Authors: Joseph Ndong
Abstract: This paper presents a comparative study of two wellknown approaches for anomaly detection based signal processing techniques, with a newer approach combining these two methods.
The analysis will show that, combining Principal Component Analysis (PCA) and Kalman filtering based statistical anomaly detection techniques, could be a good basis to find a suitable
model to achieve more higher improvement and performance for anomaly detection than when the PCA and Kalman filter are
used separately for the same detection issue. We first calibrate a predictive model using model based principal component analysis, by means of sub-space identification algorithms, which we re-used to learn the parameters of a dynamic linear model. Thereafter a Kalman filter is applied on this state state model to form a decision variable where we perform anomaly detection using thresholds. We had studied the false alarm rate vs. detection
rate trade-off by means of the ROC (Receiver Operating Characteristics) curve to show the high performance obtained via this newer methodology. We validate the approach over different realistic data.