A hybrid approach of ConvLSTMBNN-DT and GPT-4 for real-time anomaly detection decision support in edge–cloud environments

Abstract

Anomaly detection is a critical requirement across diverse domains to promptly identify abnormal behavior. Conventional approaches often face limitations with uninterpretable anomaly detection results, impeding efficient decision-making processes. This paper introduces a novel hybrid approach, the convolutional LSTM Bayesian neural network with nonparametric dynamic thresholding (ConvLSTMBNN-DT) for prediction-based anomaly detection. In addition, the model integrates fine-tuned generative pre-training version 4 (GPT-4) to provide human-interpretable explanations in edge–cloud environments. The proposed method demonstrates exceptional performance, achieving an average $F1-score$ of 0.91 and an area under the receiver operating characteristic curve ($AUC$) of 0.86. Additionally, it effectively offers comprehensible decision-support explanations.