Network fault detection: classifier training method for anomaly fault detection in a production network using test network information

Abstract

We have prototyped a hierarchical, multi-tier, multi-window, soft fault detection system, namely the Generalized Anomaly and Fault Threshold (GAFT) system, which uses statistical models and neural network based classifiers to detect anomalous network conditions. In installing and operating GAFT, while both normal and fault data may be available in a test network, only normal data may be routinely available in a production network, thus GAFT may be ill-trained for the unfamiliar network environment. We present in detail two approaches for adequately training the neural network classifier in the target network environment, namely the re-use and the grafted classifer methods. The re-use classifier method is better suited when the target network environment is fairly similar to the test network environment, while the grafted method can also be applied when the target network may be significantly different from the test network.