Intelligent Fault Detection and Isolation Based on NARX Neural Networks

Intelligent systems have become an essential component of modern technological landscapes. Their reliability is a major concern, as faults can lead to catastrophic consequences on their behavior and overall performance, making Fault Detection and Isolation (FDI) a critical task for such systems. The complex nonlinear dynamics involved in their operation make this task more challenging. In this context, this paper proposes an intelligent fault detection and isolation methodology, using an incubator system as a representative case study. The proposed method employs a Nonlinear Autoregressive Exogenous (NARX) neural network in a parallel structure to model the complex nonlinear system dynamics. Different implementations of the NARX model, based on Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Elman networks, were compared to assess their modeling performances and determine the best-performing one. The differences between these predictions and the actual outputs are referred to as residuals. For fault classification, a comparative study is conducted between five machine learning (ML) methods: Multilayer Perceptron, Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Linear Discriminant Analysis (LDA). These methods analyze the residuals to isolate the specific fault from predefined potential faults, including both actuator and sensor faults. The results demonstrate the effectiveness of the intelligent FDI methodology presented. The NARX models proved effective with high modeling performance, particularly the MLP-NARX, which outperformed the other models. The comparative study of classifiers highlights the performance differences among the five methods, with the MLP classifier achieving the best results across all metrics, thereby reinforcing its suitability for practical FDI applications, particularly given its strong ability to capture complex nonlinear relationships in the data.