Distributed Deep Fusion Predictor for a Multi-Sensor System Based on Causality Entropy.

Area of Science:

• Data Science
• Machine Learning
• Sensor Networks

Background:

• Trend prediction using sensor data is crucial but challenged by increasing data volume without performance gains.
• High-dimensional sensor data often contains noise and irrelevant information, hindering accurate predictions.
• Traditional deep learning models can overfit sensor noise, limiting their effectiveness.

Purpose of the Study:

• To address the challenge of improving prediction performance in multi-sensor systems with large datasets.
• To develop a distributed predictor capable of overcoming unrelated data and sensor noise.
• To enhance the accuracy and reliability of trend prediction models.

Main Methods:

• Defined causality entropy to quantify measurement causality.
• Proposed the series causality coefficient (SCC) for selecting high-causal input measurements.
• Employed Bayesian methods to obtain weight distribution characteristics for sub-predictor networks, preventing overfitting to sensor noise.
• Utilized a multi-layer perceptron (MLP) as a fusion layer to integrate results from sub-predictors.

Main Results:

• The proposed distributed predictor effectively models big measurement data from multi-sensor systems.
• Experimental results using meteorological data from Beijing demonstrated improved prediction performance.
• The method successfully filtered unrelated data and mitigated the impact of sensor noise.

Conclusions:

• The developed distributed predictor offers an effective solution for trend prediction in complex multi-sensor environments.
• Causality-based measurement selection and Bayesian-regularized sub-predictors enhance prediction accuracy.
• The approach provides a robust framework for leveraging big data from sensor networks.