Consistency-based diagnosis using data-driven residuals and limited training data

Abstract

Effective fault diagnosis is crucial for improving the durability and reliability of automotive systems. Developing a diagnostic system with desirable fault isolability requires an accurate model and/or representative training data covering all possible faults. One promising approach involves using physics-based neural network residuals, known as grey-box models. These networks, designed to represent the system’s nominal behavior and trained on fault-free data, are particularly advantageous when training data from faults are scarce. By incorporating causal relationships derived from physical insight, grey-box models retain the structural fault sensitivity of model-based residuals, enabling consistency-based diagnosis decision logic. However, despite their high accuracy in supervised learning benchmarks, neural networks often struggle with misclassification due to out of distribution data, a significant concern in diagnostic applications where false alarms are costly. This study highlights the importance of uncertainty quantification in neural network-based regression models and examines the interplay between different types of uncertainty in diagnostics. To address both epistemic and aleatoric uncertainties and achieve desirable fault isolation, the study applies adaptive thresholds and a measure for testing the validity of the residuals. Additionally, it proposes a consistency-based diagnosis framework using data-driven residuals, with its effectiveness demonstrated on an aftertreatment system of a heavy-duty truck under various drive cycles and fault scenarios.