Anomaly detection through Probabilistic Support Vector Machine Classification

Probabilistic Support Vector Machine Classification (PSVC) is a real time detection and prediction algorithm that is used to overcome assumptions regarding the distribution of the data. Its classification output is complemented with a probabilistic cost function and incorporates a degree of uncertainty in its predictions. Computationally, this algorithm is fast because it performs the classification using only a fraction of the original data. This project investigates the use of support vector machines (SVMs) to detect anomalies and isolate faults and failures in electronic systems. The output of the SV Classifier is calibrated to posterior probabilities thus improving the classical SVM deterministic predictor model to a more flexible probabilistic “soft” predictor model. This result is desirable because it is anticipated to reduce the false alarm rate in the presence of outliers and allow for more realistic interpretation of the system health. This report also investigates the use of a linear principal component decomposition (PCA) of the input data into two lower dimension subspaces in order to decouple competing failure modes in the system parameters and uncover hidden degradation features. The SV classification is then used in the two extracted orthonormal subspaces to determine a predictor model for each subspace respectively. A final decision function is constructed with the joint output of the two predictor models. The approach is tested on simulated and real data and the results are compared to the popular LibSVM software. This project will attempt to address some present issues in health monitoring of electronic systems: a) reduction of false negative alarms in the training stage b) reduction false positive alarms in the training stage c) identify hidden degradation of system parameters and d) investigate the presence of intermittent faults in healthy system performance.