Reliability prediction using a weighted temporal convolutional autoencoder based on limited claim data

Abstract

Many product manufacturing companies offer warranty services that cover costs incurred due to product failures during the warranty period. To maximize savings on warranty costs, researchers have attempted to predict field reliability using short-term claim data. Due to the limited information available, however, the prediction performance for long warranty periods has been unsatisfactory. This study proposes a weighted temporal convolutional autoencoder (WTCAE) model designed to predict the number of claims and field reliability over the entire warranty period using limited initial claim data. The WTCAE model compensates for the limited information from initial claim data by effectively capturing temporal patterns through a temporal convolutional network-based encoder–decoder structure. The proposed WTCAE model demonstrated superior performance even under conditions of short-term claim data, where traditional lifetime distribution-based methods fail to provide predictions. It also consistently outperformed conventional deep learning-based methods. The effectiveness and practicality of the proposed WTCAE model were validated using real-world data from millions of televisions and refrigerators, confirming its consistent performance across various data conditions within the warranty period.