Tool condition monitoring is essential to maintain the final product quality and machining efficiency of the manufacturing process. However, traditional physics-based and data-driven approaches have limitations either on prediction efficiency or performance generalization, due to the nature of the respective approaches. To address these issues, in this article, a physics-informed deep learning approach is developed, which integrates the tool wear mechanism into the data-driven model. First, some representative physical information is selected for the task learning. Then, four practical physics-informed methods are proposed to integrate various physical information into the data-driven models. Based on these physical constraints, a physics-informed deep learning model is specially designed for tool wear monitoring. Compared with previous studies, more diverse physical information can be effectively utilized to guide the hypothesis space, thereby improving the generality of the model. The effectiveness and feasibility of this model under various working conditions are verified in high-speed milling experiments. The results show that the wear prediction of the proposed approach is more accurate and consistent under unknown machining conditions.

The metal-based additive manufacturing (MAM) processes have great potential in wide industrial applications, for their capabilities in building dense metal parts with complex geometry and internal characteristics. However, various defects in the MAM process greatly affect the precision, mechanical properties and repeatability of final parts. These defects limit its application as a reliable manufacturing process, especially in the aerospace and medical industries where high quality and reliability are essential. MAM process monitoring provides a technical basis for avoiding and eliminating defects to improve the build quality. Based on of the nature of the MAM build defects, this article conducts a thorough investigation of monitoring methods, and proposes a machine learning (ML) framework for process condition monitoring. According to the structure of ML models, they are divided into shallow ML-based and deep learning-based methods. The state-of-the-art ML monitoring approaches, as well as the advantages and disadvantages of their algorithmic implementations, are discussed. Finally, the prospects of ML based process monitoring researches are summarized and advised.

Accurate and fast prediction of tool conditions is fundamental to improve the machining accuracy and consistency in smart machining systems. The current tool condition monitoring methods, i.e., physics-based and data-driven approaches, have either low prediction accuracy or model generalization. To solve the shortcomings and utilize the benefits of both sides, a novel physics-guided deep learning model is developed in this study. The physics of cutting mechanics and tool wear model is applied to guide the model construction and regulate the network learning process. The model first generates large labeled simulation dataset for the pretraining of the deep network, and solves the problem of labeled-data insufficiency for network training in practice. It then fine-tunes the model through the monitored signal to optimize the network weight and alleviate the deviation between the physical model and the actual machining process, which improves the physical consistency and generalization of the model. Additionally, with the introduction of attention mechanism to the deep residual network, discriminant features can be extracted to distinguish wear values and working conditions. The experimental results show that by learning the physics, the physics-guided deep network can accurately predict the tool wear even under limited training sets and varied working conditions, and it outperforms the original data-driven models and the physical models.

Tool condition monitoring is important to guarantee the product quality and improve the productivity in high-performance computer numerical control (CNC) machining. Due to the harsh working conditions and non-stationary milling process, it is difficult to establish a monitoring model suitable for complex conditions. In this article, a pyramid long short-term memory (LSTM) auto-encoder is proposed to monitor tool wear. Unlike the classic stacked LSTM, the pyramid model is constructed based on the frequency spectrum of cutting signals. Each layer focuses on one periodic scale, and each unit focuses on one periodic fluctuation. The features are compressed layer by layer according to the frequency spectrum. The learned patterns are restricted by the spectrum-based structure, which simplifies the monitoring task and reduces the model complexity. The length of the cutting signal is also no longer limited by the memory capacity of LSTM. In the meantime, the efficiency of long-term signal processing is also greatly improved by reducing the number of units. In addition, the introduction of auto-encoder can further improve the accuracy of the model under complex working conditions through unsupervised learning. The good performance on high-speed milling experiments shows the accuracy of the model under unknown tools and milling parameters. Compared with the classic stacked LSTM, the pyramid LSTM auto-encoder has advantages in computational speed, stability, and prediction accuracy.