Network Fault Prediction Method based on CNN-LSTM-Attention

Aiming at the limitation of fixed time window in traditional network fault prediction methods, this paper proposes a CNN-LSTM-Attention hybrid model based on dynamic window optimization. Firstly, by designing an event density-driven dynamic windowing strategy, the model can dynamically adjust the observation range according to the log generation rate, node load state and log level, effectively balancing data sparsity and semantic integrity. Then the pre-processed data is fed into the CNN-LSTM-Attention model for prediction, which combines the local feature extraction capability of Convolutional Neural Networks (CNNs), the time-dependent modeling capability of Long Short-Term Memory Networks (LSTMs), and the focusing capability of the Attention mechanism for critical information. Experiments based on the real dataset GAIA show that the comprehensive performance of the proposed model under dynamic window is significantly better than that of the existing methods, and the F1-Score reaches 71.6%, which is 3.6% and 7.6% higher than that of CNN-LSTM and LSTM-only models respectively. This experiment shows good real-time performance and engineering applicability.