Diamond wire jumper detection method for polysilicon cutting machine based on small samples

Abstract

The present invention belongs to the technical field of diamond wire jumper detection, and specifically relates to a polysilicon cutting machine diamond wire jumper detection method based on small samples. The invention expands the number of samples through a jumper sample generation algorithm based on gradient changes, solves the problem of insufficient training samples, improves the C2f module of the YOLOv8 model, reduces the number of parameters and the amount of calculation of the model, and extracts more context information through multi-scale extended convolution, enhances the model's ability to distinguish jumpers from background noise, introduces a CBAM attention module, effectively suppresses background interference, enhances sensitivity to small targets, and improves the detection accuracy of the model. Combined with experimental results, it is shown that the improved algorithm is compared with YOLOv8n, and the mAP is correspondingly improved, while the number of parameters and the amount of calculation are also correspondingly reduced, which fully meets the actual needs of jumper detection.

Description

Technical Field

[0001] The invention belongs to the technical field of diamond wire jumper detection, and in particular relates to a diamond wire jumper detection method for a polysilicon cutting machine based on a small sample. Background Art

[0002] With the rapid development of the photovoltaic industry, silicon cells, as the core material for photovoltaic power generation, are increasingly becoming an important part of energy transformation. Polysilicon cutting machines are key equipment for silicon cell production. Their working principle is to evenly arrange the diamond wire in the groove of the wire roller through the wire retracting and unreeling wheel to form a diamond wire net. The motor drives the wire roller to rotate at high speed to achieve the diamond wire cutting of silicon rods. During the high-speed cutting process, the diamond wire may detach from the wire roller groove due to wear or uneven tension, resulting in wire jumps, leaving empty grooves on the diamond wire net, and the diamond wire jumps seriously affect the cutting efficiency of the silicon rod and even cause wire breakage and shutdown.

[0003] At present, although there are some diamond wire jumper detection methods, they still face the following difficulties:

[0004] (1) The number of samples is limited. Due to the randomness of jumper generation and the limitation of production cycle, the number of jumper samples collected is very limited;

[0005] (2) The target size is small, occupying only 150 to 300 pixels in a 2500W resolution image;

[0006] (3) The background is complex and there is a large amount of splashing coolant in the machine cavity during the cutting process, resulting in a lot of noise in the collected images. Summary of the invention

[0007] The purpose of the present invention is to provide a polysilicon cutting machine diamond wire jumper detection method based on small samples, which can effectively overcome the difficulties of limited sample number, small target size and complex background in the prior art, and realize accurate detection of polysilicon cutting machine diamond wire jumpers.

[0008] The technical solution adopted by the present invention is as follows:

[0009] A polysilicon cutting machine diamond wire jumper detection method based on a small sample includes:

[0010] Obtain the diamond wire mesh image of the polysilicon cutting machine, and by counting the gradient changes of the diamond wire mesh brightness in the diamond wire mesh image, remove the parts with abnormal gradient and brightness, restore the complete diamond wire mesh background, and record it as the basic sample;

[0011] Expand the basic samples and output them as defect sample data sets;

[0012] Build a deep learning model and train it with defect sample data sets to identify and locate diamond wire jumper defects;

[0013] Among them, the basic samples are expanded and output as defect sample data sets including:

[0014] Divide the area around the jumper into different sub-areas, calculate the average brightness and area of ​​each sub-area, and record them as the first condition parameter and the second condition parameter respectively;

[0015] Filtering out a jumper sub-image from the diamond wire mesh image according to a first condition parameter and a second condition parameter;

[0016] A geometric transformation operation is performed on the jumper sub-image, and the jumper sub-image after geometric transformation is fused with the diamond wire mesh background based on transparency to generate a jumper defect sample.

[0017] In a preferred embodiment, in the diamond wire mesh image, the gradient change of the diamond wire mesh brightness is for:

[0018] ;

[0019] in, and Diamond wire mesh direction, The gradient change in direction, For diamond wire mesh and The angle between the directions is taken as the edge, and the area with significant gradient changes is used to divide the diamond wire network into multiple connected domains, and the average brightness value of each connected domain is calculated. :

[0020] ;

[0021] in, is the connected domain in the image, is the number of pixels in the connected domain, is the average brightness of the connected domain. For the connected domain below the preset threshold, the average brightness of the left and right neighboring domains and the gradient change are used to fuse and fill, and a complete diamond wire mesh background image is obtained;

[0022] ;

[0023] in, is the connected domain with abnormal brightness, and are the gradient values ​​of the left and right neighbors of the connected domain, respectively. and are the average brightness of the left and right neighbors of the connected domain, The brightness filled in the connected domain.

