Polycrystalline silicon production equipment surface defect detection method and device, medium and equipment

Abstract

The invention discloses a polycrystalline silicon production equipment surface defect detection method, a polycrystalline silicon production equipment surface defect detection device, a medium and polycrystalline silicon production equipment, and relates to the technical field of metal equipment surface defect detection. The method comprises the following steps: integrating a full-dimension dynamic convolution algorithm and a spatial pyramid pooling module into a deep octave detection network of a YOLOv8 network, adding an efficient channel attention module into a triple octave detection head network of the YOLOv8 network, and replacing an intersection-to-union ratio function in the YOLOv8 network with a weighted intersection-to-union ratio loss function to obtain an improved YOLOv8 network; inputting the surface defect image of the polycrystalline silicon production equipment with the label into the improved YOLOv8 network for training to obtain an equipment defect detection model; and inputting the acquired surface defect image of the polycrystalline silicon production equipment into the equipment defect detection model to obtain a detection result of the surface defect of the polycrystalline silicon production equipment. According to the scheme, the detection precision and the detection efficiency of the surface defects of the polycrystalline silicon production equipment are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of metal equipment surface defect detection, and in particular to a method, device, medium and equipment for detecting defects on the surface of polysilicon production equipment. Background Art

[0002] In recent years, with the continuous development of my country's photovoltaic industry, my country's polysilicon industry has shown a rapid development trend, and its output has continued to grow. Polysilicon distillation towers and reduction furnaces are important equipment for manufacturing polysilicon ingots. In the production process of polysilicon, due to the usually harsh working environment, defects are prone to appear on the surface of the equipment. These defects have an important impact on product quality, production efficiency and even worker safety. In the manufacturing process of polysilicon distillation towers, various metal materials are required, among which steel is the main structural material. The tower body, brackets, connectors, etc. of the distillation tower often use stainless steel or special steel, and the distillation tower operates under high temperature and high pressure, so the integrity and corrosion resistance of its metal structure are crucial. Through surface inspection, potential cracks, corrosion, wear and other problems can be discovered in time to prevent equipment failure. The surface inspection technology of metal equipment can ensure the quality of the steel used in the distillation tower, thereby ensuring the safety of the polysilicon production process. Therefore, timely and accurate target detection of surface defects of polysilicon production equipment has great practical application value.

[0003] Traditional surface defect detection methods rely on experienced technicians to conduct visual inspections or use basic image processing techniques to identify defects. These methods include using magnifying glasses, microscopes, or simple cameras to magnify the details of the object surface so that the human eye can recognize irregularities. However, this manual detection method is mostly manual feature extraction, which is usually active and difficult, not only time-consuming and labor-intensive, but also lacks accuracy and repeatability. It relies heavily on the experience and judgment of the inspector, resulting in high error rates and low confidence in distinguishing defect-free products. With the development of artificial intelligence technology, deep learning models can be widely used in various fields, especially the emergence of real-time target detection algorithms such as YOLO (You Only Look Once), which makes industrial automation and high-precision detection possible. Deep learning models extract features through multi-layer neural networks, and then after multiple rounds of training, use multi-layer networks to optimize end-to-end performance to enhance target recognition and positioning. There are usually two types of deep learning-based target detection methods: two-stage methods and one-stage methods. The former, such as Convolutional Neural Networks (CNN), has achieved remarkable success in image recognition and target detection. CNN can learn feature expressions from a large number of samples and has strong generalization ability. The former generates a large number of prior frames that may contain the object to be detected during detection, and then uses a classifier to determine whether the bounding box corresponding to each prior frame contains the object to be detected, as well as the probability or confidence of the category to which the object belongs. At the same time, it is necessary to process and correct the bounding box, and finally filter out the bounding boxes with low confidence and high overlap based on some criteria to obtain the inspection results. However, various defect detection algorithms based on convolutional neural network methods have more or less model parameters and slow detection speed, which makes it difficult to meet the needs of real-time detection in industrial production. The latter is a detection method represented by YOLO. Its core idea is to replace the two-stage algorithm of generating the region of interest (RoI) and target detection with a one-stage algorithm of a set of networks, and directly regress the position and category of the bounding box (bbox) in the output layer. YOLO divides the input image into S*S grids, and each grid is responsible for predicting the target in the area. When the center point of an object falls into a grid, the grid will be responsible for detecting the object. The YOLO algorithm adopts an end-to-end training method, directly regressing from image to bounding box, without additional region generation steps. This method makes training faster and can optimize model performance under the standard regression loss function. As the latest version of the YOLO series algorithm, YOLOv8 (You Only Look Once version8) has the characteristics of fast detection speed and high accuracy, and is very suitable for real-time detection in industrial sites.Compared with other target detection algorithms, YOLOv8 can achieve real-time detection while maintaining high accuracy, which is of great significance for real-time monitoring in industrial production. Although the target detection algorithm based on YOLOv8 has made great progress in the detection of surface defects of metal equipment in polysilicon production, there are still some challenges and problems in many cases. For example, the surface defects of the equipment may vary in size, small defects are often difficult to detect, and it is easy to have unclear boundary demarcation, false detection and misdetection problems. It is necessary to more accurately identify and locate the surface defects of the equipment. And because the industrial site environment is complex, the real-time requirements are high, and the detection system needs to process a large amount of image data, it is necessary to design an efficient image processing algorithm to ensure that the system can respond quickly. The surface defects of polysilicon production equipment may be affected by factors such as lighting, angle, and occlusion. Certain technologies need to be used to improve the robustness of the model in complex environments.

[0004] In summary, the existing methods for detecting surface defects of polysilicon production equipment have the problems of low detection accuracy and slow detection speed. Summary of the invention

[0005] Based on this, it is necessary to provide a method, device, medium and equipment for detecting defects on the surface of polysilicon production equipment to address the problems of low detection accuracy and slow detection speed in existing methods for detecting surface defects of polysilicon production equipment.

[0006] The present invention adopts the following technical solutions:

[0007] In a first aspect, the present invention provides a method for detecting defects on the surface of a polysilicon production device, the method comprising:

[0008] The full-dimensional dynamic convolution algorithm and the spatial pyramid pooling module are integrated into the deep octave detection network of the YOLOv8 network, and the efficient channel attention module is added to the triple octave detection head network of the YOLOv8 network, and the intersection-over-union function in the YOLOv8 network is replaced with a weighted intersection-over-union loss function to obtain an improved YOLOv8 network;

[0009] Inputting the annotated surface defect image of the polysilicon production equipment into the improved YOLOv8 network for training to obtain an equipment defect detection model;

[0010] The collected surface defect images of the polysilicon production equipment are input into the equipment defect detection model to obtain the detection results of the surface defects of the polysilicon production equipment.

