A Hybrid Classifier Using Reduced Signatures for Automated Soft-Failure Diagnosis in Network End-User Devices

Abstract

We present an automated system for the diagnosis of both known and unknown soft-failures in end-user devices (UDs). Known faults that cause network performance degradation are used to train the classifier-based system in a supervised manner while unknown faults are automatically detected and clustered to identify the existence of new categories of soft-failures. The supervised classifier used in the system can be retrained by including the newly detected faults to enhance its performance. The system uses 460 features to construct Normalized Statistical Signatures (NSSs) for fault characterization. Due to the high dimensionality of NSSs, EigenNSS was proposed to reduce the complexity without losing important information. Because of the natural network inconsistencies that exist in communication links, we propose FisherNSS, a reduced signature that provides improved linear separability between classes to further enhance classification performance. The system is evaluated over a live campus network using 17 emulated UD faults. The results show that the best overall classification accuracy of up to 97% was achieved by using FisherNSS with a dimensionality reduction of 96.74%. In comparison, both EigenNSS and FisherNSS have faster training and diagnosis time compared to NSS, which makes them suitable for on-demand as well as real time diagnostic applications. Furthermore, FisherNSS compared to EigenNSS has a higher diagnostic accuracy and quicker diagnosis time (order of microseconds).