[0024] In a preferred solution, the step of dividing the area around the jumper into different sub-areas and calculating the average brightness and area of ​​each sub-area includes:

[0025] Acquire the image area around the jumper, and divide the area around the jumper into multiple sub-areas of different sizes according to the shape and position of the jumper;

[0026] Counting the pixel values ​​in each of the sub-regions, and calculating the average brightness value of the internal pixels according to the pixel values ​​in each of the sub-regions, and recording it as the first condition parameter;

[0027] Extracting edge curves of each sub-region, constructing a virtual coordinate system based on the edge curves, and calibrating the edge inflection point coordinates of the edge curves in the virtual coordinate system;

[0028] The area of ​​each sub-region is calculated according to the edge inflection point coordinates, and then the average value of the area of ​​each sub-region is calculated to obtain the average area of ​​each sub-region, and the average area is recorded as the second condition parameter.

[0029] In a preferred embodiment, the step of segmenting the jumper sub-image from the diamond wire mesh image according to the first condition parameter and the second condition parameter comprises:

[0030] Obtain a first condition parameter and a second condition parameter;

[0031] Acquire a first screening threshold value corresponding to the first condition parameter and a second screening threshold value corresponding to the second condition parameter;

[0032] comparing the first screening threshold with a first condition parameter, and comparing the second screening threshold with a second condition parameter;

[0033] When the first condition parameter is greater than or equal to a first screening threshold value, and the second condition parameter is greater than or equal to a second screening threshold value, recording the corresponding sub-region as a region containing jumper defects, and segmenting the region from the diamond wire mesh image as a jumper sub-image;

[0034] When the first condition parameter is less than a first screening threshold, or the second condition parameter is less than a second screening threshold, the corresponding sub-region is recorded as a region not containing jumper defects and is removed.

[0035] In a preferred solution, the geometric transformation operation performed on the jump line sub-image includes image rotation, image scaling, and image affine.

[0036] In a preferred embodiment, the image fusion formula based on transparency of the jumper sub-image and the diamond wire mesh background is:

[0037] ;

[0038] in, , are the coordinates of the diamond wire mesh background image, , are the coordinates of the jumper subgraph; is the brightness of the diamond wire mesh background image, is the brightness of the jump line sub-image after transformation, is the transparency of the jumper sub-image, , are the width and height of the jumper sub-image respectively, , They are respectively the jumper sub-image relative to the background image , The offset of the direction.

[0039] In a preferred embodiment, the deep learning model is an improved YOLOv8n model, and when training the deep learning model, an Extra DW module consisting of two depth-separable convolutional layers in series is introduced. When the ExtraDW module is executed, a 3×3 convolution kernel is used to perform a deep convolution operation on the input feature map, and then a Swish activation function is used to promote the transmission of high-level features, wherein the expression of the Swish activation function is:

[0040] ;

[0041] in, Compared with the RELU function in YOLOv8, the Sigmoid function can adaptively adjust the activation value during training, provide nonlinear transformation, and make high-level features smoothly transmitted in the network;

[0042] Through point-by-point convolutional layers, channel information is fused and the dimension of feature maps is expanded to enhance the network’s perception of high-level and contextual information of skipped lines. The expanded feature maps will be subjected to additional depth convolutions of 3×3, 5×5, and 7×7 convolution kernels respectively.

[0043] After passing through the Swish activation function again, channel fusion and compression are performed with the point-by-point convolution layer to remove redundant information, output the compressed feature map, and participate in subsequent feature extraction and fusion.

[0044] In a preferred embodiment, the deep learning model further introduces a CBAM module consisting of two independent channel attention modules and a spatial attention mechanism module connected in series;

[0045] Among them, the channel attention module is used to adaptively enhance the deep learning model's attention to key feature channels, and the spatial attention module is used to capture the weight of each spatial position in the image to assist the network in focusing on the area where the diamond wire mesh is located.

[0046] In a preferred scheme, after the deep learning model training is completed, the accuracy, recall rate, average accuracy, number of parameters, detection speed and number of floating-point operations per second are used as performance indicators to evaluate the deep learning model after training, and when the deep learning model meets the preset requirements, it is determined that the deep learning model training is completed.

[0047] And, an electronic device, the electronic device comprising:

[0048] at least one processor;

[0049] and a memory communicatively coupled to the at least one processor;

[0050] The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the above-mentioned polysilicon cutting machine diamond wire jumper detection method based on small samples.

