Yolo Algorithm-based Monocrystalline Silicon Crystal Line Inspection System

The booming development of deep learning has not only made a great change in our daily life,but also greatly accelerated the automation and intelligence of industrial production.As an important raw material for the photovoltaic industry,monocrystalline silicon occupies a crucial position in the development of solar energy,chips,and other fields.Industrial production requires the production process of monocrystalline silicon to ensure defect-free crystals and a low misalignment rate.By analyzing the characteristics of the monocrystalline silicon production process,combining deep learning algorithms and the detection of macroscopic features of monocrystalline silicon single crystal state,through online real-time detection of macroscopic features of single crystal state and indirect detection of its microscopic quality,can effectively improve the efficiency of monocrystalline silicon production and promote the development of industrial automation.In this paper,we focus on the inspection of crystal line in the process of single-crystal silicon production.By analyzing the image features of single crystal silicon production process,the data taken by industrial camera are truncated and the picture data containing the target area are selected.Then the image data are cropped and enhanced,image scaled and data widened,and finally the single crystal silicon crystal line dataset is made by Labelimg manual annotation.The dataset contains two types of data,crystallization process and isometric growth process,with a total of 3193 images.By creating a crystal line dataset and further optimizing it based on Yolov4-tiny and Yolo X-tiny models,we propose a lightweight inspection model that can increase the inspection speed and further reduce the number of model parameters while maintaining the inspection accuracy and implementing it in an embedded hardware system.The contents are as follows:1.Further optimization based on Yolov4-tiny models,the algorithm is improved by combining the Inception module with the feature enhancement module,improving the Neck feature enhancement network to improve the processing ability of the model for feature maps,replacing the depthwise separable convolution to further reduce the number of model parameters.The model m AP is increased from 68.70% to 98.43%,and the number of parameters is reduced from 5.87 M to 2.05 M.The detection speed FPS of the model only decreases from 65.30 to 55.25,which can still meet the requirement of real-time detection.2.For the Yolo X-tiny model,the number of parameters is further reduced by replacing the Shuffer Net V2 lightweight backbone network and replacing the depthwise separable convolution without affecting the accuracy,applying the h-swish activation function to accelerate the computation,and applying the attention mechanism to let the model focus on more important feature information.The m AP of the improved model is increased from 97.50% to 99.15%,the FPS is increased from45.34 to 56.51,and the number of parameters is reduced from 5.03 M to 2.45 M.The performance of our improved model is verified by comparing with Nano Det,Efficient Det-d0,and Yolov7-tiny models on the single-crystal silicon wafer dataset.3.Build the environment on the Jetson Nano development board and conduct improved model performance tests to verify the superior performance of the proposed model on the embedded development board.With guaranteed detection accuracy,the detection speed of the crystallization process reaches 12 FPS on the development board,and the detection speed of the isometric process reaches 8 FPS,which meets the actual needs of the plant.By analyzing the actual process of single-crystal silicon production,the single-crystal silicon crystal line detection system is designed on this basis,and the graphical interface is developed through Py QT5 to realize the display of detection results and the selection of parameters,and the number of detected crystal line targets is counted at regular intervals,which can judge the current growth status of the crystal column in a timely manner.The experimental results show that the improved algorithm model proposed in this paper has an excellent performance in terms of detection speed and detection accuracy on the single crystal silicon crystal line data set,and the single crystal silicon crystal line detection system can basically meet the needs of industrial production.