Automatically Designing Network-Based Deep Transfer Learning Architectures Based on Genetic Algorithm for In-Situ Tool Condition Monitoring

Abstract

<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">In-situ</i> tool condition monitoring ( <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">in-situ</i> TCM) is vital for metal removal manufacturing which realizes on-machine diagnosis in a real-time manner. The limitation of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">in-situ</i> TCM based on traditional deep learning lies in several aspects: the requirement of sufficient labeled data of health conditions, the empirically manual designed architecture, and the labor-intensive tuning of hyperparameters. Network-based deep transfer learning (NDTL) partially solves the problem of limited labeled data. However, the time-consuming architecture design and the tuning of hyperparameters also pose a great impact on the schedule of real <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">in-situ</i> TCM projects. In this article, a new NDTL method is proposed, which is automatically designed by the genetic algorithm. A degradation monitoring experiment is conducted employing edge computing devices under multiple working conditions, where the texture of the machined surface is collected during the whole life cycle of milling cutters. The experimental results suggest that the proposed method possesses competitive performance evaluated by both the robustness metrics (e.g., the area under the curve of precision-recall) and the efficiency metrics (e.g., multiply–accumulates), where the trial-and-errors are reduced significantly by the automatic architecture design and the selection of hyperparameters.