[0051] The technical effects achieved by the present invention are:

[0052] The invention expands the number of samples and solves the problem of insufficient training samples through a jumper sample generation algorithm based on gradient changes, improves the C2f module of the YOLOv8 model, reduces the number of parameters and the amount of calculation of the model, extracts more context information through multi-scale extended convolution, enhances the model's ability to distinguish jumpers from background noise, introduces the CBAM attention module, effectively suppresses background interference, enhances the sensitivity to small targets, and improves the detection accuracy of the model. Combined with experimental results, it is shown that the improved algorithm has a corresponding improvement in mAP compared with YOLOv8n, while the number of parameters and the amount of calculation are also reduced accordingly, which fully meets the actual needs of jumper detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a schematic diagram of the method flow in Embodiment 1 of the present invention;

[0054] Figure 2 is a jumper defect sample containing jumper defect information according to the first embodiment of the present invention;

[0055] Figure 3 It is a diagram of the improved YOLOv8 network structure in Embodiment 1 of the present invention;

[0056] Figure 4This is a module diagram of EDW-C2f in Embodiment 1 of the present invention;

[0057] Figure 5 is a diagram of the CBAM attention module in the first embodiment of the present invention;

[0058] Figure 6 This is a jumper detection result diagram of the original YOLOv8 model in the second embodiment of the present invention;

[0059] Figure 7 : is a jumper detection result diagram of the improved YOLOv8 model in the second embodiment of the present invention;

[0060] Figure 8 It is a schematic diagram of the structure of an electronic device in Embodiment 4 of the present invention. DETAILED DESCRIPTION

[0061] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0062] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0063] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure or characteristic that may be included in at least one implementation of the present invention. The phrase "in a preferred embodiment" that appears in different places in this specification does not refer to the same embodiment, nor is it a separate or selective embodiment that is mutually exclusive with other embodiments.

[0064] At present, there is a lack of relevant data sets for diamond wire jumper detection of polysilicon cutting machines. An image acquisition device was installed on a polysilicon cutting machine. It took 5 months to obtain 43 jumper samples. However, the number of samples still cannot meet the training requirements of the deep learning model. Based on this, a data enhancement algorithm based on gradient changes was proposed to improve the problem of poor detection effect caused by insufficient samples.

[0065] Embodiment 1

[0066] See also Figure 1 As shown, it is a first embodiment of the present invention, which provides a polysilicon cutting machine diamond wire jumper detection method based on a small sample, comprising:

[0067] S1. Obtain a diamond wire mesh image of a polysilicon cutting machine, and statistically analyze the gradient changes of the diamond wire mesh brightness in the diamond wire mesh image, remove the parts with abnormal gradient and brightness, restore the complete diamond wire mesh background, and record it as a basic sample;

[0068] In step S1, when it is necessary to perform adaptive detection on the diamond wire jumper of the polysilicon cutting machine, it is first necessary to obtain the diamond wire mesh image of the polysilicon cutting machine, and then identify and remove the areas with abnormal gradient and brightness by analyzing the gradient change of the diamond wire mesh brightness in the image, so as to restore a complete diamond wire mesh background image, and record this complete diamond wire mesh background image as the basic sample, wherein the gradient change of the diamond wire mesh brightness in the diamond wire mesh image for:

[0069] ;

[0070] in, and Diamond wire mesh direction, The gradient change in direction, For diamond wire mesh and The angle between the directions is taken as the edge, and the area with significant gradient changes is used to divide the diamond wire network into multiple connected domains, and the average brightness value of each connected domain is calculated. :

[0071] ;

[0072] in, is the connected domain in the image, is the number of pixels in the connected domain, is the average brightness of the connected domain. For the connected domain below the preset threshold, the average brightness of the left and right neighboring domains and the gradient change are used to fuse and fill, and a complete diamond wire mesh background image is obtained;

[0073] ;

[0074] in, is the connected domain with abnormal brightness, and are the gradient values ​​of the left and right neighbors of the connected domain, respectively. and are the average brightness of the left and right neighbors of the connected domain, The brightness filled in the connected domain.

[0075] Specifically, when analyzing the diamond wire mesh image, we need to pay attention to the gradient change of the diamond wire mesh brightness. By identifying the areas with significant gradient changes, the diamond wire mesh can be divided into multiple connected domains, and the average brightness value of each connected domain is calculated. For the connected domains whose average brightness is lower than the preset threshold lower limit, the corresponding processing method is adopted, that is, the average brightness of the left and right neighbors are merged and filled with the gradient changes to obtain a complete diamond wire mesh background image. During the processing, special attention should be paid to the connected domains with abnormal brightness, and the gradient values ​​of their left and right neighbors should be calculated respectively. At the same time, the average brightness of the left and right neighbors will also be calculated. Finally, the brightness of the connected domain will be determined and filled based on this information.