[0011] Furthermore, the integration of the full-dimensional dynamic convolution algorithm and the spatial pyramid pooling module into the deep octave detection network of the YOLOv8 network specifically includes:

[0012] Integrate the first full-dimensional dynamic convolution algorithm layer, the second full-dimensional dynamic convolution algorithm layer, the third spatial pyramid pooling layer, the fourth full-dimensional dynamic convolution algorithm layer, the fifth spatial pyramid pooling layer, the sixth full-dimensional dynamic convolution algorithm layer, the seventh spatial pyramid pooling layer, the eighth full-dimensional dynamic convolution algorithm layer, the ninth spatial pyramid pooling layer, the tenth spatial pyramid pooling and feature splicing and spatial convolution layer connected in sequence into the deep octave detection network;

[0013] Among them, the first full-dimensional dynamic convolution algorithm layer is used to accept a four-dimensional tensor as input, the second full-dimensional dynamic convolution algorithm layer is used to receive the features of the first full-dimensional dynamic convolution algorithm layer, the third spatial pyramid pooling layer is used to receive the features of the second full-dimensional dynamic convolution algorithm layer, the fourth full-dimensional dynamic convolution algorithm layer is used to receive the features of the third spatial pyramid pooling layer, the fifth spatial pyramid pooling layer is used to receive the features of the fourth full-dimensional dynamic convolution algorithm layer, the sixth full-dimensional dynamic convolution algorithm layer is used to receive the features of the fifth spatial pyramid pooling layer, the seventh spatial pyramid pooling layer is used to receive the features of the sixth full-dimensional dynamic convolution algorithm layer, the eighth full-dimensional dynamic convolution algorithm layer is used to receive the features of the seventh spatial pyramid pooling layer, the ninth spatial pyramid pooling layer is used to receive the features of the eighth full-dimensional dynamic convolution algorithm layer, and the tenth spatial pyramid pooling and feature splicing and spatial convolution layer is used to receive the features of the ninth spatial pyramid pooling layer.

[0014] Furthermore, the full-dimensional dynamic convolution algorithm combines the dynamic convolution submodule and the attention mechanism submodule to:

[0015] Performing a convolution operation on the input data through the dynamic convolution submodule to obtain a first feature;

[0016] Attention is adjusted on the first feature by the attention mechanism submodule to obtain a second feature;

[0017] A weighted operation is performed on the second feature to obtain an output feature corresponding to the input data.

[0018] Further, before the dynamic convolution submodule performs a convolution operation on the input data, and after the dynamic convolution submodule performs a convolution operation on the input data, the method further includes:

[0019] Using the full-dimensional dynamic convolution algorithm layer to perform a two-dimensional convolution operation on the input data to obtain features corresponding to the input data;

[0020] Performing a convolution operation on the features corresponding to the input data using a dynamically adjusted convolution kernel to obtain convolved features;

[0021] The convolutional features are batch normalized using a batch normalization layer, and a nonlinear factor is introduced into the batch normalized features through an activation function.

[0022] Furthermore, the attention adjustment of the first feature by the attention mechanism submodule to obtain the second feature specifically includes:

[0023] Calculate the channel attention, spatial attention and kernel attention of the first feature through the attention mechanism;

[0024] According to the channel attention, spatial attention and kernel attention, the weights of each convolution kernel in the dynamic convolution submodule are adjusted to obtain a dynamic convolution submodule with updated weights;

[0025] The first feature is convolved using the dynamic convolution submodule with updated weights to obtain the second feature.

[0026] Furthermore, an efficient channel attention module is added to the triple octave detection head network of the YOLOv8 network, specifically including:

[0027] Add the sequentially connected eleventh upsampling layer, the twelfth splicing layer, the thirteenth spatial pyramid pooling layer, the fourteenth upsampling layer, the fifteenth splicing layer, the sixteenth spatial pyramid pooling layer, the seventeenth efficient channel attention layer, the eighteenth full-dimensional convolution algorithm layer, the nineteenth splicing layer, the twentieth spatial pyramid pooling layer, the twenty-first efficient channel attention layer, the twenty-second full-dimensional convolution algorithm layer, the twenty-third splicing layer, the twenty-fourth spatial pyramid pooling layer, and the twenty-fifth efficient channel attention layer to the triple octave detection head network;

[0028] Among them, the eleventh upsampling layer is used to enlarge the feature size of the last layer of the backbone network by 2 times, and directly copy each pixel value in the feature map to the upsampled feature map; the twelfth splicing layer is used to splice the upsampled features with the output features of the seventh spatial pyramid pooling layer; the thirteenth spatial pyramid pooling layer is used to fuse features from different layers; the fourteenth upsampling layer is used to receive the output features of the thirteenth spatial pyramid pooling layer; the fifteenth splicing layer is used to splice the features output by the fourteenth upsampling layer with the output features of the fifth spatial pyramid pooling layer; the sixteenth spatial pyramid pooling layer is used to fuse features from different layers; the seventeenth efficient channel attention layer is used to adjust the channel weights of the feature map; The eighteenth full-dimensional convolution algorithm layer is used to fuse features from different layers; the nineteenth splicing layer is used to splice the output features of the full-dimensional dynamic convolution algorithm with the output features of the thirteenth spatial pyramid pooling layer; the twentieth spatial pyramid pooling layer is used to receive the output features of the splicing layer; the twenty-first efficient channel attention layer is used to adjust the channel weights of the feature map; the twenty-second full-dimensional convolution algorithm layer is used to fuse features from different layers; the twenty-third splicing layer is used to splice the output features of the twenty-second full-dimensional convolution algorithm layer with the output features of the spatial pyramid pooling module; the twenty-fourth spatial pyramid pooling layer is used to receive the output features of the twenty-third splicing layer; the twenty-fifth efficient channel attention layer is used to adjust the weights of the feature map channels.

[0029] Furthermore, the efficient channel attention module specifically includes:

[0030] Feature compression module and weight learning module;

[0031] The feature compression module is used to compress the spatial information of each feature map into a channel descriptor that reflects the overall activity level using a global average pooling operation; the weight learning module is used to capture the dependencies between channels using a convolutional layer and learn the weights of each channel to enhance or suppress specific feature responses.

