- Demand Planning
- Machine Learning
Spare parts management is the backbone of asset intensive industries such as telecommunications companies, which operate in a highly competitive environment. Network reliability is a strategic goal as it ensures high customer service level and connectivity. Although companies utilize information related to the expected life of assets and plan maintenance activities, unplanned maintenance is still driven ad hoc. This has an impact not only on the company’s operations, inventory levels and cost but also on customers’ satisfaction. This capstone studies how telecommunications companies can improve the prediction of site failures and introduces a proactive maintenance approach. Based on our sponsor’s pilot project, we apply the MIT’s digital supply chain framework to define the value proposition and use the last 3 years of data to develop predictive models for site failures. To approach this case, we start by using the k-means algorithm and cluster the sites in three groups based on variability and demand for spare parts. To predict site failures, we apply time series models (exponential smoothing, Holt Winters and ARIMA) and assess the forecast accuracy based on RMSE and MAPE. In the last stage, we use supervised machine learning classification algorithms (Naive Bayes, Decision Tree, and Random Forest) and assess the accuracy using the correlation matrix. Based on our pilot project, we found that, while time series have a high percentage of error, machine learning algorithms can predict assets failures with accuracy between 60% to 85% and drive predictive maintenance and reduction of inventory levels and ageing. Nevertheless, companies should consider high quality and real time data prerequisites for machine learning. Our findings can be useful for other asset intensive companies that currently use traditional maintenance methods and are seeking to improve their predictive capabilities