International Conference on Advanced and Competitive Manufacturing Technologies milling tool wear prediction using unsupervised machine learning

Abstract

Degraded or defect machine components and consumables negatively impact manufacturing quality and productivity. Diagnosing and predicting the wear or degradation status of critical machine components or parts are therefore of general interest. To tackle this challenge, data-driven approaches based on supervised machine learning principles have demonstrated promising results. However, supervised learning models capable of degradation identification require large quantities of data. In practice, run-to-failure data in large amounts is usually not available and expensive to obtain. To overcome this issue, this study proposes an unsupervised learning approach for degradation prognostics of machine tool components and consumables. It uses time series of multi-sensor signal data, which are transformed into a feature representation. The features consist of various characterizations of the time series, allowing to make different signal measurements comparable, and cluster them according to their feature values. The herewith obtained density-based clustering model is used to diagnose and predict the degradation states of components and parts in unknown conditions. The novelty in the proposed approach lies within the identification of continuous component and part degradation states based on unsupervised learning principles. The proposal is verified and demonstrated on an exemplary data set containing a small sample of run-to-failure multi-sensor signals of milling inserts and their corresponding wear state. By the application of the proposed procedure on the exemplary data set, we demonstrate that an unsupervised clustering approach is capable of separating wear data such that meaningful and accurate estimations of the part condition are possible. The advantages are its ability to cope with scarce data sets, its limited engineering and hyperparameter tuning effort, and its straightforward implementation to a multitude of degradation and wear diagnostics scenarios.

Investigation of Tribological Behaviour on DLC Coatings for AA5051 using DC Sputtering

“…There are many methods to improve the hardness and strength of the material like stress hardening and shot-peening. Still, these methods are inefficient in effectively improving the wear resistance and improving the lubrication simultaneously [3]. The coatings are proposed on the substrates.…”

Milling cutter fault diagnosis using unsupervised learning on small data: A robust and autonomous framework

“…The applicability, fault rate and feasibility of clustering method were discussed within the framework of state of the art studies. Gittler et al [7] predicted the tool wear condition during milling via unsupervised machine learning approach. The tool condition was predicted with meaningful and accurate values according to the results.…”

A Robust Tool Condition Monitoring System Based on Cluster Density under Variable Machining Processes

“…However, there are few reports on TCM research based on unsupervised methods due to their poor nonlinear fitting ability and real-world supervising ability. Thomas Gittler et al [25] proposed an unsupervised approach for degradation prognostics of machine tool components and consumables. However, this was not a completely unsupervised method although a clustering algorithm was used as the core algorithm, since the model should be trained on the selected samples and their respective feature distributions.…”