[0032] In a second aspect, the present invention provides a defect detection device for a surface of a polysilicon production device, comprising:

[0033] An optimization module is used to integrate the full-dimensional dynamic convolution algorithm and the spatial pyramid pooling module into the deep octave detection network of the YOLOv8 network, and add the efficient channel attention module to the triple octave detection head network of the YOLOv8 network, and replace the intersection-over-union function in the YOLOv8 network with a weighted intersection-over-union loss function to obtain an improved YOLOv8 network;

[0034] A training module, used for inputting the surface defect images of polysilicon production equipment with annotations into the improved YOLOv8 network for training to obtain an equipment defect detection model;

[0035] The detection module is used to input the collected surface defect images of the polysilicon production equipment into the equipment defect detection model to obtain the detection results of the surface defects of the polysilicon production equipment.

[0036] The present invention provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the method for detecting defects on the surface of polysilicon production equipment is implemented.

[0037] The present invention provides a computer device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the method for detecting defects on the surface of the polysilicon production equipment is implemented when the processor executes the program.

[0038] The at least one technical solution adopted by the present invention can achieve the following beneficial effects:

[0039] The present invention applies the full-dimensional dynamic convolution algorithm module to the feature extraction network of the YOLOv8 network. For defects with complex and irregular shapes, more refined local features can be captured, and the ODConv algorithm can optimize the distribution of the convolution kernel to reduce the amount of parameters and calculations, achieve higher parameter utilization efficiency and lower computational complexity, and improve the efficiency of defect detection. In addition, the efficient channel attention (EffectiveSE) module is introduced to recalibrate these local features, emphasize important features and suppress unimportant features, improve the accuracy of defect detection, and replace the loss function with WIOU to enhance the multi-scale fusion of defect features, which can improve the detection accuracy and efficiency of surface defects of polysilicon production equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0041] Figure 1 A schematic flow chart of a defect detection method for the surface of a polysilicon production device provided by the present invention;

[0042] Figure 2 A schematic diagram of defects on the surface of the polysilicon production equipment provided by the present invention;

[0043] Figure 3 A schematic diagram of the improved equipment defect detection model architecture provided by the present invention;

[0044] Figure 4 A schematic diagram of the architecture of the full-dimensional dynamic convolution algorithm provided by the present invention;

[0045] Figure 5 A schematic diagram of the structure of an efficient channel attention module provided by the present invention;

[0046] Figure 6 A schematic diagram of a defect detection device for the surface of a polysilicon production device provided by the present invention;

[0047] Figure 7 A schematic diagram of a computer device for implementing a defect detection method on the surface of polysilicon production equipment provided by the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0049] The server mentioned in the present invention can be a server set on a business platform, or a device such as a desktop computer, a laptop computer, etc. that can execute the solution of the present invention. For the convenience of description, the following description is only based on the server as the execution subject. The following is a detailed description of the technical solutions provided by various embodiments of the present invention in conjunction with the accompanying drawings.

[0050] Figure 1 The present invention is a schematic flow chart of a defect detection method on the surface of a polysilicon production device, which specifically includes the following steps:

[0051] S10: The full-dimensional dynamic convolution algorithm and the spatial pyramid pooling module are integrated into the deep octave detection network of the YOLOv8 network, and the efficient channel attention module is added to the triple octave detection head network of the YOLOv8 network, and the intersection-over-union function in the YOLOv8 network is replaced with the weighted intersection-over-union loss function to obtain an improved YOLOv8 network.

[0052] In this embodiment, the YOLOv8 network includes an input end, a deep octave detection network, a triple octave detection head network and a prediction head. The input end is used to receive images related to surface defects of polysilicon production equipment, the deep octave detection network refers to DeepOctaveDetNet, the triple octave detection head network refers to TriOctaveDetHe-adNet, and the prediction head refers to the Head part of the equipment defect detection model, which is used to perform classification tasks.

[0053] In this embodiment, the Omni-Dimensional Dynamic Convolution (ODConv) algorithm learns attention on the four dimensions (space, input channel, output channel, number of cores) of the convolution kernel through a multi-dimensional attention mechanism, thereby enhancing the feature extraction capability. The spatial pyramid pooling module refers to the SPPFCSPC (SpatialPyramid Pooling with Feature Concatenate and Spatial Convolution) module, which uses the spatial pyramid pooling (SpatialPyramid Pooling, SPP) to capture features of different scales and enhance the expressiveness of features through splicing and convolution operations. The complete intersection-over-union function refers to the CIOU (Complete Intersection overUnion) function, and the weighted intersection-over-union loss function refers to the WIOU (WeightedIntersection overUnion) function.

[0054] S20: Input the labeled surface defect images of polysilicon production equipment into the improved YOLOv8 network for training to obtain an equipment defect detection model.

[0055] In this embodiment, surface defect images of different types of polysilicon production equipment are labeled respectively, and then the labeled surface defect images of polysilicon production equipment are input into the improved YOLOv8 network for training to obtain an equipment defect detection model.

[0056] S30: Inputting the collected surface defect image of the polysilicon production equipment into the equipment defect detection model to obtain the detection result of the surface defects of the polysilicon production equipment.

[0057] In this embodiment, the detection results include but are not limited to: defect type and defect location.

[0058] Among them, defect categories include but are not limited to: atmospheric corrosion, water corrosion, high temperature corrosion and physical wear.

[0059] Schematically, images related to surface defects of polysilicon production equipment are shown in Figure 2 shown.

[0060] based on Figure 1 The defect detection method for the surface of polysilicon production equipment shown in the figure applies the Omni-Dimensional Dynamic Convolution (ODConv) module to the feature extraction network of the YOLOv8 network. For defects with complex and irregular shapes, more refined local features can be captured, and the ODConv algorithm can optimize the distribution of convolution kernels to reduce the amount of parameters and calculations, achieve higher parameter utilization efficiency and lower computational complexity, and improve the detection efficiency of defects. In addition, the Effective SE module is introduced to recalibrate these local features, emphasize important features and suppress unimportant features, improve the accuracy of defect detection, and replace the loss function with WIOU to enhance the multi-scale fusion of defect features, which can improve the detection accuracy and efficiency of surface defects of polysilicon production equipment.

