Defect Detection On EL Images Of Photovoltaic Cells With Complex Surfaces Based On Convolutional Neural Networks

Crack defects in photovoltaic cells can cause the failure of the cells.It also affects the normal use of the cell module,and are not conducive to maintaining the stability of the photovoltaic power generation system.The features such as non-uniform complex surface,the low contrast between randomly distributed background grains and defective targets,and the different shapes and scale of defects in the surface of photovoltaic cells in electroluminescence image,bring great challenge to the traditional method for accurate and robust detection of crack defects.Three different deep learning models are proposed in this paper,which effectively solve the effect of complex backgrounds on crack defects.Finally,a set of effective intelligent detection schemes are formed for crack defection in non-uniform surface.This paper has gradually realized image-level classification and pixels-level segmentation.The specific research contents and contributions are as follows:(1)A crack defect inspection algorithm for polycrystalline silicon solar cells based on a kind of structural decoupling convolutional neural network is proposed.The algorithm integrates the steerable evidence filter in the first convolutional layer of the basic AlexNet to form an end-to-end model.At the end of the model,the class activation map is used to verify the feature extraction capability of the model.The experimental results show that the proposed model improves the accuracy of crack defect classification and significantly improves the feature extraction capability of the basic CNN model.Experiments also show that the integration of steerable evidence filter and CNN has given CNN structural decoupling function,which weakens the interference of complex backgrounds on defective structures,and improves the robustness of defect recognition.The algorithm proposed in this paper solves the problem of insufficient generalization ability of traditional filter algorithms in the face of large amounts of data,and is expected to provide ideas for the application of other traditional filters in deep learning.(2)In order to further realize pixel-level localization and segmentation of crack defects in polycrystalline silicon solar cells under non-uniform texture background,a hierarchically rich feature defect detection algorithm based on residual network is constructed in this paper.The residual network solves the degradation problem of model training.In this paper,the residual network is divided into five stages according to the pooling layer.In the last convolution layer in each stage,a side output layer is connected,and all the side output feature maps are fused to get the final fusion prediction result.The proposed model realizes the holistic image to image prediction.And it makes full use of multi-level and multi-scale convolution features.However,the problem of missing detection or false detection of crack defects exists.The feature extraction capability and detection performance of the model are expected to be further improved.(3)In response to the above problems,this paper encapsulates the convolutional features into a richer and more robust expression method to propose an RRCF framework structure.It masterly uses multi-level and multi-scale convolutional features by integrating more higher-level features and reducing fusion of lower-level features,and significantly improved feature representation capabilities.This paper also designs a novel precise loss function that combines the weighted cross-entropy loss function with the dice loss function to achieve accurate prediction from the image level to the pixel level.The precise loss function solves the problem that the prediction results are more biased towards the background pixels due to the uneven distribution of pixels(mostly non-cracked pixels).Both subjective analysis and objective experimental results show that the proposed model is very effective for solving the problem of crack segmentation under complex background interference.And the proposed loss function is helpful for the precise prediction of cracks.