We are pleased to share that our team recently presented at the 18th Hellenic Database Management Symposium, which took place on July 1-2, 2024, in Athens, Greece.
Our presentation focused on "Exploring Unsupervised Anomaly Detection for Vehicle Predictive Maintenance with Partial Information," a study by Apostolos Giannoulidis, Anastasios Gounaris, and Ioannis Constantinou.
Presentation Highlights:
The talk focused on predicting maintenance needs in vehicle fleets to enhance safety and minimize downtime. While built-in alert systems from vehicle manufacturers often fail to notify drivers of potential issues, we explored how data analytics and real-time signals could address this problem. In a challenging real-world setting with limited and partial failure data, we proposed a non-supervised approach that detects behavioral changes related to failures without relying directly on raw signals, thereby handling variability in driving behavior and weather conditions.
Our solution calculates differences in the correlations of collected signals across two periods and dynamically creates reference profiles of normal operational conditions, effectively tolerating noise. Initial experiments showed promising results, achieving 78% precision in detecting nearly half of the failures, outperforming a state-of-the-art deep learning technique. Furthermore, we presented our approach as an instance of a broader framework, evaluating a wide range of alternatives.
Key Takeaways:
- Non-supervised anomaly detection approach for predictive maintenance
- Effective handling of driving behavior and weather volatility
- Dynamic creation of reference profiles for operational conditions
- Promising results with 78% precision, outperforming existing techniques
The symposium provided a fantastic platform to share our findings and exchange ideas with experts in the field. We are excited about the potential impact of our approach and look forward to further developments.
