Multi-Scale Convolutional Gated Recurrent Unit Networks for Tool Wear Prediction in Smart Manufacturing

Abstract

As an integrated application of modern information technologies and artificial intelligence, Prognostic and Health Management (PHM) is important for machine health monitoring. Prediction of tool wear is one of the symbolic applications of PHM technology in modern manufacturing systems and industry. In this paper, a multi-scale Convolutional Gated Recurrent Unit network (MCGRU) is proposed to address raw sensory data for tool wear prediction. At the bottom of MCGRU, six parallel and independent branches with different kernel sizes are designed to form a multi-scale convolutional neural network, which augments the adaptability to features of different time scales. These features of different scales extracted from raw data are then fed into a Deep Gated Recurrent Unit network to capture long-term dependencies and learn significant representations. At the top of the MCGRU, a fully connected layer and a regression layer are built for cutting tool wear prediction. Two case studies are performed to verify the capability and effectiveness of the proposed MCGRU network and results show that MCGRU outperforms several state-of-the-art baseline models.

Improving Tool Wear Prediction with Synthetic Features from Conditional Generative Adversarial Networks

“…Yan et al replaced LSTMs with more recent recurrent architecture BiGRU units for RUL estimation of CNC (Computer Numerical Control) milling tool [11]. In [24], Xu et al proposed an MCGRU model where six separate CNN branches are used for feature extraction from multi-sensor readings in parallel but with different sizes of convolution kernels in each branch. Incorporating a multi-scale CNN design, their MCGRU model outperformed peers over the same task.…”

“…When different types of sensors simultaneously monitor a cutting operation, an intrinsic correlation should exist among these readings, such as vibration and acoustic, within a temporal vicinity W . However, if pattern learning branches are applied independently to each sensor data, for example, one CNN branch with two convolution+ReLU layers and one maxpooling layer for each sensor as in [24], the opportunity to learn the supposed intrinsic correlation among different sensor signals could be weakened or missed in the second tier of a hierarchical network.…”

Improving Tool Wear Prediction with Synthetic Features from Conditional Generative Adversarial Networks

“…Their description is given in Table IVa. To compare the proposed methods with [11], [12], [15], [24], we used three cutter records C 1 , C 4 and C 6 , as described in Table IVb. And the average value of flank wear from three flutes was used for model training and performance evaluation.…”

“…To generate synthetic features, the truth used for training cGAN was used again to produce synthetic In summary, from this second experiment, we find our synthetic feature generation proposal effectively reduces prediction error, despite a simple CNN network architecture exhibiting a disadvantage to GRU-based methods when the complexity of input features increases. Another potentially contributing factor here is that our features are extracted from the temporal domain only, unlike in [24], features extracted from both temporal and frequency domains were used in training MCGRU prediction models, in both RMSE and MAE terms.…”

Text sentiment classification based on Vit-BiGRU-attention model

“…For long data sequences, traditional unidirectional RNN networks usually suffer from gradient explosion and gradient disappearance, so researchers proposed BiGRU [8] . BiGRU is an improved bidirectional RNN network that not only solves the above problems, but also captures long-term dependencies [9] . The Vit-BiGRU-Attention model proposed in this paper uses a word2vec word vector representation to obtain local features of text in the convolutional layer of CNN.…”