Deep transfer residual variational autoencoder with multi-sensors fusion for tool condition monitoring in impeller machining

The data-driven deep learning models show incomparable superiority in tool wear monitoring. However, facing the monitoring environment of real-time changes in tool model, workpiece size, and cutting parameters in the machining of centrifugal compressor impeller, the flexibility of the model between different tools is rarely considered. In this manuscript, a novel deep transfer learning model based on residual variation autoencoder (Tresvae) with multi-sensors fusion signals is proposed. The sound pressure signals, acceleration signals, and spindle motor current signals are collected and confused into images as monitoring samples. Moreover, the residual network is used to optimize the encoding part of the variational autoencoder network while the decoding part remain unchanged. Then the parameter transfer strategy is adopted to realize the model transfer diagnosis between different tools. Experiments show that the proposed model achieves superior identification results than other methods on the same types and different types tools monitoring in impeller machining. Among them, the maximum recognition accuracy of the same type of tools is 97.25%, and that of different types of tools is 92%.