Protecting Pulsed High-Power Lasers with Real-Time Object Detection


Protecting Pulsed High-Power Lasers with Real-Time Object Detection

Kelling, J.; Gebhardt, R.; Helbig, U.; Bock, S.; Schramm, U.; Juckeland, G.

In Petawatt laser systems, firing at 10Hz, suddenly appearing scatterers can damage components. Damage(-spreading) can be avoided by suspending operation immediately on occurrence of such an event. This poster presents our approach for the automatic detection of critical failure states in real-time, employing state-of-the-art object localization on intensity profiles of the laser beam.

Learn, how we fitted the You Look Only Once (YOLO) approach, which is suited to low-latency object detection, to our problem and how we adapted the required multi-step training protocol to the available experimental data.
In this application accuracy trumps high recall, as false positives would severely impede productivity or even render our system useless. This had us refrain from general anomaly detection and thus we also present different ways in which we tune the object-detection for minimal false-positive rates.

  1. extended abstract
High-power lasers are operated at our research center for investigations of exotic states of matter and medical applications, among others. This project to improve the automatic shutdown/interlock system of two lasers (one in operation, one currently under construction) has the goal of reducing the probability of, potentially expensive, damage-spreading scenarios, while at the same time avoiding false alarms. In order to achieve high precision, we train for high recall only for known indicators, instead of using anomaly detection.
After we presented a proof-of-concept for this type of failure-state-detection at GTC 2018, where the main challenge was in dealing with a far too small dataset, we are now working on a pure deep-learning approach driven by systematic experimental data. In the new design, intended for production use, the classification takes place on differences between a running average of non-signaling images and the current shot. This is required, because no images can be obtained which can be classified as "good" without context. In order to achieve fast object-detection, to highlight potential problems for the operator, the you look only once (YOLO) approach[1], which we modify by removing the final layers for bounding-box prediction and train the network to directly produce an expressive feature map (lazy YOLO).
From this talk, the audience can learn how we adapted the well-known YOLO approach to our real-world application, from the employed network to the multi-step training protocol. Another topic is the design for short response times, to which end we employ Caffe, OpenCV on GPU and use C++ as main programming language instead of python.
[1] Redmon, J., Farhadi, A.: YOLO9000: Better, Faster, Stronger, ArXiv e-prints, 2016

Keywords: Image Classification; Caffe; automatic Laser-safety shutdown; Object Detection

Permalink: https://www.hzdr.de/publications/Publ-28992
Publ.-Id: 28992