GAF-AlexNet based Incipient Agglomerates Monitoring of Fluidized Bed Reactor*

For the complex chemical processes, traditional abnormal monitoring methods show low alarm accuracy and poor real-time performance due to the limitation of environmental interference and operational experience. Moreover, the monitored data are mostly time series with the important signal features masked by noise. This study proposes a kind of deep learning-based monitoring method to alarm the incipient agglomeration in a gas-solid fluidization polymer production. The proposed method takes the acoustic signal as the monitoring signal. First the time-frequency analysis is used to extract the fault features hidden in the original acoustic wave and enrich the tiny agglomeration information. Then, the time series in the time-frequency domain are converted into Gramian Angular Field (GAF) images which encode the one-dimension data into 2D pictures by preserving the temporal dependency features. Finally, the improved AlexNet is proposed to monitor the incipient agglomerates. The proposed method is verified in an actual gas-solid fluidized bed equipment, and the experimental results show that it effectively increases the fault detection rate and decreases the false alarm rate.