BiLSTM-MLAM: A Multi-Scale Time Series Prediction Model for Sensor Data Based on Bi-LSTM and Local Attention...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Accurate time series prediction is crucial for various applications.
• Existing models often struggle to capture complex patterns across multiple time scales.

Purpose of the Study:

• To introduce BiLSTM-MLAM, a novel multi-scale time series prediction model.
• To enhance the accuracy and robustness of time series forecasting.

Main Methods:

• Utilizing bidirectional long short-term memory (BiLSTM) for temporal feature extraction.
• Implementing a multi-scale patch segmentation module for diverse time scale pattern recognition.
• Employing a local attention mechanism for enhanced feature weighting and fusion.

Main Results:

• BiLSTM-MLAM demonstrated superior performance over six baseline methods across multiple datasets.
• Achieved significant improvements in Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) for aircraft engine life prediction, LTE, and load datasets.
• Ablation studies confirmed the positive contribution of each model component to prediction accuracy.

Conclusions:

• The BiLSTM-MLAM model effectively captures sequence information across multiple time scales.
• The multi-scale fusion strategy, combining different segment lengths, enhances prediction accuracy.
• BiLSTM-MLAM offers a powerful solution for complex time series prediction tasks.