Belt deviation detection method based on deep learning

Abstract

The invention belongs to the technical field of detection methods, and particularly relates to a belt deviation detection method based on deep learning, which comprises the following steps of: accessing an algorithm into a real-time video stream for processing, and processing each frame of video image by the algorithm to obtain whether a belt deviation image exists in a current frame or not; if deviation occurs within continuous 10 frames, belt deviation alarming is carried out; the deviation rate K of the belt is obtained through calculation; the network is modified based on a traditional UNet network, and the network finally outputs a 572 * 572 * 4 feature. According to the method, corresponding treatment is conducted on the belt with the supporting roller and the belt without the supporting roller, and universality and integrity of the method are guaranteed. According to the method, efficient monitoring and early warning of the belt deviation behavior can be achieved, and a more reliable and safer guarantee is provided for enterprise production.

Description

Technical Field

[0001] The present invention belongs to the technical field of detection methods, and specifically relates to a belt deviation detection method based on deep learning. Background Art

[0002] With the continuous improvement of industrial automation, belt conveyors play an increasingly important role in the field of material transportation. However, with the increase of service time and wear, belt conveyors often have belt deviation problems during operation. Belt deviation not only affects the normal operation of the conveyor, but may also cause equipment failure and safety hazards, and even cause production accidents. Therefore, studying and realizing the detection and prevention of belt deviation has become an important research direction.

[0003] Traditional belt deviation detection methods mainly rely on manual inspections and regular maintenance. This method has problems such as limited monitoring coverage, poor real-time performance, reliance on manpower, and easy to miss detection. In recent years, with the development of machine vision and intelligent algorithms, some automatic monitoring methods based on sensor technology and artificial intelligence have gradually been introduced. These methods monitor the running status of the belt by installing sensors, and combine computer vision and data processing algorithms to achieve real-time detection and early warning of belt deviation, thereby improving monitoring efficiency and accuracy. However, sensor-based methods have high hardware costs, complex installation, and high subsequent maintenance costs. Artificial intelligence-based methods have problems such as complex models, the need for multi-model coordination, high reliance on prior knowledge, low recognition accuracy, insufficient reliability, high cost, and weak environmental adaptability.

[0004] Patent CN 115924455 A uses the YOLOv5 model and adds a segmentation head to the head part to perform target detection and semantic segmentation at the same time, which complicates the network structure of the entire task. At the same time, this method is heavily dependent on the roller and cannot solve the problem of belt deviation without the roller.

[0005] Patent CN 114313883 B requires pre-definition of standard lines, which adds pre-processing operations and deployment operations. At the same time, the method is implemented based on traditional image processing methods and has poor adaptability.

[0006] Patent CN 117699375 A realizes the calculation of belt angle based on optical flow and the detection of knot roller based on detection method. This method requires the combination of the two methods and can only identify deviation based on video, but cannot do anything for single-frame images.

[0007] Patent CN 117566378 A obtains the mask of the belt based on the segmentation model, and then obtains the edge points of the belt based on post-processing operations, and then compares them with the pre-set standard edge points to determine whether the belt is offset. Although the entire process is simple to operate, the algorithm has poor adaptability and requires deployment personnel to pre-calibrate different standard edge points for different cameras. At the same time, this method is also powerless to deal with changes in standard edge points caused by camera deflection, jitter and other problems.

[0008] The method of patent CN 115908273 A relies heavily on a pre-built belt deviation dataset, and compares the belt status in the real-time video with the belt status in the database to determine whether the belt is offset. This method requires building an independent database for each camera, has poor adaptability, and has high initial deployment costs.

[0009] Patent CN 117689651 A requires the use of a detection model and a segmentation model at the same time, which increases complexity and computing resource consumption, and makes it difficult to achieve real-time performance.

