Asynchronous standard deviation method for fault detection

This paper introduces an asynchronous-standard-deviation (ASD) method for particle-filter based fault detection. In order to detect the anomalous condition of present states compared with the past healthy states of a system, we design the ASD by the samples/data (with sample mean M₂) in time period T₂, but using sample mean M₁ in different time period T₁ instead of M₂. The ASD reflects the deviation of data in T₂ from M₁ in T₁ (e.g., T₁ is the healthy period of a system). In our method, a state-space model is adopted to describe the system, and a classical particle filter algorithm is used to track the system condition. The healthy degree for the health evaluation of a system is proposed based on the comparison between the ASD in present time period and the standard deviation from the past healthy system. To describe the possibility of a potential fault, the fault degree of a system is proposed based on the ASD of the prediction data for a past period. These prediction data are obtained by the inverse prediction (from present to past along reverse direction of the time axis) of the particle filter algorithm. A fault is detected if the fault degree exceeds the healthy degree continuously. Our method is validated on a simulated fault case of a three-vessel water tank system and a real fault case of a UH-60 planetary gear plate. The experimental results show that our method has better performance on fault detection, and detects the faults earlier than the compared methods.