中国有色冶金

Aiming at the problem that the production state of rare earth electrolytic cell is difficult to perceive, which leads to the inability to accurately control it and affects stable production, this study proposes a hybrid neural network diagnosis model based on signal feature extraction to identify the operation state of electrolytic cell. Firstly, the normalized cell voltage signal is obtained by calculating the cell voltage and cell current, and its time-frequency analysis is carried out to extract the energy values and energy ratios of each frequency band with the time-domain and frequency-domain features in different cell states, and to construct the feature data set. Secondly, a cell state diagnosis model (CNN-BiLSTM-SA) combining convolutional neural network (CNN), bidirectional long-term and short-term memory network (BiLSTM) and self-attention mechanism is established. The local spatial features are captured by CNN, the global dependency is modeled by BiLSTM, and the feature relevance is enhanced by the self-attention mechanism to realize the state diagnosis of the electrolytic cell. Simultaneously, improvements were made to the hippo algorithm by incorporating optimal point set initialization alongside crossover and mutation operations, aiming to enhance the algorithm's hyperparameter optimization capabilities. Finally, the improved hippo algorithm is used to optimize the hyperparameters of the model to improve the diagnostic accuracy of the model. Industrial validation experiments demonstrate that the proposed method achieves a diagnostic accuracy of 94.14% and a Macro-F1 score of 91.08% in cell condition diagnosis. Compared to the unoptimized model, these results represent improvements of 8.28% and 14.19%, respectively. It has advantages in diagnosing the state of rare earth molten salt electrolysis cell, and can provide a basis for optimal control and stable operation of rare earth molten salt electrolysis process.