Fault Diagnosis and Prognostics of Stochastic Distribution Systems

A novel fusion method is proposed for fault diagnosis and prognosis in nonlinear stochastic distribution control systems. Technically, a mapping relationship between fault increments and system output increments is established. The fault diagnosis algorithm that relies only on system output increments is obtained using the least squares theory, effectively addressing the diagnosis problem of fast-varying faults. In the case of unknown fault forms, a long short-term memory network is utilized to capture the long-term dependency relationship of the system output increment. This establishes a self-mapping relationship of the system output increment, thereby enabling multistep ahead prediction of the system output. The fault diagnosis algorithm is integrated with long short-term memory networks in a particle filtering framework, achieving multistep ahead prediction of fault magnitude and remaining useful life. The effectiveness of the proposed algorithm is validated in the wet end control system of a paper machine.