Faulty sensor data often occur in structural health monitoring (SHM) systems using wireless smart sensor networks (WSSN). Serious sensor faults in the raw data may negatively affect SHM analysis and subsequent informed decisions. This issue is more critical for WSSN running decentralized data acquisition, because raw data are not transmitted back, and hence, data quality assessment can be extremely difficult for users. Therefore, developing efficient techniques to autonomously detect, identify, and recover sensor faults is essential. In the vibration data collected from the Jindo Bridge, some data sets are corrupted with sensor faults, which can be categorized as one of three types: drift, spikes, and bias. Accordingly, this paper presents a three-stage strategy to address this problem in WSSN. First, a distributed similarity test is employed to detect sensor faults; this test is based on the similarity of the power spectral density of the data among sensors within a cluster of nodes. Second, an artificial neural network model is trained to identify the types of sensor faults. Third, sensor data are recovered from the identified faults by applying a correction function or replacing faulty data with estimated values. Subsequently, numerical analysis is performed using a set of field measurements collected from the Jindo Bridge. Decentralized system identification is conducted using the original data and data recovered using the proposed strategy. The three-stage strategy is shown to successfully detect, identify, and recover sensor faults, which improves the results of system identification. Finally, the benefits and limitations of this strategy are discussed.