Abstract
The ubiquitous adoption of Machine Learning (ML) and Artificial Intelligence (AI) solutions in a diverse range of industries has created a massive inrush in the demand to expand solutions into safety-critical industries such as manufacturing. While promising, some challenges presented by manufacturing processes include their real time nature, tremendous amounts of high-variability data, and humans in the loop which cause randomness and safety concerns. Some of the presented potential improvements of AI integration into manufacturing include improvements in efficiency, safety, and a deeper process understanding. To further the advancement of human knowledge on this subject this dissertation presents six different model architectures that were applied on the Amalgamated Sugar steam dryer to preempt and find the root cause of a process-stopping event known as a plug. The novel research contributions include an industry-focused project of deploying AI on real time (10 second interval) manufacturing data, unique data preprocessing, a time-focused AdaBoost decision tree model, a hybrid model, time focused scoring metrics, model comparisons, and a custom Artificial Intelligence Predictive Appliance (AIPA). The data for this dissertation is grounded in the real world and has high noise and randomness introduced by sensors and actuators along with operators making changes. This requires a large amount of novel data preprocessing to format the data into a viable format for the models. The model architectures include an AdaBoost decision tree ensemble, a custom AdaBoost offset decision tree ensemble, long short-term memory (LSTM) model, a custom hybrid model, neural basis expansion analysis for interpretable time series forecasting (N-BEATS), and a temporal fusion transformer (TFT). The novel hybrid model includes a LSTM model performing regression predictions on the important features fed to a decision tree performing classification. All model types are compared with both standard scoring metrics and author-created time-focused scoring metrics. One highly important metric for this dissertation was the explainability aspect due to the safety concerns. Conclusively, for this dissertation the machine learning-based AdaBoost offset decision tree ensembles performed the best for event preemption and root cause analysis. This analysis included finding a slide gate on the beet input side of the steam dryer, which was the primary indicator of an issue event. Lastly, these models provide resilience to process changes by using a minimal number of features. The hybrid models show high promise for event preemption but suffer from both computational limitations and a lack of resilience due to the use of the full feature set. All deep learning models (LSM, N-BEATS, and TFT) struggled with event preemption and are unable to be used due to the aforementioned lack of resilience. The developed models are integrated into the AIPA, which is deployed at the Amalgamated Sugar factory.