关键词: KDP晶体, 表面形貌, 深度学习, 连续小波变换

Abstract:

The surface morphology of potassium dihydrogen phosphate（KH₂PO₄， KDP） machined by single point diamond turning was the research objects. The low， mid， and high-frequency wavelengths and amplitudes of the three-dimensional surface were extracted by CWT and power spectral density（PSD） as sample sets. Cutting parameters were treated as key variables， and bi-directional long short-term memory （BiLSTM）， gated recurrent units（GRU）， random forest（RF）， and convolutional neural network（CNN） were developed to predict the wavelengths and amplitudes of different frequency bands， ultimately enabling the prediction of the three-dimensional machined surface topography. The results indicate that the BiLSTM model demonstrate superior prediction performance for mid and high-frequency wavelengths， as well as low and high-frequency amplitudes， with average percentage errors of 2.14% and 3.03% for mid and high-frequency wavelengths， and 4.62% and 7.19% for low and high-frequency amplitudes， respectively. The GRU model excelles in predicting low-frequency wavelengths and mid-frequency amplitudes， with errors of 3.83% and 5.68%. The predicted three-dimensional surface topography closely matches experimental results from the validation sets. The correspondence between cutting parameters and the three-dimensional machined surface of KDP crystals was revealed by combining continuous wavelet transform， power spectral density， and deep learning methods and the correctness was verified.

Key words: potassium dihydrogen phosphate（KDP） crystal, surface topography, deep learning, continuous wavelet transform（CWT）