This paper introduces a novel approach towards sensor fault diagnosis in wireless structural health monitoring systems. As compared to traditional fault diagnosis approaches, a number of innovations are reported in this paper. First, by embedding fault models and algorithms directly into wireless sensor nodes, sensor faults can be self-detected by the nodes in a distributedcooperative fashion. Second, no redundant sensor installations are required for fault diagnosis, because the wireless sensor nodes exploit the redundant information of correlated sensors already installed in the monitored structure ("analytical redundancy"). Third, instead of using raw time series of sensor data for analytical redundancy, the sensor data is first transformed from the time domain into the frequency domain on-board the sensor nodes, entailing significantly reduced data traffic. Fourth, nonlinearities in the data sets (e.g. due to measurement factors) are handled by implementing the analytical redundancy approach in terms of feedforward backpropagation neural networks embedded into the wireless sensor nodes. Fifth, due to the adaptation abilities of the embedded neural networks, sensor fault diagnosis remains efficient and accurate even if the monitored structure is subject to structural changes (or damage) as reflected in the sensor data. Sixth, no a priori knowledge about the structure or about the sensor instrumentation is required because the neural networks, representing a purely data-driven approach, take previously collected sensor data as the sole basis for fault diagnosis. This paper presents the fault diagnosis methodology, followed by the implementation into a fault-tolerant wireless structural health monitoring system prototype. Validation experiments on a laboratory test structure demonstrate the efficiency and accuracy of the proposed approach.
