SENSOR FAULT DIAGNOSIS COUPLING DEEP LEARNING AND WAVELET TRANSFORMS

Sensor networks facilitate collecting measurement data necessary for decision making regarding structural maintenance and rehabilitation in structural health monitoring (SHM) systems. Nevertheless, the reliability of decision making in SHM systems depends on the proper operation of the sensors. Sensors may exhibit faults, entailing faulty data and incorrect judgment of structural conditions. Therefore, fault diagnosis (FD), comprising detection, isolation, identification, and accommodation of sensor faults, has been introduced in SHM systems, enabling timely detection of faulty data while advancing reliable operation of SHM systems. Traditional FD approaches based on “analytical redundancy” take advantage of correlated sensor data inherent to the SHM system, sometimes neglecting the fault identification step, and are implemented for specific sensor data. In this paper, an analytical redundancy FD approach for SHM systems, coupled with machine learning algorithms and wavelet transforms, capable of processing any type of sensor data is presented. A machine learning (ML) regression algorithm is proposed for fault detection, fault isolation, and fault accommodation, and an ML classification algorithm is proposed for fault identification. Continuous wavelet transform (CWT) is used as a preprocessing step of fault identification, exposing fault patterns in the data. The ML-CWT-FD approach is validated using data from a real-world SHM system in operation at a railway bridge implementing a deep neural network as ML regression algorithm and a convolutional neural network as ML classification algorithm. As a result of this paper, the ML-CWTFD approach is demonstrated to be capable of ensuring reliable SHM systems.