[0061] When applying the defect detection method for the surface of polysilicon production equipment provided by the present invention, it is not necessary to Figure 1 The steps are executed in the order shown. The specific execution order of the steps can be determined according to needs, and the present invention does not limit this.

[0062] In addition, in one or more embodiments of the present invention, the full-dimensional dynamic convolution algorithm and the spatial pyramid pooling module are integrated into the deep octave detection network of the YOLOv8 network, specifically including:

[0063] The first full-dimensional dynamic convolution algorithm layer, the second full-dimensional dynamic convolution algorithm layer, the third spatial pyramid pooling layer, the fourth full-dimensional dynamic convolution algorithm layer, the fifth spatial pyramid pooling layer, the sixth full-dimensional dynamic convolution algorithm layer, the seventh spatial pyramid pooling layer, the eighth full-dimensional dynamic convolution algorithm layer, the ninth spatial pyramid pooling layer, the tenth spatial pyramid pooling and feature splicing and spatial convolution layer, which are connected sequentially, are integrated into the deep octave detection network.

[0064] In this embodiment, reference Figure 3, the structure of the deep octave detection network is shown on the far left, including the sequentially connected input layer, ODConv (first full-dimensional dynamic convolution algorithm layer), ODConv (second full-dimensional dynamic convolution algorithm layer), C2f (third spatial pyramid pooling layer), ODConv (fourth full-dimensional dynamic convolution algorithm layer), C2f (fifth spatial pyramid pooling layer), ODConv (sixth full-dimensional dynamic convolution algorithm layer), C2f (seventh spatial pyramid pooling layer), ODConv (eighth full-dimensional dynamic convolution algorithm layer), C2f (ninth spatial pyramid pooling layer), SPPFCSPC (tenth spatial pyramid pooling and feature concatenation and spatial convolution layer (Spatial Pyramid Pooling with FeatureConcatenate and Spatial Convolution, SPPFCSPC)).

[0065] Among them, the first full-dimensional dynamic convolution algorithm layer accepts a four-dimensional tensor as input, whose shape is usually (batch_size, 3, height, width), the number of output channels is 32, the convolution kernel size is 3x3, the stride is 2, and the shape of the output feature is (batch_size, 32, height / 2, width / 2). The second full-dimensional dynamic convolution algorithm layer receives the features of the first full-dimensional dynamic convolution algorithm layer, the number of output channels is 64, the convolution kernel size is 3x3, the stride is 2, and the shape of the output feature is (batch_size, 64, height / 4, width / 4). The third spatial pyramid pooling layer receives the features of the second full-dimensional dynamic convolution algorithm layer, repeats 3 times, the number of output channels is 64, uses the attention mechanism, and the shape of the output feature is (batch_size, 64, height / 4, width / 4). The fourth full-dimensional dynamic convolution algorithm layer receives the features of the third spatial pyramid pooling layer, with an output channel number of 128, a convolution kernel size of 3x3, a stride of 2, and an output feature shape of (batch_size, 128, height / 8, width / 8). The fifth spatial pyramid pooling layer receives the features of the fourth full-dimensional dynamic convolution algorithm layer, repeats 6 times, has an output channel number of 128, uses an attention mechanism, and has an output feature shape of (batch_size, 128, height / 8, width / 8). The sixth full-dimensional dynamic convolution algorithm layer receives the features of the fifth spatial pyramid pooling layer, has an output channel number of 256, a convolution kernel size of 3x3, a stride of 2, and has an output feature shape of (batch_size, 256, height / 16, width / 16). The seventh spatial pyramid pooling layer receives the features of the sixth full-dimensional dynamic convolution algorithm layer, repeats 6 times, has an output channel number of 256, uses an attention mechanism, and has an output feature shape of (batch_size, 256, height / 16, width / 16). The eighth full-dimensional dynamic convolution algorithm layer receives the features of the seventh spatial pyramid pooling layer, with an output channel number of 512, a convolution kernel size of 3x3, a stride of 2, and an output feature shape of (batch_size, 512, height / 32, width / 32). The ninth spatial pyramid pooling layer is used to receive the features of the eighth full-dimensional dynamic convolution algorithm layer, repeated 3 times, with an output channel number of 512, using the attention mechanism, and an output feature shape of (batch_size, 512, height / 32, width / 32). The tenth spatial pyramid pooling and feature splicing and spatial convolution layer receives the features from the ninth spatial pyramid pooling layer, with an output channel number of 512, and an output feature shape of (batch_size, 512, height / 32, width / 32).

[0066] The scheme shown in this embodiment, by introducing the ODConv module in the deep octave detection network and introducing an offset in the standard convolution, can adapt to defects of different shapes and more accurately locate the positions of tiny defects in polysilicon production equipment. This module also reduces the number of parameters and calculations by optimizing the distribution of convolution kernels, achieving higher parameter utilization efficiency and lower computational complexity. In addition, by replacing the SPPF module with SPPFCSPC, through the splicing of features at different scales, it can provide rich contextual information, which helps the model to better utilize multi-scale information and improve the accuracy and robustness of detection.

[0067] In addition, in one or more embodiments of the present invention, the full-dimensional dynamic convolution algorithm integrates the dynamic convolution submodule and the attention mechanism submodule to:

[0068] The input data is convolved through the dynamic convolution submodule to obtain the first feature.

[0069] The attention mechanism submodule is used to adjust the attention of the first feature to obtain the second feature.

[0070] Perform a weighted operation on the second feature to obtain the output feature corresponding to the input data.

[0071] In this embodiment, reference Figure 4 , the ODConv layer combines dynamic convolution and attention mechanisms, divides the input into two parts: ordinary convolution and attention adjustment for feature extraction, and finally performs weighted operations to obtain the final output. Normalization and activation function layers are used before and after the convolution operation. The ODConv structure includes a convolution layer, a batch normalization layer, and a SiLU activation function layer connected in sequence. The input features are first convolved with ODConv2d, mainly adjusted by the attention module, and then convolved with the dynamically adjusted convolution kernel. The convolved features are batch normalized through the normalization layer to stabilize the values ​​during training. Finally, the features are activated through the SiLU function to introduce nonlinearity and enhance the expressiveness of the model. Among them, the relationship between ODConv and ODConv2d is:

[0072] ODConv is a sequence model that connects the ODConv2d layer, batch normalization layer (norm_layer) and activation function (nn.SiLU) in series. This structure is a higher-level abstraction that provides encapsulated modules that can be directly inserted into the neural network model for feature extraction. ODConv2d is a lower-level implementation. "2d" refers to the fact that this class performs a two-dimensional convolution operation, focusing on how to implement the convolution operation through the attention mechanism and expand and enhance it.