[0010] Patent CN 117985425 A relies on traditional Hough transform to obtain the belt edge line. The method has poor adaptability and many false detections. At the same time, the method relies on a predefined belt edge reference line and is greatly affected by the deflection and jitter of the camera. Summary of the invention

[0011] In view of the technical problems that the above-mentioned artificial intelligence-based methods have complex models, require multi-model coordination, high reliance on prior knowledge, low recognition accuracy, insufficient reliability, high cost and weak environmental adaptability, the present invention provides a belt deviation detection method based on deep learning, which can achieve efficient monitoring and early warning of belt deviation behavior, and provide more reliable and safe protection for enterprise production.

[0012] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0013] A belt deviation detection method based on deep learning comprises the following steps:

[0014] S1. The algorithm accesses the real-time video stream for processing. Each frame of the video image is processed by the algorithm to determine whether there is belt deviation in the current frame;

[0015] S2: If the deviation occurs within 10 consecutive frames, the belt deviation alarm will be issued;

[0016] S3, calculating the belt deviation rate K;

[0017] S4. The network is modified based on the traditional UNet network, and the final output of the network is 572*572*4 features.

[0018] The method for processing each frame of video image in S1 is as follows: The video frame image is processed by the UNet model to obtain the segmentation result of the current frame, including the left idler pulley segmentation feature map, the right idler pulley segmentation feature map, and the belt segmentation feature map.

[0019] The calculation of the belt running deviation rate K in S3 includes processing belts with left and right idler pulleys; processing belts without idler pulleys; processing belts with only one side idler pulley; processing belts with only one idler pulley on one side.

[0020] The method for processing belts with left and right idler pulleys is as follows:

[0021] Based on the belt segmentation feature map, the center point coordinates P of the belt are obtained according to the findContours method in opencv center =(x1,y1); Based on the left idler pulley segmentation feature map and the right idler pulley segmentation feature map, the center points of multiple left and right idler pulleys are obtained using the same method, and are respectively represented as P left_i =(x1_i,y1_i), P right_j =(x1_j,y1_j), where 1 < i < m, 1 < j < n, and m and n respectively represent the numbers of left and right idler pulleys;

[0022] Define the left idler pulley straight line equation as y = kx + b, and use the least squares method to fit the parameters k and b based on the points of multiple left idler pulleys to finally obtain the straight line equation of the left idler pulley; here, the L1 distance is selected as the distance in the least squares fitting process to better avoid the interference of outlier sample points; similarly, the straight line equation of the right idler pulley is obtained;

[0023] Based on the intersection equation of two straight lines, find the intersection of the left and right idler pulley straight lines to obtain the intersection point P cross =(x,y);

[0024] Based on the belt center point P center and the left and right idler pulley intersection point P cross the center line equation of the idler pulley can be calculated;

[0025] Based on the point-to-point distance formula, calculate the average distance D from the belt center point P center to the center points of the left and right idler pulleys left, D right ; D left =mean(dist(P center ,P left )), D right =mean(dist(P center ,P right)), where dist represents the distance between two points and mean represents the average;

[0026] Based on the distance D between the left and right support rollers left, D right According to the formula, the offset rate K = scale*(D left -D right ) / min(D left, D right ) calculates the belt deviation rate of the final roller image; where scale represents the scaling factor, which is 1.2 here. A negative deviation rate indicates that the belt is deviating to the left, and a positive deviation rate indicates that the belt is deviating to the right.

[0027] The method for processing a belt without a support roller is as follows:

[0028] Based on the belt segmentation feature map, the center point coordinate P of the belt is obtained according to the findContours method in opencv. center =(x1,y1);

[0029] First, generate a vertical line through the center point of the belt to divide the belt into two parts. Flip the left half of the belt axially symmetrically based on the center line and calculate the IOU of the two halves of the belt.

[0030] Similarly, rotate the center line of the belt by 0-179 degrees, and calculate the IOU according to the above method, and finally get 180 IOU values; calculate the maximum value of these IOU values, and the rotation angle corresponding to this value is the center line angle of the belt, so as to obtain the slope k of the belt;

[0031] The final offset rate is obtained according to the offset rate formula of the unsupported belt: K=1.0-IOU.

