Fault detection in sensors using single and multi-channel weighted convolutional neural networks.

The success of sensor based application depends on the availability of clean data and its fail safe operation. However, due to their inherent dynamical behavior, detecting faults in sensors can be challenging. To this end, we propose a signal processing and nonlinear dynamics based fault detection approach for learning the normal/abnormal states of sensors using machine learning. The characteristic traits in detecting faults are the textured images that are generated from the time series of sensor measurements. Our approach uses these textured images that capture the dynamical properties of the sensor systems along with spectrograms with weighted Convolutional Neural Networks (wCNNs). We first test our approach on a radar sensor system since identifying faults in such a system is difficult due to their inherent complex dynamics. An evaluation of the multi-channel wCNN model for radar fault detection shows its robustness to tolerate noise. The efficiency of our approach is validated on a real data set from a gearbox sensor system. Results demonstrate that the unique patterns from the image representation convey valuable information about the dynamical states of the sensor systems in improving fault detection.