[0073] In addition, in one or more embodiments of the present invention, before the dynamic convolution submodule performs a convolution operation on the input data, and after the dynamic convolution submodule performs a convolution operation on the input data, it also includes:

[0074] Use the full-dimensional dynamic convolution algorithm layer to perform a two-dimensional convolution operation on the input data to obtain the features corresponding to the input data.

[0075] The dynamically adjusted convolution kernel is used to perform a convolution operation on the features corresponding to the input data to obtain the convolved features.

[0076] The batch normalization layer is used to batch normalize the convolutional features, and the activation function is used to introduce nonlinear factors into the batch normalized features.

[0077] In this embodiment, reference Figure 4 , normalization and activation function layers are used before and after the convolution operation. The structure includes a convolution layer, a batch normalization layer, and a SiLU activation function layer connected in sequence. The input features are first convolved with ODConv2d, mainly adjusted by the attention module, and then convolved with the dynamically adjusted convolution kernel. The convolved features are batch normalized through the normalization layer to stabilize the values ​​during training. Finally, the features are activated by the SiLU function to introduce nonlinearity and enhance the expressiveness of the model.

[0078] In addition, in one or more embodiments of the present invention, the attention adjustment is performed on the first feature by the attention mechanism submodule to obtain the second feature, which specifically includes:

[0079] The channel attention, spatial attention and kernel attention of the first feature are calculated through the attention mechanism.

[0080] According to channel attention, spatial attention and kernel attention, the weights of each convolution kernel in the dynamic convolution submodule are adjusted to obtain a dynamic convolution submodule with updated weights.

[0081] The dynamic convolution submodule with updated weights is used to perform a convolution operation on the first feature to obtain the second feature.

[0082] In this embodiment, ODConv2d mainly introduces the attention mechanism to dynamically adjust the weight of the convolution kernel, thereby improving the feature extraction ability of the model. The first step is to calculate the channel attention, filter attention, spatial attention and kernel attention of the input features through the attention mechanism. These attentions are used to dynamically adjust the weight of the convolution kernel. The formula can be expressed as:

[0083] Attention=Func(Input) (1)

[0084] Among them, Attention represents attention, Func represents attention calculation function, and Input represents input features.

[0085] The second step of dynamic convolution is to weight the input features according to the attention map, and then perform convolution operation with the dynamically adjusted convolution kernel. The formula can be expressed as:

[0086] Output=Input*Weight (2)

[0087] Among them, Output represents the convolution result, * represents the convolution operation, and Weight is the convolution kernel weight adjusted according to attention.

[0088] The third step is attention adjustment, which is to weight the feature map obtained by ordinary convolution through spatial attention, kernel attention and filter attention to obtain the final output features.

[0089] In order to enhance the expressive power of the model, the traditional convolution is divided into two steps. First, the importance of the input features is calculated through the attention mechanism, and then a dynamic convolution operation is performed on this basis. Specifically, the calculation of the first step of the attention mechanism can be expressed as:

[0090] ChannelAttention=Sigmoid(FC(Input)) (3)

[0091] FilterAttention=Sigmoid(FC(Input)) (4)

[0092] SpatialAttention=Sigmoid(FC(Input)) (5)

[0093] KernelAttention=Softmax(FC(Input)) (6)

[0094] Among them, ChannelAttention refers to channel attention, which is a module used to adjust the importance of each channel in the feature map. FolterAttentoon refers to filter attention, which is a module used to adjust the importance of different filters (or convolution kernels). SpatialAttention refers to spatial attention, which is a module used to emphasize the spatial position in the feature map, that is, which areas are more important for subsequent processing. KernelAttention refers to kernel attention, which is a module used to assign weights to different convolution kernels, indicating their relative importance in the convolution operation. FC refers to the fully connected layer, which is used to perform weighted summation of the input. Sigmoid is an activation function used to compress the input between 0 and 1. Softmax is an activation function that converts the input into a probability distribution in which the value of each element is between 0 and 1 and the sum of all elements is 1.

[0095] The second step dynamic convolution operation formula is:

[0096] Output=Input*(SpatialAttention*KernelAttention*Weight)(7)

[0097] The third step of attention adjustment operation formula is:

[0098] Final Output=Output*(SpatialAttention*KernelAttention*

[0099] FilterAttention)(8)

[0100] Among them, Final Output refers to the final output, which is the final result obtained after the dynamic convolution operation and the attention adjustment operation.

[0101] Optionally, SPPFCSPC uses spatial pyramid pooling (SPP) to capture features of different scales, and enhances the expressiveness of features through splicing and convolution operations. First, the upper layer input features are mapped to the middle channel number through a 1x1 convolution layer, and then a convolution operation is performed through a 3x3 convolution layer. The spatial pyramid pooling operation is performed through SPP, and different sizes of pooling kernels are applied to different areas of the feature map to capture features of different scales. The pooled features are mapped back to the middle channel number through a 1x1 convolution layer, and a convolution operation is performed through a 3x3 convolution layer. The features processed by SPP are spliced ​​with the original features, and the spliced ​​features are mapped to the output channel number through a 1x1 convolution layer to obtain the final output features.

[0102] In addition, in one or more embodiments of the present invention, an efficient channel attention module is added to the triple octave detection head network of the YOLOv8 network, specifically including:

[0103] The eleventh upsampling layer, the twelfth splicing layer, the thirteenth spatial pyramid pooling layer, the fourteenth upsampling layer, the fifteenth splicing layer, the sixteenth spatial pyramid pooling layer, the seventeenth efficient channel attention layer, the eighteenth full-dimensional convolution algorithm layer, the nineteenth splicing layer, the twentieth spatial pyramid pooling layer, the twenty-first efficient channel attention layer, the twenty-second full-dimensional convolution algorithm layer, the twenty-third splicing layer, the twenty-fourth spatial pyramid pooling layer, and the twenty-fifth efficient channel attention layer are added to the triple octave detection head network.

