture set; Use an enhancement method to perform data enhancement on the original foreign body image set to form a foreign body amplified image set; Establish a foreign body data set of foreign bodies in coal mines; that is, use the Labelme tool to manually annotate the image data in the foreign body amplification image set and generate a label set, divide the image set and the label set into a training set, a validation set, and a test set according to a preset ratio to form a foreign body data set; Construct a FEM-YOLOv8 model that integrates the FC2 module, EMA mechanism, and MPDIoU loss function. Use the foreign body dataset as the input of the FEM-YOLOv8 model to train foreign body detection on conveyor belts in coal mines and obtain the detection model weights. The YOLOv8 model, the constructed FEM-YOLOv8 model and the obtained foreign body dataset are deployed to the operating system. Based on the detection model weights, the coal mine foreign bodies on the conveyor are detected, and the corresponding data are output according to the detection results.

2. A method for identifying foreign matter on a trestle belt conveyor according to claim 1, characterized in that: The FEM-YOLOv8 model is formed based on the YOLOv8 model, and specifically includes the following formation process: A1. Replace the C2f module in the YOLOv8 model with the FC2 module; A2. Introduce the EMA mechanism into the Neck part of the YOLOv8 model; A3. Replace the CIoU loss function of the Head part in the YOLOv8 model with the MPDIoU loss function.

3. A method for identifying foreign matter in a trestle belt conveyor according to claim 2, characterized in that: The formation of the FC2 module includes the following: replacing the Bottleneck module in the C2f module of the YOLOv8 model with a partial convolution PConv module in FasterNet to form the FC2 module; The calculation formula of the partial convolution PConv is: Among them, h represents the height of the feature map, w represents the width of the feature map, k represents the convolution kernel size of the feature map, and c p Represents the number of regular convolution channels.

4. A method for identifying foreign matter in a trestle belt conveyor according to claim 2, characterized in that: The formation of the EMA mechanism includes the following contents: in the Neck part based on the YOLOv8 model, an EMA module is integrated at the input end of each Head.

5. The method for identifying foreign matter in a trestle belt conveyor according to claim 2, characterized in that: The calculation formula of the MPDIoU loss function is: In the formula, Represents the square of the distance between the upper left corner of the predicted bounding box and the true bounding box, Represents the square of the distance between the lower right corner of the predicted bounding box and the true bounding box, Represents the upper left corner of the predicted bounding box, represents the lower right corner of the predicted bounding box, represents the upper left corner of the true bounding box, represents the lower right corner of the true bounding box, A represents the predicted bounding box, B represents the true bounding box, w represents the width of the minimum bounding rectangle of the predicted bounding box and the true bounding box, and h represents the height of the minimum bounding rectangle of the predicted bounding box and the true bounding box.

6. A method for identifying foreign matter on a trestle belt conveyor according to claim 1, characterized in that: The establishment of the foreign body data set of coal mine foreign bodies includes the following contents: B1. Use the Labelme tool to manually label the pictures in the picture collection one by one with codes and corresponding label names, wherein the codes include the stone code set to 0, and its corresponding label name set to stone; the bolt code set to 1, and its corresponding label name set to bolt; the angle iron code set to 2, and its corresponding label name set to angle.iron; B2. Generate a corresponding .json file for any completed image, and obtain an image set and a label set consisting of .json files; B3. Divide the image set and label set into training set, validation set and test set according to the preset ratio; B4. Read the training set, validation set, and test set through the torch.utils.data function of the Pytorch library to confirm that the dataset can be loaded correctly, that is, it is a foreign dataset.

7. A method for identifying foreign matter in a trestle belt conveyor according to claim 6, characterized in that: The preset ratio of the training set, validation set and test set in step B3 is 8:1:

1.

8. A method for identifying foreign matter on a trestle belt conveyor according to any one of claims 1 to 6, characterized in that: The data enhancement of the original foreign body image set is performed using the OpenCV tool and six data enhancement techniques are applied to the images.

9. A method for identifying foreign matter on a trestle belt conveyor according to claim 8, characterized in that: The six data enhancement techniques include the following: Rotate the data; that is, keep the image size unchanged and rotate the image content 20°~25° clockwise; RGB channel adjustment data: that is, the R, G, and B channels of the original image are increased by 20 to 25 values; Scaling data; that is, keeping the image size unchanged, enlarging the image content to 1.2 to 1.3 times the original size; Brightness value adjustment data; that is, reducing the brightness value of the image to 65% to 70% of the original; Shift the data; that is, keep the image size unchanged, shift the image content 80 to 90 pixels to the right, and 20 to 30 pixels upward; Add salt and pepper noise data, that is, add 30% to 35% salt and pepper noise to the image.

Images