[0032] The method for processing a belt with only rollers on one side is as follows: when only one side of the rollers is detected, it can be directly determined that the belt is seriously deviated, and the belt deviation rate is directly defined as 100%; the number of rollers on both sides is determined, if the number of rollers on the left side is 0, the belt is seriously deviated to the left; if the number of rollers on the right side is 0, the belt is seriously deviated to the right.

[0033] The method for processing a belt with only one roller on one side is as follows: when only one roller is detected on one side of the belt, it indicates that the belt is seriously deviated, and the belt deviation rate is directly defined as 100%; the number of rollers on both sides is judged, if the number of rollers on the left side is 1, the belt is seriously deviated to the left; if the number of rollers on the right side is 1, the belt is seriously deviated to the right.

[0034] The method for modifying the network in S4 based on the traditional UNet network is:

[0035] The input image size of the network is 572*572*3. It is first processed by the inc module, and then processed by three downsampling modules (down1, down2, down3). Each time it is downsampled, the feature resolution is halved and the feature channel is doubled; then it is processed by three upsampling modules (up1, up2, up3). Each time it passes through an upsampling module, the feature resolution is doubled and the feature channel is halved; finally, the features are output by the outc module; the inc operation is Conv+BN+Relu+Conv+BN+Relu, the outc operation is Conv+Softmax operation; the downsampling operation is Max pooling+Conv+BN+Relu+Conv+BN+Relu, the upsampling operation up is Upsample+Conv+BN+Relu+Conv+BN+Relu+Concat, and the feature fusion fuse operation is Conv1*1+Conv3*3+Conv1*1+Conv3*3+Conv1*1+Conv3*3. Assuming that the input feature of each operation is x, and the input features of the up operation are x1,x2, the corresponding operations are expressed as follows:

[0036] inc=Relu(BN(Conv(Relu(BN(Conv(x))))))

[0037] outc = Softmax(Conv(X))

[0038] down=Relu(BN(Conv(Relu(BN(Conv(Max pooling(x)))))))

[0039] up=concat(Relu(BN(Conv(Relu(BN(Conv(Upsample(x1))))))),x2)

[0040] x1_1=Conv1*1(x)+Conv3*3(x)

[0041] x2_1=Conv1*1(x1_1)+Conv3*3(x1_1)

[0042] x3_1=Conv1*1(x2_1)+Conv3*3(x2_1)

[0043] fuse=Conv1*1(x3_1)+Conv3*3(x3_1)

[0044] The final output of the network is 572*572*4 features, including belt segmentation feature map, left roller segmentation feature map, and right roller segmentation feature map.

[0045] The absolute value of the deviation rate K is within a normal acceptable range of 0-20%, 20%-50% is a slight deviation, and 50%-100% is a serious deviation.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] The present invention performs corresponding processing for both the belt with support rollers and the belt without support rollers, ensuring the universality and integrity of the method. For the belt with support rollers, the belt and support rollers are identified by the UNet model, and the center point of the belt and the center point of the support roller are obtained based on the segmentation results based on traditional image processing. The left support roller equation is obtained based on the least squares method based on the multiple center points of the support rollers on the left, and the right support roller equation is obtained similarly. The intersection of the left and right support roller equations is obtained, and the belt centerline equation is obtained based on the intersection and the belt center point. Finally, the distance from the center point of the belt to the left and right support roller straight lines is calculated, and then the final belt offset rate is obtained based on the predefined belt offset rate formula. For the belt without support rollers, the center point of the belt is also obtained based on traditional image processing, and then the center line slope and IOU value of the belt are obtained based on the method of rotating the center line, and finally the belt offset rate is obtained based on the belt offset rate equation without support rollers. The present invention can realize efficient monitoring and early warning of belt deviation behavior, providing more reliable and safe protection for enterprise production. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the implementation methods or the description of the prior art. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.