[0104] In this embodiment, reference Figure 3The middle part shows the structural diagram of the triple octave detection head network (TriOctaveDetHeadNet), which mainly includes the eleventh upsampling layer (UpSample), which enlarges the feature size of the last layer from the backbone by 2 times and directly copies each pixel value in the feature map to the upsampled feature map. The twelfth splicing layer (Concat) splices the upsampled features with the output features of the seventh spatial pyramid pooling layer from the backbone. The thirteenth spatial pyramid pooling layer (C2f) is used to fuse features from different layers to enhance the expressiveness of the features, and the number of output feature channels is 768. The fourteenth upsampling layer (UpSample) receives the output features of the previous layer and repeats the same operation as the eleventh layer. The fifteenth splicing layer (Concat) splices the features output by the fourteenth upsampling layer with the output features of the fifth layer from the backbone. The sixteenth spatial pyramid pooling layer (C2f) repeats operations similar to the thirteenth layer (C2f), and the number of channels becomes 128 after the C2f layer. The seventeenth efficient channel attention layer (EffectiveSE) adjusts the channel weight operation of the feature map to enhance the features. The eighteenth full-dimensional convolution algorithm layer (ODConv) includes a convolution layer, an attention layer, and a splicing layer, which is used to fuse features from different layers to enhance the expressiveness of the features. The number of output channels is 128, the convolution kernel is 3*3, and the step size is 2. The nineteenth splicing layer (Concat) is used to splice the output features of ODConv with the output features of the thirteenth spatial pyramid pooling layer. The twentieth spatial pyramid pooling layer (C2f) receives the output features of the splicing layer, repeats 3 times, and the number of output channels is 256. The twenty-first efficient channel attention layer (EffectiveSE) repeats the channel weight operation of the feature map to enhance the features. The twenty-second full-dimensional convolution algorithm layer (ODConv) repeats the operation of the eighteenth full-dimensional convolution algorithm layer, and the number of output channels is 256. The twenty-third splicing layer (Concat) splices the output features of the previous layer with the output features of the SPPFCSPC layer. The twenty-fourth spatial pyramid pooling layer (C2f) continues to receive the output features of the upper Concat layer, repeating 3 times, and the number of output channels is 512. The twenty-fifth efficient channel attention layer (EffectiveSE) repeats the feature map channel weight adjustment and performs feature enhancement operations.

[0105] In addition, in one or more embodiments of the present invention, the efficient channel attention module specifically includes:

[0106] Feature compression module and weight learning module.

[0107] Among them, the feature compression module is used to compress the spatial information of each feature map into a channel descriptor that reflects the overall activity level using a global average pooling operation; the weight learning module is used to capture the dependencies between channels using a convolutional layer and learn the weights of each channel to enhance or suppress specific feature responses.

[0108] In this embodiment, reference Figure 5 , the efficient channel attention module is used to enhance the model's attention to key features and effectively improve the model's expressiveness. It is divided into two parts. The Squeeze part uses a global average pooling operation to compress the spatial information of each feature map into a channel descriptor, reflecting the overall activity level of the channel. The Excitation part captures the dependencies between channels through a convolutional layer (fully connected layer) and learns the weights of each channel to enhance or suppress specific feature responses. The mechanism is simple in design and easy to integrate into different CNN architectures, especially in the Head part of YOLOv8, to improve the representation ability of features.

[0109] It should be noted that the specific fusion process of the deep octave detection network and the triple octave detection head network in the present invention is as follows:

[0110] (a) M1 feature map and M2 feature map after upsampling module Figure 1 The input concatenation layer performs dimension concatenation operation to obtain feature map T2.

[0111] (b) Feature map T2 after passing through the C2f module and the upsampling module and M3 feature map Figure 1 The input concatenation layer performs dimension concatenation operation to obtain feature map T3.

[0112] (c) The feature map T2 and T3 after the C2f module, EffectiveSE module and downsampling module are input into the concatenation layer for dimensional concatenation operation to obtain the feature map T4.

[0113] (d) The feature map M1 and T4 after the C2f module, EffectiveSE module and downsampling module are input into the concatenation layer for dimensional concatenation operation to obtain the feature map T1.

[0114] (e) Feature map R3 is obtained by passing feature map T3 through the C2f module and the EffectiveSE module; feature map R2 is obtained by passing feature map T4 through the C2f module and the EffectiveSE module; feature map R1 is obtained by passing feature map T1 through the C2f module and EffectiveSE.

[0115] The specific mathematical expression is:

[0116] R 3 =C2F(T3 ) (9)

[0117] R 2 =C2F(T 4 ) (10)

[0118] R 1 =C2F(T 1 ) (11)

[0119] T 3 =Concat{M 3 ,upasmpling[C2F(T 2 )]} (12)

[0120] T 2 =Concat{M 2 ,upasmpling(M 1 )} (13)

[0121] T 4 =Concat{doWnasmpling[C2F(T 3 )],C2F(T 2 )} (14)

[0122] T 1 =Concat{downasmpling[C2F(T 4 )],M 1} (15)

[0123] In the formula, the meaning of each parameter is as follows:

[0124] M1: After being processed by the tenth SPPFCSPC layer, the output feature map is 20x20 in size and 512 in number of channels.

[0125] M2: After processing by the seventh C2F layer, a feature map with a size of 40x40 and 256 channels is obtained.

[0126] M3: After the fifth C2F layer, the output feature map is 80x80 in size and 128 in number of channels.

[0127] T1: After the twenty-third Concat layer, the output feature map has a size of 20*20 and 768 channels.

[0128] T2: After the twelfth Concat layer, the output feature map has a size of 40*40 and 768 channels.

[0129] T3: After the fifteenth Concat layer, the output feature map has a size of 80*80 and 384 channels.

[0130] T4: After the nineteenth Concat layer, the output feature map has a size of 40*40 and 384 channels.

[0131] R1: Outputs a feature map of size 20x20 with 512 channels for detecting small objects.

[0132] R2: Outputs a feature map of size 40x40 with 256 channels for detecting medium-sized objects.

[0133] R3: Outputs a feature map of size 80x80 with 128 channels for detecting large objects. Concat represents the dimension concatenation operation corresponding to the concatenation layer, upsampling represents the upsampling operation corresponding to the upsampling module, C2f represents the function corresponding to the C2f module, and downsampling represents the downsampling operation corresponding to the downsampling module.

[0134] In a specific embodiment, in terms of category classification loss, VFL Loss is used as the classification loss. VFL Loss measures the difference between the probability distribution predicted by the model and the actual label. The formula is as follows:

[0135]

[0136] Where b is the Intersection over Union (IoU) between the bounding box (bbox) and the groundtruth (gt), and a is the probability. When b is a positive sample, the classification loss is the Binary Cross-Entropy Loss (BCE) loss, and when it is a negative sample, it is FL. Where bbox is a rectangular box used to indicate the location of an object in an image. gt contains the actual location of the object (usually in the form of a bounding box) and category information.

