Machine-Learning Sensor Validation for Digital Twins

Abstract

The rapid growth of digital twins (DTs), built upon Internet of Things (IoT) and Industrial IoT systems, demands a large variety of networked sensors’ solutions. Indeed, networked sensors enable various sophisticated applications of DT by gathering/integrating sensor data, meanwhile, sensor failures can potentially undermine DT representativeness and cause serious consequences. In this thesis, we propose three generic sensor fault detection, isolation and accommodation (SFDIA) architectures capable of promptly detecting sensor failures, identifying faulty sensors, and replacing their faulty data with reliable estimations. More specifically, the first modular architecture is built upon a series of neural network (NN) estimators and a classifier, which allows the selection of the most suitable models among diverse NN models with respect to the application. Estimators correspond to virtual sensors of all unreliable sensors (to reconstruct normal behavior and replace the isolated faulty sensor within the system), whereas the classifier is used for detection and isolation tasks. This architecture is enhanced further to fully exploit the spatio-temporal correlation of sensor data and provide real-time detection, isolation and accommodation of multiple faulty sensors. A multi-dimensional classifier in the enhanced architecture is responsible for interpreting residual signals (from previous stages) to detect and identify faulty sensors, and provide feedback to a controller block. The controller is policing inputs-outputs of two banks of NNs which are providing estimations and predictions of all unreliable sensors within the system, thus supporting nearly-instantaneous SFDIA performance. In the third proposed architecture, for the first time, we address the problem of SFDIA in large-size networked systems. Current available machine-learning solutions are either based on shallow networks unable to capture complex features from input graph data or on deep networks with overshooting complexity in the case of large number of sensors. To overcome these challenges, we propose a new framework for sensor validation based on a deep recurrent graph convolutional architecture (DRGCA) which jointly learns a graph structure and models spatiotemporal inter-dependencies. the proposed two-block DRGCA (i) constructs the virtual sensors in the first block to refurbish anomalous (i.e. faulty) behavior of unreliable sensors and to accommodate the isolated faulty sensors and (ii) performs the detection and isolation tasks in the second block by means of a classifier. A detailed performance evaluation on different real-world datasets is conducted. Results prove the effectiveness of the proposed architectures in detection, isolation and accommodation of faults. Performance comparison shows their superiority over state-of-the-art machine-learning-based architectures.