[0049] The structures, proportions, sizes, etc. illustrated in this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with the technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantial technical significance. Any structural modification, change in proportion or adjustment of size shall still fall within the scope of the technical contents disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention.

[0050] Figure 1 is a flow chart of the method of the present invention;

[0051] Figure 2 is a network structure diagram of the present invention;

[0052] Figure 3 The effect diagram of the present invention having left and right supporting rollers;

[0053] Figure 4 This is a rendering of the present invention without a support roller;

[0054] Figure 5 It is the belt segmentation characteristic diagram of the present invention;

[0055] Figure 6 This is a rendering of the present invention with only one-side support roller;

[0056] Figure 7 The figure is a rendering of the present invention in which one side of the support roller is only one. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. These descriptions are only to further illustrate the features and advantages of the present invention, rather than to limit the claims of the present invention. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0058] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0059] In this embodiment, if Figure 1 As shown in the figure, the algorithm accesses the real-time video stream for processing. Each frame of the video image is processed by the algorithm to determine whether there is belt deviation in the current frame. If deviation occurs within 10 consecutive frames, a belt deviation alarm is issued.

[0060] The processing flow for each frame of video image is as follows: First, the video frame image is processed by the UNet model to obtain the segmentation result of the current frame, including the left support roller segmentation feature map, the right support roller segmentation feature map, and the belt segmentation feature map.

[0061] The processing flow for belts with left and right rollers is as follows:

[0062] Based on the belt segmentation feature map, the center point coordinate P of the belt is obtained according to the findContours method in opencv. center =(x1,y1). Based on the segmentation feature map of the left support roller and the segmentation feature map of the right support roller, the center points of the left and right support rollers are obtained using the same method, which are represented as P left_i =(x1_i,y1_i),

[0063] P right_j=(x1_j, y1_j), where 1 < i < m, 1 < j < n, and m and n represent the numbers of the left and right supporting rollers respectively.

[0064] Define the straight line equation of the left supporting roller as y = kx + b. Based on the points of multiple left supporting rollers, the parameters k and b are fitted by the least squares method, and finally the straight line equation of the left supporting roller is obtained. Here, the L1 distance is selected as the distance in the least squares fitting process, which can better avoid the interference of outlier sample points. Similarly, the straight line equation of the right supporting roller is obtained.

[0065] Based on the intersection equation of two straight lines, find the intersection of the left and right supporting roller straight lines to obtain the intersection point P cross =(x, y).

[0066] Based on the center point P of the belt center and the intersection point P of the left and right supporting rollers cross the center line equation of the supporting roller can be calculated.

[0067] Based on the distance formula from point to point, calculate the average distance D from the center point P of the belt center to the center points of the left and right supporting rollers left, D right . D left = mean(dist(P center , P left ))), D right = mean(dist(P center , P right ))), where dist represents the distance between two points and mean represents taking the average.

[0068] Based on the distance D between the left and right supporting rollers left, D right , according to the formula offset rate K = scale * (D left - D right ) / min(D left, D right ) calculate the belt offset rate of the final supporting roller image. Among them, scale represents the scaling factor, which takes the value of 1.2 here. A negative offset rate indicates that the belt is deflected to the left, and a positive offset rate indicates that the belt is deflected to the right.

[0069] For the processing flow of the belt without supporting rollers is as follows:

[0070] Based on the belt segmentation feature map, obtain the center point coordinates P of the belt according to the findContours method in opencv center =(x1, y1).

[0071] First, generate a vertical line through the center point of the belt to divide the belt into two parts. Flip the left half of the belt axially symmetrically based on the center line and calculate the IOU of the two halves of the belt.

[0072] Similarly, rotate the centerline of the belt by 0-179 degrees, and calculate the IOU according to the above method, and finally get 180 IOU values. Calculate the maximum value of these IOU values, and the rotation angle corresponding to this value is the centerline angle of the belt, so as to obtain the slope k of the belt.

