A spin on active learning analysis for health monitoring.

Decisions on the operation and maintenance of high-valued structures are critical, collecting and analysing data that can inform these decisions motivates the development of structural health monitoring (SHM) and tool monitoring (TM) systems. Descriptive health-state information labels for measured data points are crucial for successful monitoring systems. However, the labels can be expensive to collect and are often unavailable. Due to this problem, fully-supervised machine learning is limited in it’s application to monitoring systems which gives rise to unsupervised and semi-supervised approaches, one of these approaches is active learning. Active learning aims to highlight which unlabelled data points have the highest value of information and would be most informative to the model if they were labelled. For risk-based active learning in SHM, the model is used to advise some maintenance decision process where querying a data point means an expert inspecting the structure to determine its health-state. The current paper applies an active learning regression technology to a machining tool health case study. The algorithm queries data points that are most important in helping determine the health-state of the tool. Querying the most informative datapoints can drastically reduce the costs of a monitoring system. The regression analysis results are used to estimate tool wear and inform a risk-based maintenance decision process. This addresses the problem of when to replace a tool based on its health status, which has applications for many monitoring system problems.