Monocrystalline silicon crystal line detection based on the improved YoloX-tiny algorithm

• Published in: Journal of intelligent manufacturing Vol. 36; no. 3; pp. 2141 - 2162
• Main Authors: She, Yuting, Li, Hongxin
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.03.2025; Springer Nature B.V
• Subjects: Algorithms; Business and Management; Control; Crystal defects; Crystal dislocations; Crystal growth; Crystallization; Crystals; Datasets; Frames per second; High temperature; High temperature environments; Industrial production; Machines; Manufacturing; Mechatronics; Parameters; Photovoltaics; Processes; Production; Raw materials; Robotics; Silicon; Single crystals; Deep learning; Lightweight model; Crystal line; YoloX-tiny
• ISSN: 0956-5515; 1572-8145
• DOI: 10.1007/s10845-023-02312-z

Monocrystalline silicon is an essential raw material for the photovoltaic industry, and industrial production requires keeping monocrystalline silicon crystals free of defects. Monocrystalline silicon production requires the staff to adjust the monocrystalline furnace parameters according to the state of the monocrystalline silicon to control the crystal growth process. The high-temperature environment inside the furnace causes the staff to observe the status of the crystal lines only through the industrial camera. The crystallization process is based on the generation of crystal lines to determine if the crystal is in a stable state. The isometric growth process determines whether dislocations have occurred in the crystal by crystal line characteristics. Therefore, it is necessary to automatically detect the status of the crystal lines through algorithms. We have built a monocrystal silicon crystal line dataset by analyzing the image features of the crystallization process and the isometry process of monocrystal silicon. Then we propose an improved YoloX-tinys model based on the YoloX-tiny model, which can detect crystal line features accurately and quickly at low arithmetic power. The backbone network is replaced with ShufferNetV2 lightweight network and the internal 3*3 convolutional kernel is replaced with a 5*5 size to improve the computational power of the model. DFC(Decoupled Fully Connected Attention) attention mechanism is added to the Neck network to enhance the feature processing capability. We also optimize the Neck network by replacing the depthwise separable convolution and applying the h-swish activation function. The improved model achieves 99.53% mAP on the proposed dataset. Meanwhile, the number of parameters of the model decreases from 5.03M to 2.35M, and the FPS(Frames Per Second) increase from 45.34 to 59.41. The results demonstrate that our model is able to achieve accurate crystal line detection with less computational consumption compared to other models, while achieving higher detection accuracy and speed.