A Multimodal Sensor-Based Self-Supervised Learning Framework for Low-Noise System State Prediction and Anomaly...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Sensor Systems

Background:

• Multimodal sensor systems face challenges including noise, asynchrony, subjective labeling, and instability.
• Existing methods struggle with complex, real-world sensor data.

Purpose of the Study:

• To develop a low-noise system state prediction and anomaly detection method.
• To enhance model stability and interpretability in multi-source sensor systems.
• To leverage self-supervised representation learning for improved performance.

Main Methods:

• Uniformly modeled diverse data (environmental, device, network, operational, logs) as system state perception signals.
• Employed temporal masking for state structure modeling.
• Utilized state-oriented contrastive learning for representation constraints.
• Implemented task alignment strategies for prediction and representation.

Main Results:

• Achieved state-of-the-art performance in prediction and anomaly detection (MSE: 0.0167, MAE: 0.0856, RMSE: 0.1291).
• Demonstrated superior state-ranking (IC: 0.494, RankIC: 0.460) and discrimination (AUC: 0.815).
• Showcased high accuracy, precision, recall, and F1-score in classification recognition.
• Verified effectiveness of multimodal fusion, temporal masking, contrastive learning, and task alignment.

Conclusions:

• The proposed method offers robust, stable, and interpretable system state features.
• It exhibits strong noise resistance and practical application potential in complex sensor systems.
• Outperformed baseline models significantly in prediction and anomaly detection tasks.