Comparison of Chronos and Conventional Models: Predicting Machine Downtime using Time Series

Abstract

This study analyzes the comparison between a pretrained transformer model (Chronos) and conventional models in predicting industrial machine downtime using time series data to achieve greater accuracy and efficiency for companies. More specifically, this research focuses on early detection before downtime occurs to reduce company losses in terms of both costs and product quality, and to ensure that Key Performance Indicator targets are met. Design/methodology/approach: The research methodology includes primary data collection, data preprocessing, and sequential data splitting (80% training, 10% validation, 10% testing) to prevent potential data leakage. Model evaluation is measured using the Mean Absolute Error loss function, focusing on the “handling machine” category, which yields 4,069 to 4,101 data rows after the preprocessing stage. Research showed that the conventional XGBoost model with tuning performed best, with the lowest Mean Absolute Error among the other models. XGBoost proved to be highly effective and was capable of outperforming advanced transformer-based models (such as Chronos), particularly when applied to a limited dataset of 4,069 data points. Conversely, transformer architectures like Chronos performed poorly on small datasets because they were designed for massive datasets. This study focuses on the application and evaluation of modern artificial intelligence technologies, specifically transformer architectures such as the Chronos model. Although previous similar studies have successfully predicted downtime accurately using conventional models (such as ARIMA, Random Forest, Support Vector Machine, and autoencoders), those earlier studies have not tested the effectiveness of transformer architectures in detecting machine downtime.

GS Cited Analysis