Distributed deep learning enabled prediction on cutting tool wear and remaining useful life

Abstract

To optimise the utilisation cost of cutting tools, it is imperative to develop an online system to efficiently and accurately predict tool wear conditions and remaining useful lives (RULs). With this aim, a novel system is proposed based on deep learning algorithms distributed over an edge-cloud computing architecture. The system is innovative in the following aspects: (i) a lightweight convolutional neural network-random forest (CNN-RF) model is designed to be executed on an edge device to assess tool wear conditions efficiently, which supports severe tool resilience and tool replacement when necessary; (ii) a convolutional neural network-long short-term memory (CNN-LSTM) model is designed to be executed on a cloud to process long-term signals to predict the RUL of the cutting tool, which supports fine-tuning tool parameters dynamically; (iii) a signal compression mechanism is developed to condense the signals of tooling conditions into 2D images so the signal volumes transferred over the network are minimised and signal security is improved. Experiments were performed in a real-world machining workshop for research methodology validation. It showed that the prediction accuracies for tool wear and RUL achieved 90.6% and 93.2%, respectively, and the volume of signals transferred over the network was reduced by 89.0%. The experiments and benchmarks with comparative algorithms demonstrated that the system and its methodology exhibited great potential to reinforce cutting tool optimisation for real-world applications.

Tool remaining useful life prediction considering wear state based on hybrid attention network

“…For models that rely on a single algorithm or similar types of algorithms, it is difficult to obtain excellent performance in both feature extraction and regression prediction. Hybrid models [21][22][23] have broader application prospects in the tool RUL prediction due to its variable structural characteristics. Combining spectral subtraction and CNN can roughly estimate the tool wear state to achieve RUL prediction.…”