Structured dataset for optical machining in machine learning, including surface shape data and process parameters

In the process planning of ultra precision optical components, multi tool combination processing decisions have long relied heavily on expert experience, resulting in high costs, low efficiency, and insufficient stability. To overcome this challenge, this study proposes an innovative method based on supervised learning neural network architecture. We have developed a dedicated process decision data collection system: using the Python QT designer platform to build a user-friendly interactive interface, and innovatively designed a dedicated process coding mechanism to ensure the uniqueness of indexes for different parameter combinations, generating structured data formats adapted to machine learning algorithms. This article uses Python provided library functions to write a program that reads the feature data of the machined component surface shape from the dat binary file measured by Zygo interferometer. It collects the machining process parameter information corresponding to the surface shape in the interactive interface software and integrates the obtained information. In order to efficiently read and save the annotated process planning dataset, it is helpful to serialize the data and store it in a set of linearly readable files. The dataset in this study is saved in TFRecord format, which is convenient for subsequent machine learning and training. At the same time, for data containing multiple processing parameter information, we ensure the uniqueness of the index for different parameter combinations through weight design, and use single heat encoding to classify and characterize each process parameter. Single hot encoding is a standard and effective method for converting unordered classification features into numerical format. By creating binary columns, text labels that cannot be directly processed by machine learning algorithms are converted into equal binary data features. After preprocessing the surface data, the RMS values of each frequency band of the surface are finally extracted through Fourier downsampling to form a structured dataset.