[0076] S2, expand the basic samples and output them as defect sample data sets;

[0077] Among them, the basic samples are expanded and output as defect sample data sets including:

[0078] Divide the area around the jumper into different sub-areas, calculate the average brightness and area of ​​each sub-area, and record them as the first condition parameter and the second condition parameter respectively;

[0079] In step S2, after eliminating the brightness difference and gradient difference in the diamond wire mesh image, the area around the jumper is divided into several different sub-areas, the brightness and area of ​​each sub-area are calculated, the calculated average brightness value is used as the first condition parameter, and the corresponding area value is recorded as the second condition parameter, wherein the area around the jumper is divided into different sub-areas, and the steps of calculating the average brightness and area of ​​each sub-area include:

[0080] The image area around the jumper is obtained, and according to the shape and position of the jumper, the area around the jumper is segmented into multiple sub-areas of different sizes by using an adaptive threshold segmentation algorithm or a K-means algorithm;

[0081] Count the pixel values ​​in each sub-region, and calculate the average brightness value of the internal pixels based on the pixel values ​​in each sub-region;

[0082] Extracting edge curves of each sub-region, constructing a virtual coordinate system based on the edge curves, and calibrating the edge inflection point coordinates of the edge curves in the virtual coordinate system;

[0083] The area of ​​each sub-region is calculated according to the coordinates of the edge inflection point, and then the average area of ​​each sub-region is calculated to obtain the average area of ​​each sub-region;

[0084] Specifically, when processing the diamond wire mesh image, it is first necessary to divide the area around the jumper into different sub-areas. This process includes multiple steps, which can be achieved by analyzing the shape and position of the jumper. In order to perform segmentation more accurately, this embodiment uses an adaptive threshold segmentation algorithm. The adaptive threshold segmentation algorithm can dynamically adjust the threshold according to the specific situation of the image, thereby dividing the area around the jumper into multiple sub-areas of different sizes. Of course, the K-means algorithm can also be used for segmentation. The K-means algorithm is an unsupervised learning clustering algorithm that can divide image data into K categories according to similarity, thereby achieving segmentation of the area around the jumper. After that, the pixel values ​​in each sub-area need to be segmented. Row statistics, in this way, the average brightness value of the pixels inside each sub-region can be obtained, thereby reflecting the brightness characteristics of the sub-region. After obtaining the average brightness value of the sub-region, it is also necessary to extract the edge curves of each sub-region. Through the edge curves, a virtual coordinate system can be constructed, and in the virtual coordinate system, the edge inflection point coordinates of the edge curve will also be synchronously calibrated. Finally, based on the edge inflection point coordinates, the area of ​​each sub-region can be calculated. Specifically, the shoelace theorem can be introduced for calculation. After calculating the area of ​​each sub-region, the average value of these areas is calculated to obtain the average area of ​​each sub-region. Finally, the calculated average brightness value is used as the first condition parameter, and the corresponding area value is recorded as the second condition parameter.

[0085] Subsequently, a jumper sub-image is segmented from the diamond wire mesh image according to the first condition parameter and the second condition parameter;

[0086] In the above steps, after the first condition parameter and the second condition parameter are output, a jumper sub-graph containing jumpers can be accurately segmented from the diamond wire mesh image, wherein the step of selecting the jumper sub-graph from the diamond wire mesh image according to the first condition parameter and the second condition parameter includes:

[0087] Obtain a first condition parameter and a second condition parameter;

[0088] Obtaining a first screening threshold corresponding to the first condition parameter and a second screening threshold corresponding to the second condition parameter;

[0089] comparing a first screening threshold with a first condition parameter, and comparing a second screening threshold with a second condition parameter;

[0090] When the first condition parameter is greater than or equal to the first screening threshold, and the second condition parameter is greater than or equal to the second screening threshold, the corresponding sub-region is recorded as a region containing jumper defects, and the region is segmented from the diamond wire mesh image as a jumper sub-image;

[0091] When the first condition parameter is less than the first screening threshold, or the second condition parameter is less than the second screening threshold, the corresponding sub-region is recorded as a region not containing jumper defects and is removed.

