An Augmented AutoEncoder With Multi-Head Attention for Tool Wear Prediction in Smart Manufacturing

Abstract

Computer numerical control (CNC) machine tools play a crucial role in the manufacturing industry, and cutting tools, as key functional components, directly impact the quality of the machining process. An improved autoEncoder with multi-head attention for tool wear prediction is proposed. MultiCNN-Attention-GRU (MCAG) consists of an encoder and decoder. The encoder contains multiple sets of Convolutional Neural Networks (CNNs) and CNNs adaptively extract signal features. The decoder includes Multi-Head Attention (MHA) and Gated Recurrent Unit (GRU), which can adaptively enhance the relevant feature weights and extract long-term, deep different features. For the model training, a monotonicity loss function is defined. The proposed method is validated on the 2010 PHM Data Challenge (PHM2010) public dataset. The original dataset is dimensionally reduced and then resampled. The experimental results demonstrate the effectiveness of the proposed algorithm, achieving an Mean Absolute Error (MAE) of 6.15 and Mean Square Error (MSE) of 79.6, which is approximately 1.6 and 13 lower than the second-place algorithm. The result validates the superior performance of the proposed model compared to other deep learning algorithms in predicting tool wear.

Keywords

• Deep learning
• multi-head attention
• milling machine tool
• remaining life prediction