Abstract
Sensors play a critical role in industrial applications, but are susceptible to faults caused by various internal and external factors. Traditional data-driven methods for sensor fault detection and isolation often rely on virtual sensor models trained with available measurements. In this paper, we propose a layered multi-scale architecture exploiting dilated convolutional neural network (CNN) as building blocks. The underlying hypothesis is that dilated convolutions are more effective in extracting complex spatial and temporal patterns from time series data generated by spatially-distributed sensors. The performance of the proposed model is evaluated against state-of-the-art techniques through a systematic
statistical analysis. Results on two datasets comprising pressure sensors in a simulated system of interconnected water tanks and temperature sensors in a research facility for carbon capture and storage demonstrate the superior performance of the proposed architecture in both fault detection and isolation tasks. Different variants of the architecture are also explored.