Dispersion transition entropy: dual-modal feature fusion based on global attention mechanism for time series feature representation

Abstract A new entropy measure-Dispersion transition entropy (DTE) has been proposed for the analysis of complex dynamic signals in mechanical fault diagnosis. The DTE method combines the static pattern analysis of Dispersion Entropy (Normalization cumulative distribution function-based symbolic mapping) with the dynamic interval characterization of Global ordinal pattern attention entropy (GOPAE), enabling enhanced feature extraction under noisy conditions. Through systematic evaluations on simulated signals (AR (1), Chirp, MIX) and real-world vibration data (bearings and planetary gears), DTE demonstrates superior stability and discriminative power compared to six existing entropy methods. Experimental results show that the DTE-ELM framework achieves the highest classification accuracy (93.7% mean) with the lowest standard deviation (1.2%), while maintaining competitive computational efficiency (3.88 s). The proposed method provides an effective solution for improving fault detection performance in rotating machinery.