Edinburgh Napier University

  The objective of this work is to design a sensor-fault detection and diagnosis system for the Internet of Things and Cyber-Physical Systems. The challenge is, however, achieving this objective within the limited computation, memory, and energy resources of the sensors. More importantly, the detection of faults is time-sensitive, whereas the diagnosis does not necessarily need to be fast. We propose a distributed sensor-fault detection and diagnosis system based on machine learning algorithms where the fault detection block is implemented in the sensor in order to achieve output immediately after data collection. This block consists of an auto-encoder to transform the input signal into a lower-dimensional feature vector, which is then provided to a Support Vector Machine (SVM) for classification as normal or faulty. Once detected, fault diagnosis is performed at a central node, such as a network server, to reduce the computational load on the sensor. In this work, a Fuzzy Deep Neural Network (FDNN) is used for diagnosis to provide further information, such as the type of fault. Here, the input propagates through a deep neural network and a fuzzy representation process. The output of these two components is then fused through densely connected layers. This multi-modal technique learns high-level representations in the data that are missed by conventional methods. To assess the performance of the proposed model, we utilize data obtained from a healthy temperature-to-voltage converter that are then injected with five types of fault: drift, bias, precision degradation, spike, and stuck faults. The performance from fault detection is analyzed in terms of detection accuracy, area under the ROC curve (AUC-ROC), false positive rate, and F1 score. Furthermore, the efficiency of fault diagnosis is shown by the classification accuracy parameter. The experimental results show the efficiency of the proposed fuzzy learning-based model over classic neuro-fuzzy and non-fuzzy learning approaches.