[0073] The final offset rate is obtained according to the offset rate formula of the unsupported belt: K=1.0-IOU.

[0074] The processing flow for a belt with only one side of the roller is as follows:

[0075] When only one side of the roller is detected, it can be directly judged that the belt is seriously deviated, and the belt deviation rate is directly defined as 100%. The number of rollers on both sides is determined. If the number of rollers on the left side is 0, the belt is seriously deviated to the left; if the number of rollers on the right side is 0, the belt is seriously deviated to the right.

[0076] The processing flow for a belt with only one roller on one side is as follows:

[0077] When only one roller is detected on one side of the belt, it means that the belt is seriously deviated, and the belt deviation rate is directly defined as 100%. The number of rollers on both sides is determined. If the number of rollers on the left side is 1, the belt is seriously deviated to the left; if the number of rollers on the right side is 1, the belt is seriously deviated to the right.

[0078] Finally, the belt deviation rate K is obtained. The absolute value of the belt deviation rate is defined as 0-20% as the normal acceptable range, 20%-50% as a slight deviation, and 50%-100% as a serious deviation.

[0079] like Figure 2As shown in the figure, the network is modified based on the traditional UNet network. The input image size of the network is 572*572*3. It is first processed by the inc module, and then processed by 3 downsampling modules (down1, down2, down3). After each downsampling, the feature resolution is halved and the feature channel is doubled. Then it is processed by 3 upsampling modules (up1, up2, up3). After each upsampling module, the feature resolution is doubled and the feature channel is halved. Finally, the features are output through the outc module. The inc operation is Conv+BN+Relu+Conv+BN+Relu, and the outc operation is Conv+Softmax operation. The down sampling operation is Max pooling+Conv+BN+Relu+Conv+BN+Relu, the up sampling operation is Upsample+Conv+BN+Relu+Conv+BN+Relu+Concat, and the feature fusion fuse operation is Conv1*1+Conv3*3+Conv1*1+Conv3*3+Conv1*1+Conv3*3. Assuming that the input feature of each operation is x, and the input features of the up operation are x1,x2, the corresponding operations are expressed as follows.

[0080] inc=Relu(BN(Conv(Relu(BN(Conv(x))))))

[0081] outc = Softmax(Conv(X))

[0082] down=Relu(BN(Conv(Relu(BN(Conv(Max pooling(x)))))))

[0083] up=concat(Relu(BN(Conv(Relu(BN(Conv(Upsample(x1))))))),x2)

[0084] x1_1=Conv1*1(x)+Conv3*3(x)

[0085] x2_1=Conv1*1(x1_1)+Conv3*3(x1_1)

[0086] x3_1=Conv1*1(x2_1)+Conv3*3(x2_1)

[0087] fuse=Conv1*1(x3_1)+Conv3*3(x3_1)

[0088] The final output of the network is 572*572*4 features, including belt segmentation feature map, left roller segmentation feature map, and right roller segmentation feature map.

[0089] like Figure 3 As shown, in the belt segmentation effect diagram, the pure black pixel value (0,0,0) represents the background, the dark gray pixel value (85,85,85) represents the belt, the light gray pixel value (170,170,170) represents the left support roller, and the white pixel value (250,250,250) represents the right support roller.

[0090] In the belt straight line equation diagram, the red dot on the belt represents the belt center point P center , the blue dots represent the center points P corresponding to each support roller on the left left_i The green dots represent the center points P of each support roller on the right. right_j The left support roller straight line, the right support roller straight line and the belt center line are shown in the figure respectively.

[0091] like Figure 4 As shown in the belt segmentation effect diagram, the pure black pixel value (0,0,0) represents the background, and the dark gray pixel value (85,85,85) represents the belt. For this video frame without a support roller, the IOU method is used to calculate the maximum IOU and the belt centerline slope according to the rotation centerline angle, and then the belt offset rate is obtained.