[0092] Specifically, when accurately segmenting a jumper sub-image from a diamond wire mesh image based on the first condition parameter and the second condition parameter, it is first necessary to accurately obtain the first condition parameter and the second condition parameter. Secondly, according to the obtained first condition parameter, the corresponding first screening threshold is determined. At the same time, according to the second condition parameter, the corresponding second screening threshold is also determined. Then, a corresponding comparison operation is performed, that is, the first screening threshold is carefully compared with the first condition parameter, and the second screening threshold is also carefully compared with the second condition parameter. On the basis of the comparison result, if and only if the first condition parameter is greater than or equal to the first screening threshold, and the second condition parameter is also greater than or equal to the second screening threshold, the corresponding sub-region will be determined as a region containing jumper defects, and it will be accurately segmented from the diamond wire mesh image to form the required jumper sub-image. On the contrary, if the first condition parameter is less than the first screening threshold, or the second condition parameter is less than the second screening threshold, the corresponding sub-region will be recorded as a region not containing jumper defects, and will be eliminated in subsequent processing to ensure the accuracy and reliability of the final segmentation result.

[0093] Finally, a geometric transformation operation is performed on the jumper sub-image, and the jumper sub-image after geometric transformation is fused with the diamond wire mesh background based on transparency to generate a jumper defect sample;

[0094] In the above steps, necessary geometric transformation operations are performed on the segmented jumper sub-image to ensure the geometric consistency between the sub-image and the diamond wire mesh background image. Here, the geometric transformation operations performed on the jumper sub-image include image rotation, image scaling, image affine and other methods. The purpose is to enable the jumper sub-image to accurately match the size and angle of the diamond wire mesh background image, thereby improving the accuracy and reliability of jumper defect detection. When performing image rotation, the sub-image can be rotated by a corresponding angle based on the difference in rotation angle between the sub-image and the background image to eliminate the rotation deviation between the two. When performing image scaling, the sub-image can be rotated by a corresponding angle based on the difference in rotation angle between the sub-image and the background image to eliminate the rotation deviation between the two. The size difference between the sub-image and the background image is that the sub-image is appropriately enlarged or reduced to ensure the size consistency between the sub-image and the background image. When performing image affine transformation, the position, shape and angle of the sub-image can be further adjusted to make it fit the background image more closely. Through geometric transformation operations, the geometric inconsistency problem between the jumper sub-image and the diamond wire mesh background image can be effectively solved, laying the foundation for transparent image fusion. Then, through the transparency-based image fusion technology, the jumper sub-image after geometric transformation is fused with the diamond wire mesh background image to generate a jumper defect sample containing jumper defect information. For details, please refer to Figure 2 As shown in the attached Figure 2 In the figure, a is a diamond wire mesh image with jumpers, b is the obtained background image, c is a jumper sub-image, and d is a generated jumper sample. The image fusion formula based on transparency of the jumper sub-image and the diamond wire mesh background is:

[0095] ;

[0096] in, , are the coordinates of the diamond wire mesh background image, , are the coordinates of the jumper subgraph; is the brightness of the diamond wire mesh background image, is the brightness of the jump line sub-image after transformation, is the transparency of the jumper sub-image, , are the width and height of the jumper sub-image respectively, , They are respectively the jumper sub-image relative to the background image , The offset of the direction.

[0097] S3, building a deep learning model, and training the deep learning model through jumper defect samples to identify and locate diamond wire jumper defects;

[0098] In step S3, see the attached Figure 3 and attached Figure 4 After the jumper defect samples are output, a deep learning model is constructed. The jumper defect samples generated above are used to train the model so that it can effectively identify and locate defects on the diamond wire jumpers, thereby improving the working efficiency and cutting quality of the polysilicon cutting machine. The deep learning model is an improved YOLOv8n model. The Extra DW module is introduced to optimize the C2f module to reduce the number of parameters and calculations. The multi-scale extended convolution is used to expand the receptive field to enhance the model's ability to distinguish jumpers from background noise. The CBAM attention module is added to the Neck part to improve the channel and space representation capabilities in the feature fusion process, suppress background interference, and enhance the accuracy of small eye detection. Due to the obvious features of the diamond wire jumper, the small target and the complex background, the model is prone to rely on the underlying features (such as color and texture), lose high-level features and context information, and cause missed detection and false detection. In order to solve this problem, the Bottleneck layer in C2f is improved, and the Extra DW module (EDW) is introduced to enhance the module's ability to extract high-level features and context information, and improve the model's ability to distinguish jumpers from background noise.

