Proactive Fault Prediction of Fog Devices Using LSTM-CRP Conceptual Framework for IoT Applications

Abstract

Technology plays a significant role in our daily lives as real-time applications and services such as video surveillance systems and the Internet of Things (IoT) are rapidly developing. With the introduction of fog computing, a large amount of processing has been done by fog devices for IoT applications. However, a fog device’s reliability may be affected by insufficient resources at fog nodes, which may fail to process the IoT applications. There are obvious maintenance challenges associated with many read-write operations and hazardous edge environments. To increase reliability, scalable fault-predictive proactive methods are needed that predict the failure of inadequate resources of fog devices. In this paper, a Recurrent Neural Network (RNN)-based method to predict proactive faults in the event of insufficient resources in fog devices based on a conceptual Long Short-Term Memory (LSTM) and novel Computation Memory and Power (CRP) rule-based network policy is proposed. To identify the precise cause of failure due to inadequate resources, the proposed CRP is built upon the LSTM network. As part of the conceptual framework proposed, fault detectors and fault monitors prevent the outage of fog nodes while providing services to IoT applications. The results show that the LSTM along with the CRP network policy method achieves a prediction accuracy of 95.16% on the training data and a 98.69% accuracy on the testing data, which significantly outperforms the performance of existing machine learning and deep learning techniques. Furthermore, the presented method predicts proactive faults with a normalized root mean square error of 0.017, providing an accurate prediction of fog node failure. The proposed framework experiments show a significant improvement in the prediction of inaccurate resources of fog nodes by having a minimum delay, low processing time, improved accuracy, and the failure rate of prediction was faster in comparison to traditional LSTM, Support Vector Machines (SVM), and Logistic Regression.