Outlier Sensitive Online Change Detection using Convex Combination of Adaptive Filters in Sensor Data Streams

Detecting changes in multivariate streaming sensor data by estimating the probability distribution of the log-likelihood over different time windows is difficult due to presence of outliers. To deal with such situations this paper proposes a modified-adaptive online change detection algorithm. This approach utilizes a convex combination a slow and a fast LMS-based adaptive filters making it robust against outliers. The robustness is ensured by finding the probability of occurrence of the outliers determined based on the change in the mean and variance of the moving averages within a threshold. The versatility of the proposed approach is tested out in two different scenarios; a synthetically generated dataset and a real-world dataset featuring sensor streams from a reefer container. The algorithm’s effectiveness is also evaluated against the state-of-the-art change detection methods with the help of appropriate metrics like False Positive Rate, False Negative Rate, and Latency which are crucial in assessing the model’s responsiveness in real-time. In particular high false positive rate in case of the proposed algorithm indicates, it is insensitive to the presence of outliers in both univariate and multi variate data streams.