A study of Machine Learning for Predictive Maintenance - A topic and programming guidance
MetadataVis full innførsel
A significant potential and interest is found for Predictive Maintenance (PdM) and Machine Learning (ML). Both fields are under development and need further research. Through collaboration with the Norwegian maintenance company, Karsten Moholt, two areas of special significance for this thesis where found. Firstly, finding more information regarding the ML approach and especially the potential of Transfer Learning (TL), and secondly, understanding and building an ML model for PdM.This thesis aims to provide an introduction and to be practical start guide for those interested in using ML for PdM. To achieve this, necessary background information has been collected, and to further learn about ML, several online courses has been completed. The online course, books and several web-guides are used to build and develop two ML models to find Remaining Useful Life (RUL) on the Turbofan Engine Degradation Data Set provided by NASA. The information and code are presented to enable the reader to understand ML and to provide tool to start building ML models. PdM is presented as a technique for monitoring operating condition to provide data that can ensure the maximum interval between repairs and minimize the number and cost of unscheduled machine failures. ML is con sidered as an important and powerful tool for finding patterns and make predictions from a vast amount of data where Neural Networks (NN) is the main technique for implementing ML. TL is considered as a powerful idea, where knowledge from one NN can be transferred to another. Further, the topic of NN is presentment and explained with examples. To enable the reader to start programming, needed tools such as Python, Jupyter, Numpy, Pandas and Keras are presented, used and recommended. The first model provides a guide on how to program an NN oneself, and includes the elements; prepare data , initialize weights and biases , forward propagation , calculating cost , backpropagation and update parameters . The second model presents how to program the same model with Keras, a NN framework, also enabling one to build a ML model with few lines of code. This guide provides examples and tools that can be used for simple demonstrations of ML for regression problems. The thesis provides arguments and justification for the need for further research. It enables a foundation for further development of ML models for different PdM RUL scenarios.