Autonomous Incident Detection on Spectrometers Using Deep Convolutional Models.

Area of Science:

• Analytical Chemistry
• Computer Vision
• Deep Learning

Background:

• Scientific instrumentation, like spectrometers, relies on effective sample introduction via spray chambers.
• Poor sample introduction, indicated by liquid beading and flooding, compromises analytical accuracy and can damage equipment.
• Existing deep learning object detection struggles with detecting small, transparent, and non-rigid incidents like beading and flooding.

Purpose of the Study:

• To develop novel computer vision frameworks for autonomous detection of beading and flooding incidents in scientific instrument spray chambers.
• To improve the operational accuracy and efficacy of chemical analysis by addressing sample introduction quality.
• To provide a real-time solution for preventing instrument malfunction and damage.

Main Methods:

• Proposed two deep learning frameworks integrating modern architectures with expert knowledge.
• Implemented region of interest localization and output refinement for incident detection.
• Utilized data augmentation, synthesis, and negative sampling for enhanced accuracy and real-time inference.

Main Results:

• Achieved high detection rates: 95% for beading and 98% for flooding.
• Surpassed widely used object detection baselines in performance.
• Demonstrated real-time processing capability at four frames per second.

Conclusions:

• The developed deep learning approach effectively detects critical incidents in spray chambers.
• This method offers a significant advancement for autonomous monitoring and control in analytical chemistry instrumentation.
• The system provides a robust, real-time solution for improving instrument performance and reliability.