Vehicle Health Indicator Based on Cumulative Reconstructed Errors

Constructing a health indicator for vehicles is becoming important for enhancing reliability and stability of operation systems. Typically, health indicator construction relies on supervised learning using each device’s end-of-life data. However, obtaining complete lifecycle data for automotive vehicles is challenging because of their long operational lifespans. This study proposes an unsupervised learning approach using a long-short term memory-based variational autoencoder (LSTM-VAE) to construct health indicators for vehicles using only early-life driving data. The VAE model learns the distribution of input data through reconstruction, allowing the detection of changes in data distribution due to aging by monitoring reconstruction errors. However, because VAE models may struggle to capture subtle data variations, this study integrates the concept of cumulative sum (CUSUM) control chart, which effectively detects minor changes of data distribution over time. Experimental results show that the CUSUM-based health indicator outperform the reconstruction error-based indicator in both quantitative and qualitative assessments. Notably, the health indicator demonstrates a clear response to specific component repairs or replacements, validating the practical effectiveness of the proposed approach in real-world vehicle operations. By enabling early detection of potential failures, this comprehensive vehicle health indicator can help reduce unexpected downtime and maintenance costs, ultimately improving the reliability and safety of automotive vehicles. This study is the first to develop a comprehensive vehicle health indicator using early-life data, verified with actual vehicle data, highlighting the practical applicability of PHM technology in the automotive industry.