Koneoppimiseen perustuvien virtuaalisten sensorien soveltaminen sellun jauhatusprosessiin

Refining is one of the most important steps in pulp manufacturing, aiming to improve the strength properties of the final product (board) by mechanically treating cellulose fibers using refining blades. The degree of refining is typically measured by Canadian Standard Freeness (CSF). However, since CSF is difficult to measure online, an inferential model for providing real-time information on CSF is desired. Due to the complexity and nonlinearity of the pulp refining process, and the availability of historical data, a machine learning based soft sensor is an attractive approach to model the process and predict pulp or board properties. In this thesis, a two-staged dynamic convolutional neural network (CNN) based soft sensor framework is developed for predicting the CSF after a low consistency refining phase in a semi-chemical (NSSC) pulping process. In the first stage, an unsupervised denoising CNN-autoencoder (CNN-DAE) is applied for learning a hidden feature representation of the input data. Moreover, the DAE serves as a robust anomaly detector for noisy data, simultaneously allowing to pretrain the model both with labeled and unlabeled (no CSF observation) samples. In the second stage, a supervised CNN with pretrained weights is trained to predict the CSF. In addition, a similar soft sensor for predicting pulp consistency is developed without the pretraining phase. The CNN-DAE soft sensor outperformed the benchmark models (incl. partial least-squares regression; PLS) in the freeness prediction, achieving satisfactory prediction accuracy using 15 carefully selected input variables, such as refining power, consistency, flow, etc. Lagged observations of the variables were included to the input samples to capture the dynamics of the process. Slightly better scores were obtained using all input variables (94), which demonstrates the superior ability of the CNN to automatically learn relevant features. The CNN soft sensor achieved good predictive performance in the consistency prediction, the results being comparable to the PLS. Furthermore, the CNN-DAE showed great potential in unsupervised anomaly detection, being able to detect abnormal process events and consistency sensor drift from the data. Overall, the results indicate that data-driven predictive models seem promising for the pulp and board industry when integrated with extensive process knowledge.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.