Novel CUSUM Methods for Repetitive Change Detection in Sensor Signals

Repetitive changes are very common in real-world systems such as in refrigeration temperature, respiration rate detected by wearable devices or human activity detected in surveillance camera. These devices generate certain meaningful repetitive changes which needs to be detected for example, the refrigeration off cycle. However, these data streams exhibit the presence of added sensor noise and external interferences which makes it difficult to detect such changes accurately in real-time. Therefore, we proposed two new methods for this problem where Method 1 is a Bidirectional cumulative sum (CUSUM) consisting of two one-sided algorithms, and Method 2 is based on adaptively updating the drift component in a CUSUM representation of a time varying process. The efficacy of the proposed methods is evaluated in terms of the actual change's detection, number of false alarms and missed changes. It is found that our proposed methods outperform other existing methods when tested on an experimental dataset on refrigerated shipping containers.