Nuclear plant inspection trials reveal machine learning gains

Researchers are using advanced analytics programs to increase the accuracy and shorten the duration of reactor examinations ahead of real-life data tests.

The increasing power of computers with graphics processing units is opening up new ways to improve inspection efficiency. (Image credit: Archy13)

Related Articles

While data analytics is being widely deployed to optimize plant output, machine and deep learning programs are also set to help reduce inspection costs.

Flaw detection studies led by the U.S. Electric Power Research Institute (EPRI) and Purdue University have demonstrated impressive results by using automated analytics technology, researchers told Nuclear Energy Insider.

At present, the data from materials inspections is usually collected in a semi- or wholly automated manner but it is then analyzed manually. This is time consuming and subject to human error because the work is often tedious. Currently, staff have to plough through hours of footage from inspection cameras to identify cracks before they become a safety concern.

Most existing systems rely on detecting cracks within a single image but these can be obscured by the play of light and shadow, while it can be difficult to differentiate tiny cracks from scratches.

Human errors in data analysis can result in additional examinations to characterize or monitor an indication of a flaw. In addition, there is a risk of unnecessary repairs being carried out if an indication is improperly evaluated.

The automation of data analysis through analytics removes human error and provides a faster, more reliable detection capability, John T Lindberg, program manager for NDE [non-destructive examination] Innovation at EPRI, told Nuclear Energy Insider.

EPRI is currently researching automated data analysis processes for NDE, including reactor vessel internal remote visual examinations, reactor vessel head ultrasonic examinations and heat exchanger tubing eddy current examinations.

“In the long term, these NDE reliability improvements can reduce component inspection and repair costs,” Lindberg said.

Inspection costs could be reduced by speeding up the data analysis process and reducing the numbers of data analysts needed to examine the data acquired in the field.

“Typically, the manual data analysis process takes longer than the time required to perform the actual examination,” Lindberg noted.

In addition, increasing flaw detection rates and reducing the number of false calls will help operators reduce the cost of lifespan extensions. According to Lindberg, improvements in NDE reliability will "build a stronger case for assuring component structural integrity” in lifespan reviews.

                            Profit margin advantage of AI adopters
                                                            (Click image to enlarge)

Source: McKinsey Global Institute (June 2017)

Neural network

A team at Purdue University in Indiana is using a deep learning framework, called a naïve Bayes-convolutional neural network, to analyze individual video frames for crack detection.

An innovative “data fusion scheme” aggregates the information extracted from each video frame to enhance the overall performance and robustness of the system.

Led by Mohammad Jahanshahi, Director of Purdue University’s Smart Informatix Laboratory and Assistant Professor in Purdue’s Lyles School of Civil Engineering, the research team is developing a CRAQ (crack recognition and quantification) system which combines accelerated graphics processing with deep learning and machine learning. Information is collated from different video frames to identify changes in steel textures via a naïve Bayes-convolutional neural network.

Convolutional neural networks are already used in facial and speech recognition.

The data fusion algorithm system is designed to track cracks from one frame to the next, providing “more robust decision making” than current processes, Jahanshahi told Nuclear Energy Insider.

Trials showed the CRAQ system had a 98.3% success rate in tracking cracks, he said.

The samples, which were provided by EPRI, were scanned at 30 frames per second.

“With the increasing computational capabilities of computers with GPUs [graphics processing units], we can now use computer vision, image processing and deep learning to tackle this problem, Jahanshahi said.

The researchers have filed a patent application for the crack-detection technology.

Growing data insights

The widespread transfer of data from physical files to digital media and increasing granularity and depth of data sets provides further opportunities for analytics and artificial intelligence efficiencies.

“EPRI research indicates that the use of automated data analysis processes through use of data analytics technologies could be integrated into other NDE examinations where the examination data is acquired and recorded on digital media,” Lindberg said.

Major nuclear operators have already embraced data analytics technologies in a bid to reduce operating costs. The nuclear sector is also collaborating with other power sectors on the use of data analytics for other applications in power plant generation and transmission monitoring as well as inspections. 

A key hurdle for the new technology developers will be regulatory approval.

“One of the main challenges to implementation of this [NDE] technology is proving to the industry, first and then the regulator, that the results from automated data analysis technologies are equivalent to data analysed by humans,” Lindberg said.

EPRI's future projects and collaborations will focus on refinement and testing of data analysis algorithms and processes. The research institute is currently collecting additional field data sets from utilities that will be used in a machine learning environment to train algorithms to improve flaw detection rates and reduce false calls, Lindberg said.

“Once the algorithms are refined and tested, we will beta-test the data analysis algorithms in pilot applications on real examination data scenarios,” he said.

By Neil Ford