PhD "Self-supervised anomaly detection and explanation in temporal data" M/F

about the role

Your role is to conduct a PhD thesis on: "Self-supervised detection and explanation of anomalies in time series data."

Global Context and Problematic of the Subject
Orange operates a large number of connected devices, from the core network to customers' LAN. The number of failures and attacks that can occur is significant, and it is becoming increasingly difficult to understand these issues, which incur substantial costs for customer services. So far, state-of-the-art work and research within Orange have focused on detecting anomalies in time series data, particularly in the communications of these devices. This detection results in a binary outcome—anomaly or not—and provides no information about the origin of the detection or any other characterization of the anomaly and/or its potential cause.

The explored solutions do not allow for action on detected anomalies, and an expert is still needed to find correlations between anomalies and hypothesize about their origins and corrective procedures.

More recently, significant advances in large language models (LLMs) promise interesting perspectives on interpretability issues. With an anomaly characterization model, it would be possible to relate detected anomalies to known issues.

1. Najari et al., RADON: Robust Autoencoder for Unsupervised Anomaly Detection, SIN 2021
2. Najari et al., Robust Variational Autoencoders and Normalizing Flows for Unsupervised Network Anomaly Detection, AINA 2022
3. Darban et al., Deep Learning for Time Series Anomaly Detection: A Survey, ACM Computing Surveys 2025

Scientific objectives – expected results and challenges
The objective of this PhD is to propose self-supervised deep learning approaches for the detection and characterization of anomalies in time series data. These characteristics should be leveraged to enable the identification of detected anomalies, for example, by utilizing business knowledge databases. The aim is to propose an end-to-end solution for anomaly detection and interpretation without human intervention.

Scientific challenges:

• Heterogeneous data and anomalies
• Uncertain presence of anomalies in the data
• Difficult correlation between anomaly characterization and business knowledge
• Potentially unknown anomalies

 Expected results

• A robust self-supervised anomaly detection model that characterizes the discrepancies between the observed sequence and the expected normal sequence
• An LLM+RAG-based approach to relate the characterization of detected anomalies with business knowledge
• Experimental evaluations on public and Orange data
• Publication of scientific articles, patents, and the thesis manuscript