LEARNING PROCESS BEHAVIOR FOR FAULT DETECTION

Recently, there has been an increased interest in self-healing systems. These types of systems are able to cope with failures in the environment they execute and work continuously by taking proactive actions to correct these problems. The detection of faults plays a prominent role in self-healing systems, as faults are the original causes of failures. Fault detection techniques proposed in the literature have been based on three mainstream approaches: process heartbeats, statistical analysis and machine learning. However, these approaches present limitations. Heartbeat-based techniques only detect failures, not faults. Statistical approaches generally assume linear models. Most machine learning techniques assume the data is independent and identically distributed. In order to overcome all these limitations we propose a new approach to address fault detection, which also gives insight into how process behavior changes over time in the presence of faults. Experiments show that the proposed approach achieves a twofold increase in F-measure when compared to Support Vector Machines (SVM) and Auto-Regressive Integrated Moving Average (ARIMA).