[0092] like Figure 5 The figure shows the belt segmentation feature map after rotating 179 degrees through the center line of the belt. At this time, the IOU value of the symmetric map of the left belt and the segmentation map of the right belt reaches 0.88, which is the maximum IOU value.

[0093] like Figure 6 As shown in the belt segmentation effect diagram, the pure black pixel value (0,0,0) represents the background, and the white pixel value (250,250,250) represents the right roller. In the belt straight line equation diagram, the red dot on the belt represents the belt center point P center , the green dots represent the center points P corresponding to each support roller on the right right_j This is a serious deviation, with a deviation rate of -64%.

[0094] like Figure 7 As shown in the belt segmentation effect diagram, the pure black pixel value (0,0,0) represents the background, and the white pixel value (250,250,250) represents the right roller. In the belt straight line equation diagram, the red dot on the belt represents the belt center point P center , the blue dots represent the center points P corresponding to each support roller on the left left_iThe green dots represent the center points P of each support roller on the right. right_j This is a serious deviation, with a deviation rate of -52%.

[0095] Only the preferred embodiments of the present invention are described in detail above, but the present invention is not limited to the above embodiments. Various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention, and various changes should be included in the protection scope of the present invention.

Claims

1. A belt deviation detection method based on deep learning, characterized in that: The following steps are involved: S1. The algorithm accesses the real-time video stream for processing. Each frame of the video image is processed by the algorithm to determine whether there is belt deviation in the current frame; S2: If the deviation occurs within 10 consecutive frames, the belt deviation alarm will be issued; S3, calculating the belt deviation rate K; S4. The network is modified based on the traditional UNet network, and the final output of the network is 572*572*4 features.

2. The belt deviation detection method based on deep learning according to claim 1 is characterized in that: The method of algorithm processing each frame of the video image in S1 is as follows: the video frame image is processed by the UNet model to obtain the segmentation result of the current frame, including the left support roller segmentation feature map, the right support roller segmentation feature map, and the belt segmentation feature map.

3. The belt deviation detection method based on deep learning according to claim 1 is characterized in that: The belt deviation rate K calculated in S3 includes processing a belt with left and right rollers; processing a belt without rollers; processing a belt with only one roller on one side; and processing a belt with only one roller on one side.

4. The belt deviation detection method based on deep learning according to claim 3 is characterized in that: The method for processing the belt with left and right rollers is as follows: Based on the belt segmentation feature map, the center point coordinates P of the belt are obtained according to the findContours method in opencv center =(x1,y1); Based on the left idler segmentation feature map and the right idler segmentation feature map, the center points of multiple left and right idlers are obtained respectively using the same method, and are represented as P left_i =(x1_i,y1_i), P right_j =(x1_j,y1_j), where 1 < i < m, 1 < j < n, and m and n represent the numbers of left and right idlers respectively; Define the linear equation of the left support roller as y=kx+b, and use the least square method to fit the parameters k and b based on multiple points of the left support roller, and finally obtain the linear equation of the left support roller; Here, the distance in the least squares fitting process is selected as L1 distance to better avoid the interference of outlier sample points; similarly, the straight line equation of the right side of the support is obtained; Based on the intersection equation of the two straight lines, find the intersection of the left and right support lines and get the intersection point P cross =(x,y); Based on the belt center point P center The intersection point P of the left and right support rollers cross The centerline equation of the support roller can be calculated; Calculate the belt center point P based on the point-to-point distance formula center The average distance D to the center point of the left and right support rollers left, D right ;D left =mean(dist(P center ,P left )), D right =mean(dist(P center ,P right )), where dist represents the distance between two points and mean represents the average; Based on the distance D between the left and right support rollers left, D right According to the formula, the offset rate K = scale*(D left -D right ) / min(D left, D right ) calculates the belt deviation rate of the final roller image; where scale represents the scaling factor, which is 1.2 here. A negative deviation rate indicates that the belt is deviating to the left, and a positive deviation rate indicates that the belt is deviating to the right.

