Sensor fusion and failure detection using virtual sensors

Fault detection has become an integral aspect of modern system design. To minimize costs, a software approach is desirable. We have designed a neural network based decentralized virtual sensor system to conduct sensor fusion and sensor failure detection. To accomplish effective fault detection, we have conceived an innovative row-wise and column-wise checking procedure that is able to detect all sensor faults while minimizing false alarms. Another measure that reduces the damage of false alarms is the addition of a sensor reinstatement procedure. This allows us to further tighten the decision threshold to prevent missed detection without compromising the goal of false alarm deterrence. We have demonstrated this virtual sensor system using a vibrating beam structure. The system has been shown to be very robust, giving a 100% detection rate in all experiments. This virtual sensor concept can be applied to large structures, such as a space platform, where robustness and autonomous reconfiguration is essential