Skip to content
Conference Open access

Dynamic, Physics-Based Wear Prediction Using Multimodal Sensor Data for Industrial Applications

2026 · Proceedings of the 23rd International Conference on Informatics in Control, Automation and Robotics · 0 citations · 21 references

Abstract

: In autonomous manufacturing, the industrial deployment of deep learning models for tool wear prediction frequently fails due to concept drift. This paper presents a dynamic, physics-informed online learning framework for the tool wear documented in the PHM 2010 dataset, built around a comparative study of Temporal Convolutional Network (TCN) and Long Short-Term Memory (LSTM) architectures, optionally regularized by Elastic Weight Consolidation (EWC). Physical domain knowledge is embedded twice: sparse measurement points are interpolated via the Paris-Erdo˘gan crack-growth law combined with an exponential function, and the predictions are constrained to monotonicity. An asymmetric Huber loss penalizes safety-critical underes-timations more heavily than overestimations. Using leave-one-cutter-out cross-validation (c1, c4, c6), static offline baselines are compared with adaptive online approaches and related to recently reported PHM 2010 methods, evaluated by predictive accuracy (MAE, MSE, RMSE, MAPE, Pearson-r , bias) and efficiency (inference time, model size). As the validation rests on the three labeled cutters of a single machine under fixed cutting parameters, the results are reported as a proof of concept under these conditions.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.