A Data-Driven Semi-Supervised Soft-Sensor Method: Application on an Industrial Cracking Furnace

The cracking furnace is the key equipment of ethylene unit. Coking in tubes results from generation coke during cracking, which will compromise heat transfer efficiency and lead to shape change tubes. In order keep operating economically safely, engineers need decoke according surface temperature tube. However, tube difficult obtain practice. Due redundant instrumentation high level process control furnaces, a large number operation data have been collected, makes it possible predict based on autocorrelation cross correlation within among variables. Traditional prediction methods rely labeled samples for training, ignoring information contained vast amount unlabeled data. this work, data-driven semi-supervised soft-sensor method proposed. Considering nonlinear dynamic relationship variables, long short-term memory network (LSTM) autoencoder (AE), deep neural suitable feature extraction long-term series, used pre-training extract features Then, principal component analysis (PCA) mutual (MI) are applied remove select related target respectively. Finally, selected utilized establish model artificial (ANN). Data an industrial unit considered validate performance proposed method. show that error about 1% successfully aid determining optimal time decoking.