[0099] Specifically, when training the deep learning model, an Extra DW module consisting of two depth-separable convolutional layers in series is introduced. When the Extra DW module is executed, a 3×3 convolution kernel is used to perform a deep convolution operation on the input feature map, and then the Swish activation function is used to promote the transmission of high-level features. The expression of the Swish activation function is:

[0100] ;

[0101] in, Compared with the RELU function in YOLOv8, the Sigmoid function can adaptively adjust the activation value during training, provide nonlinear transformation, and make high-level features smoothly transmitted in the network;

[0102] By fusing channel information through point-by-point convolutional layers and expanding the dimension of feature maps, the network’s perception of high-level and contextual information of skipped lines is enhanced. The extended feature maps are further convolved with additional depth convolutions of 3×3, 5×5, and 7×7 kernels to improve the network’s perception of contextual information and the model’s ability to distinguish skipped lines from background noise.

[0103] After passing through the Swish activation function again, channel fusion and compression are performed with the point-by-point convolution layer to remove redundant information, output the compressed feature map, and participate in subsequent feature extraction and fusion.

[0104] In a preferred embodiment, please refer to Figure 5 ,The deep learning model also introduces a CBAM module which consists of two independent channel attention modules and a spatial attention mechanism module in series;

[0105] Among them, the channel attention module is used to adaptively enhance the deep learning model's attention to key feature channels, and the spatial attention module is used to capture the weight of each spatial position in the image to assist the network in focusing on the area where the diamond wire mesh is located.

[0106] In this embodiment, during the cutting process of the polysilicon cutting machine, the silicon powder ground from the diamond wire will mix with the sprayed coolant and adhere to the diamond wire, causing changes in the color and brightness characteristics of the wire mesh and jumper wires. To solve this problem, the CBAM attention mechanism is introduced and connected to the Neck part of the YOLOv8 network, so that the network feature fusion process can focus more accurately on the target area and improve the model's sensitivity to small targets. The channel attention module can adaptively enhance the model's attention to key feature channels, improve the feature expression ability, enable the network to better capture the changes in jumpers, suppress background noise, and improve the accuracy and robustness of detection. The spatial attention module focuses on capturing the weight of each spatial position in the image, allowing the network to focus on the area where the diamond wire mesh is located, and improve the positioning accuracy of the model.

[0107] Secondly, after the deep learning model training is completed, the accuracy, recall rate, average accuracy, number of parameters, detection speed and number of floating-point operations per second are used as performance indicators to evaluate the deep learning model after training. When the deep learning model meets the preset requirements, the deep learning model training is determined to be completed.

[0108] In this implementation, in order to better evaluate the model performance and understand the pros and cons of the model, the accuracy (P), recall (R), average accuracy (mAP), parameter quantity (Params), detection speed (FPS) and floating point operations per second (GFLOPs) are used as performance indicators, and the evaluation function involved is:

[0109] ;

[0110] ;

[0111] ;

[0112] in, Represents the number of samples that are actually positive and predicted to be positive. Indicates the number of samples that are actually negative but predicted to be positive. Indicates the number of samples that are actually negative but predicted to be negative. is the sample category, and the calculated and As the ordinate and abscissa respectively, draw curve, It is the average value after the integration of the curves of each category. FPS is the number of frames processed by the model per second, which is used to measure the real-time performance of the deep learning model. Params and GFLOPs represent the number of model parameters and the amount of calculation, respectively, which are used to measure the model complexity and computing resource requirements.

[0113] Embodiment 2

[0114] See attached Figure 6 and attached Figure 7 ,This embodiment provides an ablation experiment of the diamond wire jump detection method based on the improved YOLOv8.

[0115] The ablation experiment uses two data sets, one is the diamond wire mesh data set collected on-site, totaling 43 images, and the other is the data set expanded by the diamond wire jumper detection method for polysilicon cutting machines based on small samples, totaling 742 images. The 699 images generated by the diamond wire jumper detection method for polysilicon cutting machines based on small samples are used as the training set, and the 43 images collected on-site are used as the test set.

[0116] In order to verify the effectiveness of the proposed diamond wire jump detection algorithm based on improved YOLOv8, 6 groups of experiments were carried out. The EDW-C2f module, CBAM module and data enhancement algorithm were added to the YOLOv8n network respectively, and the combined effect of multiple improved modules was given; the ablation experiment results are shown in Table 1.

[0117] Table 1. Comparison of ablation experiment structures

[0118]

[0119] As shown in Table 1, the various improvement methods proposed on the basis of the YOLOv8n algorithm have played a certain role in improving the detection performance of the deep learning model, and the detection algorithm has a significant advantage in the jumper detection task. The detection effect of the improved algorithm and the YOLOv8n model is compared. Figure 7 shown.

