Entropy space quantum behaved dung beetle optimized long short-term memory network for industrial process fault...

To address the challenges posed by the high dimensionality and strong nonlinearity of complex industrial process data, as well as the inherent defects in feature learning and hyperparameter dependency of the traditional long short-term memory network (LSTM), this paper proposes a novel hierarchical fault diagnosis method termed entropy space quantum behaved dung beetle optimized LSTM (ES-QLSTM). The proposed method employs kernel entropy component analysis (KECA) to extract features from raw data in the entropy space (ES). Benefiting from the strong nonlinear mapping capability and high-order information retention property of the KECA, the proposed method effectively preserves critical high-order statistical information that cannot be captured by conventional methods, which compensates for the insufficient feature mining ability of the LSTM. Then, this proposed method integrates entropy space feature extraction with a quantum behaved mechanism into the dung beetle optimization algorithm, constructing a systematic synergistic framework for entropy space feature dimensionality reduction and intelligent hyperparameter optimization. This framework solves two key shortcomings of the LSTM simultaneously, weak feature extraction from high-dimensional nonlinear data and unstable performance caused by empirical hyperparameter settings. Subsequently, a quantum behaved mechanism is incorporated into the dung beetle optimization (DBO) algorithm, resulting in the quantum behaved dung beetle optimization (QDBO) algorithm, which is utilized to optimize the key hyperparameters of the LSTM network. The ultimately constructed ES-QLSTM method enhances diagnostic performance through the synergistic integration of entropy-space feature purification and intelligent hyperparameter optimization. Experimental validation on the Tennessee Eastman Process (TEP) and grid-connected photovoltaic system (GPVS) datasets demonstrates that the proposed method outperforms traditional models and improved models in diagnosing complex faults, with the average diagnostic accuracy reaching 93.70% and 83.91% on the two datasets, respectively.