Mid Sweden University

Industry 4.0 is the new era of digitization in manufacturing. Intelligent manufacturing uses integrated sensors to gain valuable insights in system dynamics and supply chain management. This work focused on methods for autonomous detection and classification of faults in industrial processes, with specific interest to the wood processing industry. The thesis is developed in collaboration with SCA and aims to investigate the possibility of implementing statistics and machine learning to speed up the troubleshooting process, avoid unnecessary downtime and maximize the yield. This study focused on the different methods used to analyze data, but do not investigated the data communications step, i.e., edge-, fog- and cloud computing. In this thesis a model for system diagnostics using machine learning and an algorithmic fault detection scheme is proposed. It is a Big-data-driven approach that handles imbalanced data in industrial IoT. Imbalanced data is a common problem in machine learning and means that the data distribution is skewed, i.e., a minority- and majority class where the number of observations is not evenly distributed. A binary classifier is used to detect quality deviations and the mathematics to identify where it has occurred. It is proven possible to implement machine learning for a system diagnostic purpose. This study recommend RUSBoosted Trees, Kernel Naive Bayes or Decision Tree to solve binary classification problems.

The importance of data is something that has proven to be of great relevance in this study, as a key component to enable big-data analysis for advanced diagnostics in the sawmill

is the availability of abundant sensor data from the production process. Then, this study highlights the requirements for such a sensor data collection, which enables the collection of training data to exploit the advantages of supervised-learning-based classification approaches.