Predictive modelling of surface roughness in precision grinding based on hybrid algorithm

Abstract

Aimed at predicting surface roughness (SR) of bearing outer rings under various grinding conditions, a model utilizing the DBO-1DCNN-LSTM algorithm was proposed. The structural parameters of the 1D convolutional neural network with long short-term memory neural networks (1DCNN-LSTM) combination model are optimized using Dung Beetle Optimization algorithm (DBO), and comparative experiments demonstrate the excellent performance in extracting features from multiple sources of signals during the grinding process. By utilizing the DBO-1DCNN-LSTM model to extract vibration and acoustic emission signal features in precision grinding of bearing outer rings, a predicting method for SR considering multi-source heterogeneous data is proposed. Signal characteristics with grinding parameters are combined to build a SR forecasting model of precision-ground outer rings under different operating conditions. Experimental results indicate that incorporating batch normalization layers and employing the grinding parameters as model input can effectively enhance the forecast accuracy. It achieves a coefficient of determination (R²) of 0.9910, average absolute error (MAE) of 0.0050, root mean square error (RMSE) of 0.0067, and mean absolute percentage error (MAPE) of 0.0491. Capable of accurately forecasting the SR of bearing outer rings across different grinding conditions by the proposed approach. The model can shorten the production cycle from machining to inspection, ensuring the qualification rate of workpieces directly during the machining process. This facilitates efficient quality control and timely comprehensive decision-making in bearing production, ultimately improving production efficiency.