[0120] Embodiment 3

[0121] This example provides a comparison between the improved YOLOv8n model and other models.

[0122] In order to further verify the effect of the proposed method on network improvement, the EDW-YOLOv8 model is compared with YOLOv5n, YOLOv6n, YOLOv7-tiny and YOLOv8n respectively. The experimental data uses the jump line dataset after data enhancement. The experimental results are shown in Table 2.

[0123] Table 2 Comparison of diamond wire jumper test results of different models

[0124]

[0125] As shown in Table 2, in wire jump detection, the improved YOLOv8n model has a slightly higher computational cost than YOLOv8n, YOLOv5n, YOLOv6n, and YOLOv7-tiny, and its FPS is slightly lower than that of YOLOv5 and YOLOv6, but its accuracy, recall, mAP, and number of parameters have all reached the best. The CBAM attention module is introduced to improve the sensitivity to small targets. The experimental results show that the improved algorithm has a 5.3% increase in mAP compared to YOLOv8n, while the number of parameters and the amount of computation are reduced by 24.1% and 6% respectively. The model detection speed reaches 80.2 frames / s, which fully meets the actual needs of wire jump detection.

[0126] Embodiment 4

[0127] See also Figure 8 , which is a fourth embodiment of the present invention, provides an electronic device, the electronic device comprising:

[0128] at least one processor;

[0129] and a memory communicatively coupled to the at least one processor;

[0130] The memory stores a computer program that can be executed by at least one processor, and the computer program is executed by at least one processor so that the at least one processor can execute the above-mentioned polysilicon cutting machine diamond wire jumper detection method based on small samples.

[0131] The processor of the above electronic device may be a central processing unit (CPU), a graphics processing unit (GPU) or a digital signal processor (DSP), etc. The memory may include a read-only memory (ROM), a random access memory (RAM), a hard disk, an optical disk, a USB flash drive, etc. The electronic device may also include an input device, an output device, a network interface, etc. The input device may include a keyboard, a mouse, a touch screen, etc. The output device may include a display, a printer, etc. The network interface may include a wired network interface and a wireless network interface, which is used for the electronic device to communicate data with other devices.

[0132] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0133] The above is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention. The structures, devices and operating methods not specifically described and explained in the present invention shall be implemented according to the conventional means in the art unless otherwise specified and limited.

Claims

1. A method for detecting diamond wire jumpers of polysilicon cutting machines based on small samples, characterized in that: include: Obtain the diamond wire mesh image of the polysilicon cutting machine, and statistically analyze the gradient changes of the diamond wire mesh brightness in the diamond wire mesh image, remove the parts with abnormal gradient and brightness, restore the complete diamond wire mesh background, and record it as the basic sample; Expand the basic samples and output them as defect sample data sets; Build a deep learning model and train it with defect sample data sets to identify and locate diamond wire jumper defects; Among them, the basic samples are expanded and output as defect sample data sets including: Divide the area around the jumper into different sub-areas, calculate the average brightness and area of ​​each sub-area, and record them as the first condition parameter and the second condition parameter respectively; Filtering out a jumper sub-image from the diamond wire mesh image according to a first condition parameter and a second condition parameter; Performing geometric transformation operation on the jumper sub-image, and performing transparency-based image fusion on the jumper sub-image after geometric transformation and the diamond wire mesh background to generate jumper defect samples; In the diamond wire mesh image, the gradient change of the brightness of the diamond wire mesh for: ； in, and Diamond wire mesh direction, The gradient change in direction, For diamond wire mesh and The angle between the two directions is taken as the edge, and the area with significant gradient changes is used to divide the diamond wire network into multiple connected domains, and the average brightness value of each connected domain is calculated. : ； in, is the connected domain in the image, is the number of pixels in the connected domain, is the average brightness of the connected domain. For the connected domain below the preset threshold, the average brightness of the left and right neighboring domains and the gradient change are used to fuse and fill, and a complete diamond wire mesh background image is obtained; ； in, is the connected domain with abnormal brightness, and are the gradient values ​​of the left and right neighbors of the connected domain, respectively. and are the average brightness of the left and right neighbors of the connected domain, The brightness filled in the connected domain.

