Dynamic characteristics of the internal flow field of a rotary centrifugal air classifier and pressure prediction through attention mechanism-enhanced CNN-LSTM

The flow field in an air classifier has complex dynamic characteristics, making it extremely unstable and difficult to monitor. Here, the time-domain pressure signals in an air classifier were measured at different rotational speeds and spatial locations using micro differential pressure dynamic sensors. The pressure signals were analyzed using the probability density function, standard deviation, and power spectral density. Moreover, a new type of predictive model combining the convolutional neural network, long and short-term memory, and the attention mechanism was proposed. The results show that, regardless of whether a screen cage is present, the pressure signals at different spatial locations in the gas-phase flow field have only one main frequency, which is caused by the quasi-forced vortex motion with the most intense fluctuations. However, PSD has only one main frequency under both load and no-load conditions, its value under load is 1.5 times higher, and the distribution of main frequencies is unaffected by solid loading at high rotational speeds (>40 Hz). Meanwhile, PSD and coherence analyses revealed that large eddies at the outlet induced low-amplitude pressure fluctuations in the high-frequency range. Finally, the proposed CNN-LSTM-A model outperformed traditional deep learning models (SVM and BPNN) in predicting the pressure fluctuations.