关键词: 磨削变形预测, 孪生神经网络, 多变量耦合条件, 特征对比学习

Abstract: Traditional models such as CNN,LSTM,and GRU have limited feature extraction capabilities when dealing with multivariate coupling and signal nonlinear complexity in the grinding process.A deformation prediction method based on twin CNN-LSTM network was proposed,which combined the spatial feature extraction ability of CNN and the time series modeling advantage of LSTM to extract local and global features of grinding signals.By using a dual input structure with shared weights,the model significantly enhanced its ability to perceive feature differences under dynamic changes in grinding parameters by comparing the initial and current signal features.The experiment showed that compared with the traditional single model(CNN,LSTM),gating optimization model(GRU)and Spatiotemporal feature fusion model(CNN-LSTM),the prediction accuracy of the validation set and test set of Siamese CNN-LSTM model and Siamese CNN-LSTM model respectively reached 95.86% and 92.56%,and increased by 21.43% and 15.64% compared with the optimal control model(CNN-LSTM).The RMSE and MAE of the validation set were significantly reduced,exhibiting excellent prediction accuracy and robustness under complex working conditions,providing a new technical means for deformation prediction under complex working conditions and having broad application potential.

Key words: grinding deformation prediction, twin neural network, multivariable coupling conditions, feature contrastive learning