Automatic detection of rare observations during production tests using statistical models

Abstract

Engines are verified through production tests before delivering them to customers. During those tests, numerous measures are taken on different parts of the engine, considering multiple physical parameters. Unexpected measures can be observed. For this very reason, it is important to assess if these unusual observations are statistically significant. However, anomaly detection is a difficult problem in unsupervised learning. The obvious reason is that, unlike supervised classification, there is no ground truth against which we could evaluate results. Therefore, we propose a methodology based on two independent statistical algorithms to double check the results. One approach is the Isolation Forest model which is specific to anomaly detection and able to handle a large number of variables. The goal of the algorithm is to find rare items, events or observations which raise suspicions by differing significantly from the majority of the data and, at the same time, it discriminates non-informative variables to improve estimation. One main issue of Isolation Forest is its lack of interpretability. Within this scope, we extend the Shapley values, interpretation indicators, to the unsupervised context to interpret the model outputs. The second approach is the Self-Organizing Map (SOM) model which has nice properties for data mining by providing both clustering and visual representation. The performance of the method and its interpretability depend on the chosen subset of variables. In this respect, we first implement a sparseweighted K-means to reduce the input space, allowing the SOM to give an interpretable discretized representation. We apply both methodologies on aircraft engines data. Both approaches show similar results which are easily interpretable and exploitable by the experts.