A systematic data science approach towards predictive maintenance application in manufacturing industry

Abstract

The use of data and obtaining insights from the data through data science approaches is invading every domain of industrial applications. Today, most of the industries harness the power of data to assess their performance and also to find any possible chance for improvement. This thesis project explores the data driven approach in order to perform a diagnostics and predictive analysis on the industrial alarm data collected from Volvo Group, using a systematic data scientific methodology. The company is currently facing with an issue of having machine stoppages due to the triggering of alarms, for which the cause is unknown. The machine performs a process of Part A1 and Part A2 used in the engines of the trucks. Therefore, the company is having challenge in finding the root cause of the alarms and also to get insights into correlation of alarms and the product types being processed in the machines, which is Part A1 and Part A2. Therefore, this thesis aims to find the combination of Part A1 arm and Part A2, which produces or triggers more alarms in the specified machine chosen for the study. CRISP-DM methodology is been followed to perform the data driven approach and thereby to answer the research questions put forward in the earlier stages of the thesis. The data is obtained from 2 sources, namely Manufacturing execution system and process data which is then integrated to perform the further analysis in which exploratory data analysis is performed and the list of Part A1 - Part A2 combination which triggers more alarms are been found out. Along with this, the behaviour of problematic alarms are analysed and more insights are obtained and reported. After the exploratory data analysis, the predictive analysis is performed using machine learning models in which the multi class classification models are generated using different machine learning models wherein the model predicted the alarm category based on the selected input variables used in the modelling. The decision tree model is selected based on the accuracy score and its corresponding rules are derived and the rule with higher accuracy is chosen to provide as a decision support to the company. As a practical contribution of this thesis, it is proposed that more historical data can be used for the analysis and also the data from the quality department can be used to explore the possible in-depth root causes. Finally, this thesis introduced the novelty of implementing predictive modelling in industrial alarm problems and hence produced an academic contribution.