Non-Stationary Process Monitoring for Change-Point Detection With Known Accuracy: Application to Wheels Coating Inspection

This paper addresses the problem of monitoring online a non-stationary process to detect abrupt changes in the process mean value. Two main challenges are addressed: First, the monitored process is non-stationary; i.e., naturally changes over time and it is necessary to distinguish those “regular” process changes from abrupt changes resulting from potential failures. Second, this paper aims at being applied for industrial processes where the performance of the detection method must be accurately controlled. A novel sequential method, based on two fixed-length windows, is proposed to detect abrupt changes with guaranteed accuracy while dealing with non-stationary process. The first window is used for estimating the non-stationary process parameters, whereas the second window is used to execute the detection. A study on the performances of the proposed method provides analytical expressions of the test statistical properties. This allows to bound the false alarm probability for a given number of observations while maximizing the detection power as a function of a given detection delay. The proposed method is then applied for wheels coating monitoring using an imaging system. Numerical results on a large set of wheel images show the efficiency of the proposed approach and the sharpness of the theoretical study.