A METHODOLOGY ON INTERPRETABLE NOVELTY DETECTION

摘要

Vibration-based Structural Health Monitoring (VSHM) systems continuously gather data from an array of sensors mounted on a structure. Features are constructed from the data measured. The aim is to monitor the vibration responses in the search for changes that may hint to damage. The continuous data acquisition generates high-dimensional feature spaces that require Data-Driven approaches to make inferences concerning the integrity of the structure. In recent years, machine learning has played an increasingly important role in VSHM. Data-driven algorithms have been successfully used to construct models capable of detecting anomalies such as damage in the features derived from the vibration signals. Mahalanobis Distance based novelty detection is a common used method to detect damage. Yet, the resulting models have been labelled “black box models” given that they lack interpretability. This becomes a relevant challenge in the presence of high-dimensional feature spaces. Using machine learning algorithms that can be interpreted would enable a more reliable novelty detection process, building trust in these methods and easing the decision-making process. Decision Trees (DT) is a widely used interpretable machine learning algorithm. The hierarchical structure of this algorithm enables the prioritisation of features that are used as predictors in the damage detection models. Furthermore, the nature of the algorithm enables the user to track the decisions and understand the classification process in detail. In this paper, we introduce the complementary use of so-called “black box models” and DT for novelty detection. The proposed damage detection approach is tested on an experimental setup with a 14.3m wind turbine blade (WTB) equipped with 24 accelerometers. A pseudo-damage was simulated by adding masses to several locations of the WTB. The pseudo-damage was detected by means of a semi-supervised novelty-detection. The novelties were later studied in detail with decision-trees to make inferences on their potential causes.