Dominant trend based logistic regression for fault diagnosis in nonstationary processes

This paper presents a fault diagnosis method called dominant trend based logistic regression (DTLR) for monitoring nonstationary processes. Different from conventional sample-wise diagnosis approaches, it uses sliding windows to collect process data and extract dominant trend features. After data preprocessing via robust sparse representation, the feature vector reflecting variation trend is obtained by solving a convex optimization problem, i.e., dominant trend extraction (DTE). Then the ℓ2-norm of the dominant trend vector is used as a detection index to quantify the dissimilarity between normal and abnormal conditions. Once it exceeds the control limit, the feature vector is used to train the weight vector of logistic regression. The fault type can be determined as the class with the maximum conditional probability. With trend information, DTLR can effectively detect and isolate faults in nonstationary processes. Simulations on a synthetic nonstationary dynamic process, a nonstationary continuous stirred tank reactor (CSTR), and the real data of a blast furnace iron-making process illustrate superior monitoring and isolation performance of DTLR, compared with conventional methods.