Study on Shoulder Edge Breakage during the CSSCGP Based on MQBNet

As a crucial step in the Czochralski silicon single crystal growth process (CSSCGP), the shoulder plays a significant role, and the shoulder edge is a vital feature for determining whether the process can smoothly proceed to the next stage. Existing methods for shoulder edge recognition, relying heavily on manual discernment, suffer from issues such as cost inefficiency and low effectiveness. To address this, this paper proposes a data-driven approach for shoulder edge recognition based on Multi-granularity Quadratic convolutional and Broadcast self-attention Network (MQBNet). First, a Quadratic convolutional network (QConv) with strong feature extraction capabilities is adopted to extract fine-grained features from the signals. QConv with different kernel sizes is used to capture features at multiple scales. Subsequently, a Long Short-Term Memory (LSTM) network is used to capture long and short-term dependencies. The broadcast self-attention mechanism is then incorporated to obtain crucial feature information. Finally, the reliability of the model is validated through industrial field data.