[0137] In a specific embodiment, in terms of border regression loss, DFL Loss+WIOU Loss is used as the border regression loss. Specifically, after YOLOv8 introduces the Center-based methods (based on the center point) idea of ​​Anchor-Free, the model changes from outputting offest to ltrb=left, top, right, bottom, that is, the distance from the four borders of the predicted box to the center point. In order to cooperate with Anchor-Free and improve generalization, DFL loss is added in YOLOv8, and its formula is as follows:

[0138] L WIoUv1 =R WIoU L IoU(17)

[0139]

[0140] Among them, Wg and Hg are the sizes of the smallest enclosing boxes. In order to prevent the generation of gradients that hinder convergence, Wg and Hg are separated from the computational graph (the superscript * indicates this operation). Because it effectively eliminates factors that hinder convergence, no new metrics such as aspect ratio are introduced. Therefore, WIoU solves the problem of CIoU when the aspect ratio is the same.

[0141] After calculating ltrb, DFL uses cross entropy to allow the network to focus on the target location and the distribution of neighboring areas more quickly.

[0142] Since the training data necessarily includes low-quality examples, and geometric factors such as distance and aspect ratio will aggravate the impact of such low-quality examples, thereby reducing the generalization performance of the model. A good loss function should reduce the penalty of geometric factors when the anchor box and the true box basically overlap, and less intervention during training can enable the model to have better generalization ability. Therefore, the distance attention is constructed based on the distance metric, and WIoUv1 with a two-layer attention mechanism is obtained, and its formula is as follows:

[0143]

[0144] Wise-IoUv3 introduces a dynamic non-monotonic attention mechanism to replace the static monotonic attention mechanism. This mechanism uses "outliers" instead of IoU to evaluate the quality of anchor boxes, and implements a wise gradient gain allocation strategy. This strategy can reduce the competitiveness of high-quality anchor boxes and reduce the harmful gradients generated by low-quality samples, so that WIoU can give priority to anchor boxes of ordinary quality and improve detector performance. The formula is as follows:

[0145] Using β to construct a non-monotonic focusing coefficient and apply it to WIoU v1, we obtain WIoU v3 with dynamic non-monotonic FM, which can dynamically optimize the loss to weight small objects to improve the detection performance of the model.

[0146] The surface defect detection model for polysilicon production metal equipment uses the SGD optimizer and is trained with a parameter setting of a learning rate of 0.01. The training data is divided into a training set, a validation set, and a test set according to a ratio of 8:1:1. The equipment defect images in the training set are input into the equipment defect detection model, and the trained equipment defect detection model is deployed and applied to realize the detection of equipment defects.

[0147] The above is a method for detecting defects on the surface of a polysilicon production equipment provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding device for detecting defects on the surface of a polysilicon production equipment, such as Figure 6 shown.

[0148] Figure 6 A schematic diagram of a defect detection device for the surface of a polysilicon production device provided by the present invention comprises:

[0149] An optimization module is used to integrate the full-dimensional dynamic convolution algorithm and the spatial pyramid pooling module into the deep octave detection network of the YOLOv8 network, and add the efficient channel attention module to the triple octave detection head network of the YOLOv8 network, and replace the intersection-over-union function in the YOLOv8 network with a weighted intersection-over-union loss function to obtain an improved YOLOv8 network;

[0150] A training module, used for inputting the surface defect images of polysilicon production equipment with annotations into the improved YOLOv8 network for training to obtain an equipment defect detection model;

[0151] The detection module is used to input the collected surface defect images of the polysilicon production equipment into the equipment defect detection model to obtain the detection results of the surface defects of the polysilicon production equipment.

[0152] For the specific definition of the defect detection device on the surface of polysilicon production equipment, please refer to the definition of the defect detection method on the surface of polysilicon production equipment mentioned above, which will not be repeated here. Each module in the defect detection device on the surface of polysilicon production equipment can be implemented in whole or in part by software, hardware and a combination thereof. Each module can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0153] The present invention also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the Figure 1 A method for detecting defects on the surface of polysilicon production equipment is provided.

[0154] The present invention also provides Figure 7 The structural diagram of the computer device shown in FIG. Figure 7 As shown, at the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory, and may also include other hardware required for the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the Figure 1 A method for detecting defects on the surface of polysilicon production equipment is provided.

[0155] Those skilled in the art can understand that all or part of the processes in the embodiment method can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0156] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of the present invention.

Claims

1. A method for detecting defects on the surface of polysilicon production equipment, characterized in that: include: The full-dimensional dynamic convolution algorithm and the spatial pyramid pooling module are integrated into the deep octave detection network of the YOLOv8 network, and the efficient channel attention module is added to the triple octave detection head network of the YOLOv8 network, and the intersection-over-union function in the YOLOv8 network is replaced with a weighted intersection-over-union loss function to obtain an improved YOLOv8 network; Inputting the annotated surface defect image of the polysilicon production equipment into the improved YOLOv8 network for training to obtain an equipment defect detection model; The collected surface defect images of the polysilicon production equipment are input into the equipment defect detection model to obtain the detection results of the surface defects of the polysilicon production equipment.

2. The method for detecting defects on the surface of polysilicon production equipment according to claim 1, characterized in that: The integration of the full-dimensional dynamic convolution algorithm and the spatial pyramid pooling module into the deep octave detection network of the YOLOv8 network specifically includes: Integrate the first full-dimensional dynamic convolution algorithm layer, the second full-dimensional dynamic convolution algorithm layer, the third spatial pyramid pooling layer, the fourth full-dimensional dynamic convolution algorithm layer, the fifth spatial pyramid pooling layer, the sixth full-dimensional dynamic convolution algorithm layer, the seventh spatial pyramid pooling layer, the eighth full-dimensional dynamic convolution algorithm layer, the ninth spatial pyramid pooling layer, the tenth spatial pyramid pooling and feature splicing and spatial convolution layer connected in sequence into the deep octave detection network; Among them, the first full-dimensional dynamic convolution algorithm layer is used to accept a four-dimensional tensor as input, the second full-dimensional dynamic convolution algorithm layer is used to receive the features of the first full-dimensional dynamic convolution algorithm layer, the third spatial pyramid pooling layer is used to receive the features of the second full-dimensional dynamic convolution algorithm layer, the fourth full-dimensional dynamic convolution algorithm layer is used to receive the features of the third spatial pyramid pooling layer, the fifth spatial pyramid pooling layer is used to receive the features of the fourth full-dimensional dynamic convolution algorithm layer, the sixth full-dimensional dynamic convolution algorithm layer is used to receive the features of the fifth spatial pyramid pooling layer, the seventh spatial pyramid pooling layer is used to receive the features of the sixth full-dimensional dynamic convolution algorithm layer, the eighth full-dimensional dynamic convolution algorithm layer is used to receive the features of the seventh spatial pyramid pooling layer, the ninth spatial pyramid pooling layer is used to receive the features of the eighth full-dimensional dynamic convolution algorithm layer, and the tenth spatial pyramid pooling and feature splicing and spatial convolution layer is used to receive the features of the ninth spatial pyramid pooling layer.

