MIT PhD Student Develops Open-Source Anomaly Detection Framework Usable by Anyone

MIT PhD student Sarah Alnegheimish, working within Principal Research Scientist Kalyan Veeramachaneni’s Data-to-AI group at the Laboratory for Information and Decision Systems (LIDS), has developed Orion, an open-source machine learning framework designed for anomaly detection. Orion aims to make machine learning more accessible and trustworthy, particularly in large-scale industrial and operational settings.

Alnegheimish’s background heavily influenced her pursuit of accessible technology. Growing up in a home where education was highly valued, she witnessed the power of shared knowledge. Her experiences with MIT OpenCourseWare further solidified her belief in open-source resources.

Orion addresses the critical need for effective time series anomaly detection. By identifying unexpected patterns in data, Orion can provide vital insights across various sectors, including cybersecurity (detecting unusual network traffic), predictive maintenance (monitoring sensor readings in machinery), and healthcare (analyzing patient vital signs).

The framework utilizes statistical and machine learning models that are continuously logged and maintained. A key feature of Orion is its user-friendly design, allowing individuals without extensive machine learning expertise to analyze signals, compare anomaly detection methods, and investigate anomalies within a unified program.

Transparency is a core principle of Orion’s design. Each step in the model is labeled and presented to the user, fostering trust and understanding. “We label every step in the model and present it to the user,” Alnegheimish explains, emphasizing that this transparency builds user confidence in the model’s reliability.

Alnegheimish is also exploring the use of pre-trained models for anomaly detection within Orion. While traditionally, machine learning models require training from scratch on specific data, Alnegheimish is repurposing models initially designed for forecasting. By using prompt engineering, she aims to leverage the existing pattern-recognition capabilities of these models for anomaly detection without additional training.

Accessibility has been a guiding principle throughout Orion’s development. Alnegheimish has focused on creating abstractions within the system to provide a universal representation for all models, simplifying the process from raw input to desired output. These abstractions have proven stable and reliable over the past six years.

Orion’s impact is already significant, with over 120,000 downloads and a large number of users marking the repository as a favorite on Github. Alnegheimish notes that open source provides real-time adoption, moving beyond traditional metrics such as citations and paper publications.