Adaptive signatures of soft-failures in end-user devices using aggregated TCP statistics

We present a new approach for effective soft-failure characterization in end-user devices (EUDs) on networks that support the TCP/IP. Our method can be employed for creating fully automated, accurate and scalable fault diagnostic systems. First, we describe Normalized Statistical Signatures (NSSs), a technique for characterizing EUD soft-failures. We create the NSSs by using aggregated statistical features extracted from TCP packet streams collected on-demand upon user complaint. We then introduce the Link Adaptive Signature Estimation (LASE) technique to minimize the number of NSSs needed to create diagnostic systems that have generalization capability for coping with communication link variations. To achieve this, we create Feature Estimator Functions (FEFs) using multivariate regression techniques and a minimal number of signatures of emulated EUD faults. We use these FEFs to generate synthetic NSSs which, can be used to train diagnostic systems with robust generalization capabilities. We expect that the combined use of NSSs and LASE technique will serve as the foundation of next-generation fault diagnosis systems.