3. The method for detecting defects on the surface of polysilicon production equipment according to claim 1, characterized in that: The full-dimensional dynamic convolution algorithm combines the dynamic convolution submodule and the attention mechanism submodule to: Performing a convolution operation on the input data through the dynamic convolution submodule to obtain a first feature; Attention is adjusted on the first feature by the attention mechanism submodule to obtain a second feature; A weighted operation is performed on the second feature to obtain an output feature corresponding to the input data.

4. The method for detecting defects on the surface of polysilicon production equipment according to claim 3, characterized in that: Before the dynamic convolution submodule performs a convolution operation on the input data, and after the dynamic convolution submodule performs a convolution operation on the input data, the method further includes: Using the full-dimensional dynamic convolution algorithm layer to perform a two-dimensional convolution operation on the input data to obtain features corresponding to the input data; Performing a convolution operation on the features corresponding to the input data using a dynamically adjusted convolution kernel to obtain convolved features; The convolutional features are batch normalized using a batch normalization layer, and a nonlinear factor is introduced into the batch normalized features through an activation function.

5. The method for detecting defects on the surface of polysilicon production equipment according to claim 3, characterized in that: The attention adjustment of the first feature by the attention mechanism submodule to obtain the second feature specifically includes: Calculate the channel attention, spatial attention and kernel attention of the first feature through the attention mechanism; According to the channel attention, spatial attention and kernel attention, the weights of each convolution kernel in the dynamic convolution submodule are adjusted to obtain a dynamic convolution submodule with updated weights; The first feature is convolved using the dynamic convolution submodule with updated weights to obtain the second feature.

6. The method for detecting defects on the surface of polysilicon production equipment according to claim 2, characterized in that: The efficient channel attention module is added to the triple octave detection head network of the YOLOv8 network, specifically including: Add the sequentially connected eleventh upsampling layer, the twelfth splicing layer, the thirteenth spatial pyramid pooling layer, the fourteenth upsampling layer, the fifteenth splicing layer, the sixteenth spatial pyramid pooling layer, the seventeenth efficient channel attention layer, the eighteenth full-dimensional convolution algorithm layer, the nineteenth splicing layer, the twentieth spatial pyramid pooling layer, the twenty-first efficient channel attention layer, the twenty-second full-dimensional convolution algorithm layer, the twenty-third splicing layer, the twenty-fourth spatial pyramid pooling layer, and the twenty-fifth efficient channel attention layer to the triple octave detection head network; Among them, the eleventh upsampling layer is used to enlarge the feature size of the last layer of the backbone network by 2 times, and directly copy each pixel value in the feature map to the upsampled feature map; the twelfth splicing layer is used to splice the upsampled features with the output features of the seventh spatial pyramid pooling layer; the thirteenth spatial pyramid pooling layer is used to fuse features from different layers; the fourteenth upsampling layer is used to receive the output features of the thirteenth spatial pyramid pooling layer; the fifteenth splicing layer is used to splice the features output by the fourteenth upsampling layer with the output features of the fifth spatial pyramid pooling layer; the sixteenth spatial pyramid pooling layer is used to fuse features from different layers; the seventeenth efficient channel attention layer is used to adjust the channel weights of the feature map; The eighteenth full-dimensional convolution algorithm layer is used to fuse features from different layers; the nineteenth splicing layer is used to splice the output features of the full-dimensional dynamic convolution algorithm with the output features of the thirteenth spatial pyramid pooling layer; the twentieth spatial pyramid pooling layer is used to receive the output features of the splicing layer; the twenty-first efficient channel attention layer is used to adjust the channel weights of the feature map; the twenty-second full-dimensional convolution algorithm layer is used to fuse features from different layers; the twenty-third splicing layer is used to splice the output features of the twenty-second full-dimensional convolution algorithm layer with the output features of the spatial pyramid pooling module; the twenty-fourth spatial pyramid pooling layer is used to receive the output features of the twenty-third splicing layer; the twenty-fifth efficient channel attention layer is used to adjust the weights of the feature map channels.

7. The method for detecting defects on the surface of polysilicon production equipment according to any one of claims 1 to 6, characterized in that: The efficient channel attention module specifically includes: Feature compression module and weight learning module; The feature compression module is used to compress the spatial information of each feature map into a channel descriptor that reflects the overall activity level using a global average pooling operation; the weight learning module is used to capture the dependencies between channels using a convolutional layer and learn the weights of each channel to enhance or suppress specific feature responses.

8. A defect detection device for the surface of a polysilicon production equipment, characterized in that: include: An optimization module is used to integrate the full-dimensional dynamic convolution algorithm and the spatial pyramid pooling module into the deep octave detection network of the YOLOv8 network, and add the efficient channel attention module to the triple octave detection head network of the YOLOv8 network, and replace the intersection-over-union function in the YOLOv8 network with a weighted intersection-over-union loss function to obtain an improved YOLOv8 network; A training module, used for inputting the surface defect images of polysilicon production equipment with annotations into the improved YOLOv8 network for training to obtain an equipment defect detection model; The detection module is used to input the collected surface defect images of the polysilicon production equipment into the equipment defect detection model to obtain the detection results of the surface defects of the polysilicon production equipment.

9. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method for detecting defects on the surface of polysilicon production equipment according to any one of claims 1 to 7 is implemented.

10. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the method for detecting defects on the surface of polysilicon production equipment as claimed in any one of claims 1 to 7 is implemented.

Images