2. The method for detecting diamond wire jumpers of polysilicon cutting machines based on small samples according to claim 1, characterized in that: The step of dividing the area around the jumper into different sub-areas and calculating the average brightness and area of ​​each sub-area comprises: Acquire the image area around the jumper, and divide the area around the jumper into multiple sub-areas of different sizes according to the shape and position of the jumper; Count the pixel values ​​in each sub-region, and calculate the average brightness value of the internal pixels according to the pixel values ​​in each sub-region, and record it as the first condition parameter; Extracting edge curves of each sub-region, constructing a virtual coordinate system based on the edge curves, and calibrating the edge inflection point coordinates of the edge curves in the virtual coordinate system; The area of ​​each sub-region is calculated according to the edge inflection point coordinates, and then the average value of the area of ​​each sub-region is calculated to obtain the average area of ​​each sub-region, and the average area is recorded as the second condition parameter.

3. The method for detecting diamond wire jumper of a polysilicon cutting machine based on a small sample according to claim 2, characterized in that: The step of selecting the jumper sub-image from the diamond wire mesh image according to the first condition parameter and the second condition parameter comprises: Obtain a first condition parameter and a second condition parameter; Obtaining a first screening threshold corresponding to the first condition parameter and a second screening threshold corresponding to the second condition parameter; comparing a first screening threshold with a first condition parameter, and comparing a second screening threshold with a second condition parameter; When the first condition parameter is greater than or equal to the first screening threshold, and the second condition parameter is greater than or equal to the second screening threshold, the corresponding sub-region is recorded as a region containing jumper defects, and the region is segmented from the diamond wire mesh image as a jumper sub-image; When the first condition parameter is less than the first screening threshold, or the second condition parameter is less than the second screening threshold, the corresponding sub-region is recorded as a region not containing jumper defects and is removed.

4. The method for detecting diamond wire jumpers of a polysilicon cutting machine based on a small sample according to any one of claims 1 to 3, characterized in that: The geometric transformation operation performed on the jump line sub-image includes image rotation, image scaling, and image affine.

5. The method for detecting diamond wire jumper of a polysilicon cutting machine based on a small sample according to claim 4, characterized in that: The image fusion formula based on transparency of the jumper sub-image and the diamond wire mesh background is: ； in, , are the coordinates of the diamond wire mesh background image, , are the coordinates of the jumper subgraph; is the brightness of the diamond wire mesh background image, is the brightness of the jump line sub-image after transformation, is the transparency of the jumper sub-image, , are the width and height of the jumper sub-image respectively, , They are respectively the jumper sub-image relative to the background image , The offset of the direction.

6. The method for detecting diamond wire jumper of a polysilicon cutting machine based on a small sample according to claim 1, characterized in that: The deep learning model is an improved YOLOv8n model, and when training the deep learning model, an Extra DW module consisting of two depth-separable convolutional layers in series is introduced. When the Extra DW module is executed, a 3×3 convolution kernel is used to perform a deep convolution operation on the input feature map, and then the Swish activation function is used to promote the transmission of high-level features, where the expression of the Swish activation function is: ； in, Compared with the RELU function in YOLOv8, the Sigmoid function can adaptively adjust the activation value during training, provide nonlinear transformation, and make high-level features smoothly transmitted in the network; Through point-by-point convolutional layers, channel information is fused and the dimension of feature maps is expanded to enhance the network's perception of high-level and contextual information of skipped lines. The expanded feature maps will be subjected to additional depth convolutions of 3×3, 5×5, and 7×7 convolution kernels respectively. After passing through the Swish activation function again, channel fusion and compression are performed with the point-by-point convolution layer to remove redundant information, output the compressed feature map, and participate in subsequent feature extraction and fusion.

7. The method for detecting diamond wire jumper of a polysilicon cutting machine based on a small sample according to claim 1, characterized in that: The deep learning model also introduces a CBAM module consisting of two independent channel attention modules and a spatial attention mechanism module connected in series; Among them, the channel attention module is used to adaptively enhance the deep learning model's attention to key feature channels, and the spatial attention module is used to capture the weight of each spatial position in the image to assist the network in focusing on the area where the diamond wire mesh is located.

8. The method for detecting diamond wire jumper of a polysilicon cutting machine based on a small sample according to claim 1, characterized in that: After the deep learning model training is completed, the accuracy, recall rate, average accuracy, number of parameters, detection speed and number of floating-point operations per second are used as performance indicators to evaluate the deep learning model after training, and when the deep learning model meets the preset requirements, it is determined that the deep learning model training is completed.

9. An electronic device, characterized in that: The electronic device comprises: at least one processor; and a memory communicatively coupled to the at least one processor; The memory stores a computer program executable by at least one processor, and the computer program is executed by at least one processor so that the at least one processor can execute the polysilicon cutting machine diamond wire jump detection method based on a small sample according to any one of claims 1 to 8.

Images