Research on dataset optimization technology of process decision-making for supervised learning network in ultra-precision optical machining

In the process planning of ultra-precision optical components, the decision-making for multi-tool combined processing has long been highly dependent on expert experience, leading to high costs, inefficiency, and insufficient stability. To overcome this challenge, this study proposes an innovative method based on a supervised learning neural network architecture. We have developed a dedicated data acquisition system for process decision-making: a user-friendly interactive interface was constructed using the Python QT-designer platform, and a specialized process encoding mechanism was innovatively designed to ensure the uniqueness of indexes for different parameter combinations, generating a structured data format suitable for machine learning algorithms. The system also supports functions for outlier data exclusion and the addition of virtually annotated data by experts. To further overcome the limitations of small-sample learning, we combined data balancing and augmentation strategies to propose an improved dataset enrichment scheme. Specifically, a random cropping and multiplication technique for optical component surface shapes tailored to different process parameter categories was adopted, combined with the weighted cross-entropy loss function method in neural network classification decision-making, effectively mitigating the issue of class imbalance in small-sample data. Experimental results demonstrate that, compared to traditional methods, the new scheme achieves significant improvements in prediction accuracy for tool size, tool material, and abrasive material.