# Real-time motor operating state recognition via multi-sensor fusion: A wavelet–neural–evidence framework for industrial condition monitoring

## TL;DR

This paper introduces a real-time system using sensors and advanced algorithms to monitor motor conditions in industrial settings, improving accuracy and reliability.

## Contribution

A novel wavelet–neural–evidence framework for multi-sensor fusion in motor state recognition is proposed.

## Key findings

- The fusion-based model achieved 92.8% overall accuracy in motor state classification.
- The system outperformed single-sensor baselines in classification robustness and confidence.
- The framework is scalable and suitable for practical predictive maintenance applications.

## Abstract

Accurate and real-time monitoring of motor operating states is essential for ensuring the reliability, safety, and efficiency of modern industrial systems. This paper presents a multi-sensor fusion framework for intelligent motor condition monitoring, which integrates wavelet-based feature extraction, shallow neural network classification, and evidence-theoretic decision fusion. A compact hardware platform is developed to synchronously acquire vibration, acoustic, and magnetic field signals under multiple motor operating conditions. The acquired signals are segmented using sliding windows and decomposed via wavelet packet transform to extract energy distribution features. These features are independently processed by BP neural networks trained on individual sensor modalities, and their softmax outputs are fused through Dempster–Shafer theory to enhance classification robustness and confidence. The proposed system is evaluated on a structured dataset covering three typical motor states: shutdown, no-load, and loaded operation. Experimental results show that the fusion-based model achieves an overall accuracy of 92.8%, outperforming all single-sensor baselines. Additional analyses of confidence distribution, confusion matrix, and ROC curves confirm the superiority of the proposed method in decision reliability and class separability. The developed framework offers a scalable, interpretable, and deployable solution for intelligent motor health monitoring and has strong potential for practical implementation in predictive maintenance applications.