A robot operational state recognition method based on SGMD-LSTM-Transformer hybrid network

Purpose The purpose of this study is to address the challenge that accurate recognition of industrial robot operational states from noisy long-sequence current signals remains difficult, as feature representations are easily degraded by noise and long-range temporal dependencies are hard to model. Design/methodology/approach This study proposes a Symplectic Geometric Mode Decomposition-Long Short-Term Memory (SGMD-LSTM)-Transformer hybrid network for robot operational state recognition. First, SGMD is used to decompose the current signals, and a weighted mutual information strategy is introduced to select the optimal mode combination for feature enhancement. Then, an LSTM network is used to capture temporal dynamics, while a stacked Transformer encoder is incorporated to strengthen long-range dependency modeling and sequential feature representation. Findings Experiments conducted on a selective compliance assembly robot arm (SCARA) industrial robot data set show that, compared with a conventional LSTM model, the proposed method reduces the Fluctuation and Label Jumps metrics by 13.5% and 42.3%, respectively, while also improving classification accuracy and prediction stability. Research limitations/implications This study focuses on planar motion states derived from current signals. The effectiveness of the proposed method under more complex three-dimensional motion patterns and more diverse industrial environments still requires further validation. Practical implications The proposed method improves the reliability and adaptability of industrial robot operational state recognition and provides useful support for robot condition monitoring, task analysis and maintenance decision-making. Originality/value This study develops a novel SGMD-LSTM-Transformer framework that combines signal decomposition, weighted information-based mode selection and hybrid sequential modeling for industrial robot operational state recognition, providing a new solution for improving both recognition accuracy and temporal stability.