Abstract

The problem of minimizing risk in industrial plant settings is a research interest. Long hours in industrial control rooms induce fatigue, drowsiness, and ease of distraction, posing a risk of failure or catastrophe. Considering the critical importance industrial plants for the modern-day economy, the minimization of risk in such settings should be prioritized. Recently, deep learning has proven successful on a host of applications, including online recommendation systems, online language translation, image recognition, robotics, and medical diagnosis. This research aimed to develop a deep learning control system that learned to attain human performance on a task (a 1990's Microsoft DOS racing game called Car & Driver) with only optical input, serving as a preliminary prototype for controlling a power plant from only optical feed of dials and gauges. A supervised learning approach was taken; a convolutional neural network (CNN) was trained on 6 hours of pre-recorded human expert play data to return the optimal keystroke given a still frame image of the task. When evaluated offline on a test dataset of still image frames of the same task, the CNN achieved 91.5% accuracy. To evaluate the CNN online on the task itself, a novel hardware interface consisting of two connected Arduino boards was developed to facilitate CNN-task interaction. When all components were incorporated into a cohesive control system, the CNN achieved near-human performance on the in-game metrics “Average Speed,” “Top Speed,” and “Lap Time,” demonstrating the system’s capacity to achieve human-performance when applied to an actual power plant scenario.

Video Presentation