A reconfigurable and element-wise ICI-based change-detection test for streaming data

Abstract

Detecting changes in data-generating processes is a primary requirement for adaptive and flexible systems endowed with computational intelligence abilities. In order to maintain/improve their performance in evolving or dynamic environments, these systems have to detect any variation in the data-generating process and react and adapt to the new operating conditions. The problem of detecting changes in streams of data is generally addressed by means of Change-Detection Tests (CDTs) and, recently, a family of CDTs based on the Intersection-of-Confidence-Interval (ICI) rule has been presented. ICI-based CDTs monitor data streams by extracting Gaussian distributed features from non-overlapping data windows. The drawback of such a window-wise operational mode is a structural delay, which is particularly evident when the change magnitude is large. We present a novel ICI-based CDT that overcomes this problem by operating in an element-wise manner thanks to a Gaussian transform of the acquired data. Such an element-wise CDT is characterized by a high change-detection ability and a reduced computational complexity, which makes it suitable for the execution on low-power embedded systems. The proposed CDT is also provided with a reconfiguration mechanism that, after any detected change, allows the CDT to be reconfigured on the new working conditions to detect further changes. A wide experimental campaign shows the effectiveness of the proposed element-wise CDT both on synthetic and real datasets.