Here's the link to the dataset mentioned in the paper.
The problem of minimizing risk in industrial plant settings is of high interest for research. 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 have on 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 match human performance on a simulated 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 ~90% 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.” This demonstrates that this system can achieve human-performance when applied to an actual power plant scenario.
A project to create a deep learning control system that learns by an air-gapped optical input stream rather than raw pixels. Applications include all sorts of infrastructure with legacy hardware (flood control, manufacturing plants, power plants).
Specifically, below is the configuration.
This will store corresponding image frames and keypresses at every timestep into .h5 files in a "dataset" folder. It will store it in batches of 500 frames as one .h5 file, and to toggle data collection press the spacebar.
Since the policy_collect.py script produces extraneous blank frames, run deleteblankframes.py to delete them to prepare for policy network training.
Drag collected data in the /sim-datasets/datasetname/fixed/fixed/ to /dataset/, renaming it with dataset_rename.py if needed.
Run policy_train.py to train on data collected in /dataset/.
Run policy_run.py to run on the trained neural network and control the simulator in real-time.
