Machine learning methods for Industrial processes

Abstract

The general machine learning algorithms are introduced and explained before being implemented in Python. A comparison of various algorithms are made, and their workings are visualized. Time series data is defined and analyzed using traditional techniques such as AR, MA, and ARIMA, and then with LSTM, a deep learning algorithm, which is a subset of machine learning.

The estimation of dynamic systems is considered extremely complex due to nonlinearity and, more importantly because dynamic systems are dependent on previous data. Nonlinear data is known to be handled by neural networks, and LSTM is a type of neural network that is known to handle sequential data such as text, speech, and time series data due to the presence of memory. These benefits make LSTM an excellent candidate for creating models that can replicate a complex dynamic model; a few case studies are provided.

Traditional linear models are matched up against newer machine learning models such as decision trees and SVM. The dynamic models estimated using LSTM methods are then compared to conventional system identification methods like ARX and nonlinear ARX. It was discovered that machine learning LSTM models could compete with traditional system identification techniques.

Anomaly detection has grown in importance in recent years in industries as it allows them to identify faults at an early stage and can aid in the prevention of large failures. Examples are used to explain and implement anomaly detection techniques such as One-Class Classification, Isolation Forest, and LSTM Autoencoders.