Multivariate Comparative Analysis of Statistical and Deep Learning Models for Prediction Hardware Failure

Abstract

The twenty-first century is witnessing a transformative shift toward digitalization, where various services are undergoing a paradigm change from traditional to digital platforms. This digital revolution, evident in the transition from commerce to e-commerce and agriculture to agrotech, necessitates the development of robust and sustainable systems to support these applications. While cloud providers like Amazon, Azure, and Google Cloud offer Infrastructure as a Service (IaaS) solutions, many organizations still prefer on-site hardware maintenance. The existing studies on system failure prediction have primarily focused on either machine learning techniques like random forest and Naive Bayes or deep learning neural networks like RNN, LSTM, and CNN. However, there has been no comprehensive comparison of prediction accuracy between statistical models and deep learning models. This research aims to identify the algorithm that yields the highest prediction accuracy while considering hardware resource utilization, including CPU, RAM, ROM, and network usage. The study utilizes multivariate time series analysis to predict system failures based on the “BitsBrain” dataset. The models evaluated include ARIMA, auto-regression, SARIMAX, exponential smoothing, LSTM, and Bi-LSTM. The results demonstrate that the Bi-LSTM model outperforms the other models, achieving the lowest mean absolute error (MAE) and the highest R² score. Conversely, the exponential smoothing model exhibits the poorest performance. Additionally, a comparison of actual versus predicted plots reveals that the Bi-LSTM model generates more accurate predictions. These findings suggest that the Bi-LSTM model can serve as a valuable tool for forecasting system failures and enhancing system reliability. Its superior prediction accuracy, coupled with relatively efficient resource utilization, makes it a compelling choice for organizations seeking to optimize their system performance and minimize downtime.