Distributed sensor networks working in harsh environmental conditions can suffer from permanent or transient faults affecting the embedded electronics or the sensors. Fault Diagnosis Systems (FDSs) have been widely studied in the literature to detect, isolate, identify, and possibly accommodate faults. Recently introduced cognitive FDSs, which represent a novel generation of FDSs, are characterized by the capability to exploit temporal and spatial dependency in acquired datastreams to improve the fault diagnosis and by the ability to operate without requiring a priori information about the data-generating process or the possible faults. This paper suggests a novel approach for fault detection in cognitive FDSs based on an ensemble of Hidden Markov Models. A wide experimental campaign on both synthetic and real-world data coming from a rock-collapse forecasting system shows the advantages of the proposed solution.