Deep learning based multi-source heterogeneous information fusion framework for online monitoring of surface quality in milling process

The multi-sensor configuration enables a comprehensive description of the machining processes and thus improves the capability of quality prediction model. However, the structural heterogeneity of various sensor data imposes barriers to information fusion as well as model construction. This study developed a novel multi-source heterogeneous information fusion framework based on deep learning for the prediction of milling quality, where thermal imaging is first attempted particularly. Specifically, the preprocessing module extracts multi-domain features from structured time series data, and the convolutional neural network based module is assigned to extract information from unstructured data. After that, the multilayer perceptron technique is employed to realize feature enhancement and fusion of cross-domain characteristics. Experimental validation was performed on a vertical machining center and comprehensive comparison experiments were conducted. The proposed approach achieves the best performance (minimum mean absolute percentage error 0.33%) and exhibits great robustness (standard deviation 0.17%). In addition, various time–frequency processing methods and convolutional neural network architectures are exploited for better configuration and prediction performance. The results revealed the great potential of thermal imaging for roughness prediction and the excellent prediction performance of the proposed framework demonstrates its superiority and effectiveness in practice.