Introduction
As computing environments scale to massive proportions, the complexity of manual system management has become unsustainable. "Self-healing" systems aim to mitigate this by autonomously detecting and recovering from faults. However, most current implementations are reactive, often requiring human intervention to identify the actual root cause of a failure. This paper addresses the challenge of moving toward proactive diagnostics by identifying specific configuration changes or hardware metrics responsible for system downtime.
Traditional approaches using Artificial Neural Networks (ANNs) or Hidden Markov Models (HMMs) often struggle with accuracy or fail to account for the historical context of system behavior. This research proposes a novel methodology leveraging Restricted Boltzmann Machines (RBMs) to predict sequences of system behavior, significantly improving root cause identification.
The Anomaly Detection Framework (ADF)
The authors developed an Anomaly Detection Framework (ADF) designed to operate in a continuous loop. The framework prioritizes efficiency through a "lazy evaluation" strategy, only triggering heavy machine learning analysis when high-level "fitness tests" (e.g., verifying if a web server is serving pages) fail. This approach recovers significant CPU resources—approximately 25 seconds per minute—that would otherwise be consumed by constant background analysis.
When a fault is detected, the system utilizes an RBM trained with Contrastive Divergence Learning (CDL). The RBM compares the current faulty state against historical "known-good" configurations to predict what system features should look like. Deviations from these predictions are flagged as "leads" and ranked by confidence to assist in troubleshooting.
Experimental Evaluation
The methodology was tested in a controlled environment using a Windows 7 Virtual Machine running IIS 7.5. Data was gathered via Windows Management Instrumentation (WMI), tracking 6,000 system features such as CPU usage, memory, and network throughput.
Two types of faults were injected to test the system:
- Adverse Configuration Changes (ACCs): Such as disabling network interfaces or sabotaging DNS settings.
- Direct Fault Injections (DFIs): Crashing services by injecting code into their address space.
The RBM's performance was benchmarked against standard ANN and HMM models to evaluate its effectiveness in lead generation and precision.
Results and Discussion
The RBM-based approach demonstrated a clear advantage in both accuracy and precision. While traditional models like HMMs and ANNs struggled to maintain precision above 30%, the RBM consistently achieved precision rates between 80% and 100%. This means that the leads generated by the RBM were far more likely to point directly to the actual root cause of the fault.
Furthermore, the RBM was significantly better at placing the actual root cause at the top of the generated leads list. Although the RBM requires more computational time (ElapsedTicks) to generate leads once a fault is detected, this is considered a beneficial trade-off given the massive reduction in background CPU overhead and the high quality of the resulting diagnostics.
Conclusion and Future Work
By utilizing RBMs to understand the trajectory of system behavior rather than isolated snapshots, this framework provides a robust "brain" for self-healing systems. Future research will focus on stabilizing RBM variability through larger datasets and expanding the framework to be "network-aware," allowing multiple machines to share learned data to identify external faults. Additionally, the authors suggest exploring Deep Belief Networks and Long Short-Term Memory (LSTM) networks for even more sophisticated sequence prediction.
Summary updated