Comparison of data driven predictive maintenance models for a manufacturing company

Maintenance method is one of the most important problems for manufacturing enterprises. Maintenance strategies applied in the industry are mainly examined in two groups. These are corrective maintenance and preventive maintenance. In corrective maintenance methods, maintenance is performed when equipment fails, while in preventive maintenance methods, maintenance is performed before failure occurs.With the development of sensor technologies in production systems in recent years, data of production equipment such as pressure, temperature, vibration can be monitored continuously. This has led to the more widespread use of predictive maintenance techniques (performing / not performing maintenance according to the condition of the equipment), which is one of the preventive maintenance methods. However, it can be costly for small and medium-sized enterprises to install such sensor systems that can continuously monitor and record the status of production equipment. Within the scope of this thesis, it is aimed to propose a predictive maintenance model based on the loss data of production lines without such recorded data for production equipment. In the study, data, belonging to an enterprise that produces PVC profiles continuously for 24 hours, such as loss (kg) on the basis of shift / production line, production equipment operating speed differences and the number of shifts passed over the last maintenance, were used. First of all, a threshold value (kg) was determined for the maintenance according to the average maintenance cost based on the planned losses (losses due to maintenance). Then, models that will estimate the possible loss amount (kg) for the production line to be run in the next shift, were trained from the production loss data. Statistical learning algorithms such as linear regression, neural networks, random forest and gradient boosting were used to train the models. When the performance of the trained models was compared, it was seen that the most successful prediction model was the neural network. In the final stage of the thesis, it is explained how to decide whether to perform maintenance or not for the production line to be operated. According to the proposed method, the possible loss (kg) in the related production line will be calculated and the estimated possible loss (kg) will be compared with the determined threshold value. If the possible loss (kg) is greater than the threshold value, maintenance will be performed, if it is less than the threshold value, no maintenance will be performed. Keywords: Predictive Maintenance, Statistical Learning, Production Losses, Prediction Models