5. The belt deviation detection method based on deep learning according to claim 3 is characterized in that: The method for processing a belt without a support roller is as follows: Based on the belt segmentation feature map, the center point coordinate P of the belt is obtained according to the findContours method in opencv. center =(x1,y1); First, generate a vertical line through the center point of the belt to divide the belt into two parts. Flip the left half of the belt axially symmetrically based on the center line and calculate the IOU of the two halves of the belt. Similarly, rotate the center line of the belt by 0-179 degrees, and calculate the IOU according to the above method, and finally get 180 IOU values; calculate the maximum value of these IOU values, and the rotation angle corresponding to this value is the center line angle of the belt, so as to obtain the slope k of the belt; The final offset rate is obtained according to the offset rate formula of the unsupported belt: K=1.0-IOU.

6. The belt deviation detection method based on deep learning according to claim 3 is characterized in that: The method for processing a belt with only rollers on one side is as follows: when only one side of the rollers is detected, it can be directly determined that the belt is seriously deviated, and the belt deviation rate is directly defined as 100%; the number of rollers on both sides is determined, if the number of rollers on the left side is 0, the belt is seriously deviated to the left; if the number of rollers on the right side is 0, the belt is seriously deviated to the right.

7. The belt deviation detection method based on deep learning according to claim 3 is characterized in that: The method for processing a belt with only one roller on one side is as follows: when only one roller is detected on one side of the belt, it indicates that the belt is seriously deviated, and the belt deviation rate is directly defined as 100%; the number of rollers on both sides is judged, if the number of rollers on the left side is 1, the belt is seriously deviated to the left; if the number of rollers on the right side is 1, the belt is seriously deviated to the right.

8. The belt deviation detection method based on deep learning according to claim 1 is characterized in that: The method for modifying the network in S4 based on the traditional UNet network is: The input image size of the network is 572*572*3. It is first processed by the inc module, and then processed by three downsampling modules (down1, down2, down3). Each time it is downsampled, the feature resolution is halved and the feature channel is doubled; then it is processed by three upsampling modules (up1, up2, up3). Each time it is upsampled, the feature resolution is doubled and the feature channel is halved; Finally, the features are output through the outc module; the inc operation is Conv+BN+Relu+Conv+BN+Relu, the outc operation is Conv+Softmax operation; the down sampling operation is Max pooling+Conv+BN+Relu+Conv+BN+Relu, the up sampling operation is Upsample+Conv+BN+Relu+Conv+BN+Relu+Concat, and the feature fusion fuse operation is Conv1*1+Conv3*3+Conv1*1+Conv3*3+Conv1*1+Conv3*3. Assuming that the input feature of each operation is x, and the input features of the up operation are x1,x2, the corresponding operations are expressed as follows: inc=Relu(BN(Conv(Relu(BN(Conv(x)))))) outc = Softmax(Conv(X)) down=Relu(BN(Conv(Relu(BN(Conv(Max pooling(x))))))) up=concat(Relu(BN(Conv(Relu(BN(Conv(Upsample(x1))))))),x2) x1_1=Conv1*1(x)+Conv3*3(x) x2_1=Conv1*1(x1_1)+Conv3*3(x1_1) x3_1=Conv1*1(x2_1)+Conv3*3(x2_1) fuse=Conv1*1(x3_1)+Conv3*3(x3_1) The final output of the network is 572*572*4 features, including belt segmentation feature map, left roller segmentation feature map, and right roller segmentation feature map.

9. The belt deviation detection method based on deep learning according to claim 1, characterized in that: The absolute value of the deviation rate K is within a normal acceptable range of 0-20%, 20%-50% is a slight deviation, and 50%-100% is a serious deviation.

Images