Interpretable fault detection using projections of mutual information matrix

This paper presents a novel mutual information (MI) matrix based method for fault detection. Given m-dimensional process, the MI is m×m in which (i,j)-th entry measures values between ith dimension and jth variables. We introduce recently proposed matrix-based Rényi’s α-entropy functional to estimate each of matrix. The new estimator avoids density estimation it operates on eigenspectrum (normalized) symmetric positive definite (SPD) matrix, makes well suited industrial process. combine different orders statistics transformed components (TCs) extracted from constitute detection index, derive simple similarity index monitor changes characteristics underlying process consecutive windows. term overall methodology “projections matrix” (PMIM). Experiments both synthetic data benchmark Tennessee Eastman demonstrate interpretability PMIM identifying root variables that cause faults, its superiority detecting occurrence faults terms improved rate (FDR) lowest false alarm (FAR). advantages also less sensitive hyper-parameters. Code available at https://github.com/SJYuCNEL/Fault_detection_PMIM.