Deep neural networks smarten inspection of nuclear power plants, other aging infrastructure
Nuclear power is a vital component of a low-carbon, sustainable energy future. The United States already is the largest supplier of nuclear power, with 99 reactors in 30 states. But our nation has an aging infrastructure, and inspection is crucial to keep current with maintenance, catch issues at the earliest stages, and remediate them for safety and performance.
My Purdue research team is applying deep neural networks to improve inspections.
Traditional inspection is a tedious, time-consuming, subjective — and intermittent — approach that depends on an inspector’s experience and focus on the task. Nuclear reactors are submerged in hot water with radiation. Looking at videos of the assets that have been captured by robotic arms and machine vision, inspectors try to identify and assess what are often very tiny cracks caused by erosion, corrosion, oxidation and other processes. This is a qualitative — not quantitative — method, and therefore not optimally exact or reliable.
For better results, we use artificial intelligence (AI) to analyze the images collected by robots. We employ deep neural networks to go through this “big data,” frame by frame, to indicate where the damage is, quantify its severity, and assign a probability value that represents the level of confidence in the accuracy of findings.
In addition to developing these deep learning algorithms, another innovation is that we fuse the information extracted from multiple image frames related to the same crack based on Bayesian inference, a method of statistical inference. This breakthrough combination of deep learning and statistical inference provides better damage detection and fewer false positives than traditional AI approaches. We have recently improved the system to enable it to process the image frames online, in near real time, as the data is captured.
This solution is helping us usher in autonomous inspection, in which a deep learning network continuously feeds analysis and recommendations as it processes the images. We are aiming toward letting the system make decisions on its own, determining such things as how to move and position the camera in the most effective manner to collect the data and derive the analytics.
The system could then issue alerts to the human inspector about the rate of a crack’s change and amount of deterioration, so the technician could focus on more problematic areas, conserving time and resources. In this way, the inspections could be conducted more frequently, and the measurements could be archived to build a history of the asset, to better evaluate the evolution of changes over time.
Deep neural networks are inspired by the human brain. We are using fully “convolutional” networks — which automatically extract features and weigh their importance to differentiate one thing from another — to segment cracks. We set up a training set with annotated labels that show the computer where the cracks are. The deep neural network is then trained, learning to distinguish between the crack and its background — somewhat like how a baby learns to differentiate between things by extracting and weighing various features that, taken together, characterize differences.
The advantage of deep learning is that the computer learns these features on its own based on the input data — of which it needs large amounts. We have worked with the Electric Power Research Institute, a nonprofit organization that provided us with inspection data. We are currently collaborating with companies globally to test the system with their data and nuclear reactors, with an eye toward commercialization and adoption into their inspection routines. Two patents have been published about this innovation.
The applications for this technology in the United States, with its aging infrastructure, are limitless. For instance, this technology is ideal for incorporation into unmanned aerial vehicles (UAVs) for inspection purposes. We also are looking to widen the system’s use to encompass buildings, sewer pipelines, roads, dams, and wind turbines, as well as additional critical assets in civil infrastructure, aerospace, oil and gas, and other sectors.
This is the future of AI: autonomous systems helping humans perform crucial tasks more accurately and efficiently, by learning from massive, exponentially expanding volumes of big data that can’t be processed into useful knowledge by the human mind alone.
Mohammad R. Jahanshahi
Assistant Professor of Civil Engineering, Lyles School of Civil Engineering
Assistant Professor of Electrical and Computer Engineering (Courtesy)
College of Engineering, Purdue University