https://www.mdu.se/

Svealandstrafiken wanted us to perform predictive maintenance based on historical maintenance data. They provided us with files containing information for three buses, each containing the variables: article number of a spare part, mileage in kilometers, order number, date of the maintenance, the number of spare parts, maintenance code, price, and total price. The purpose for this research is to see if it is possible to perform predictive maintenance using maintenance records in order to improve the company's sustainability work from an economical perspective and improve their maintenance work. This type of research is interesting because of the sparse types of features in the maintenance records and it is not uncommon for businesses to lack the necessary information that is needed for predictive maintenance. Most related works we found used sensor data for predictive maintenance. We tried to perform predictive maintenance based on the similarities between the maintenance records. We started by preparing the data set using K-means clustering and created features that was used for similarity calculations. Later, we divided the prepared data set into 3 different sets: test, validation, and training set. After the data preparation, we performed k-nearest neighbors on the data, in order to find similarity between the records in the test and validation sets against the training set. For each row in the test and validation sets, similarity calculations would yield top five similar records that would be our predictions. As a result, we managed to predict which article numbers should be changed with a very low accuracy and precision, around 2 and 3\% respectively. We managed to predict which maintenance code the maintenance should focus on, with an accuracy of around 95\% and a precision around 80\%. In conclusion, while it is possible to perform predictive maintenance based on patterns in the maintenance records, it is inaccurate when predicting which articles should be changed, but gives an accurate prediction on which maintenance code the maintenance should focus on.