Industrial application of artificial intelligence for solving complex PHM problems

Abstract

This thesis proposes effective fault detection and explainable artificial intelligence methods for solving highly complex PHM problems, each method in diagnosis and prognosis perspective. First, a practical fault detection and prediction method is proposed by addressing a margin-maximized hyperspace. The proposed method is effective for a highly imbalanced dataset without any supervision, which is a frequently occurring and challenging problem in real-world applications. Specifically, knowledge-based feature manipulation is executed to provide sufficient information for a neural network. Next, variational autoencoder with regulated latent space transforms distinct input features into a latent space, which ensures high accuracy and robustness. Lastly, the obtained latent space is confirmed to statistically allocate two extremes of major (normal) and minor (faulty) clusters at an origin and unity, maximizing the sensitivity to classify faults. The effectiveness of the proposed method is demonstrated through field measurements of elevator door-strokes and showed high sensitivity to separate each cluster along with locational constancy compared to other autoencoders. Therefore, the proposed method is effective for real-world applications with scarce fault measurements. Second, this thesis proposes a novel explainable remaining useful life (RUL) prediction method. The main goal of the proposed method is precisely predicting RUL while tracking the origin of the prediction simultaneously to solve black-box characteristics of deep neural network (DNN). Specifically, effective DNN model for RUL prediction is designed and trained with manipulated features providing effective information regarding RUL of the object of interest. Next, the relevance distribution of each feature on RUL prediction is quantitatively calculated by addressing layer-wise relevance propagation and postprocessing methods specially designed for time-series degradation data interpretation. Lastly, the acquired relevance is used in two ways: feature selection and failure mode decomposition. The effectiveness of the proposed method is validated using a degradation dataset of aircraft turbofan engines, which is widely used for assessing the performance to predict RUL in the prognostics and health management area. The proposed method is greatly practical in diagnosing complex mechanical systems in the real world because the method could be a useful solution for selecting sensors only related to health states, which requires both domain knowledge and a massive amount of effort.