Machine Learning Vision Techniques for Nondestructive Evaluation
26 Mar 2019
Limiting the human error factor of phased array testing data analysis
Nondestructive evaluation (NDE) has progressively required a higher level of automatization to be able to handle the amounts of data generated with modern data acquisition systems. Data analysis, especially the detection and classification of defects and faults, has traditionally relied on the skills of experts and typically requires many hours of manual labor to get reliable results. As computer power has increased, many classification machine learning vision algorithms have become available to speed up this time-consuming process. These algorithms reduce errors and misclassifications due to human error. Intertek offers an automatic defect recognition system methodology for NDE data using artificial neural networks/deep learning and similar algorithms for pattern recognition-based automated-signal classification. This system is capable of training models using Intertek's own (and other) databases on NDE data. This approach can lead to improved efficiency and higher reliability of detection when analyzing large amounts of sensing, monitoring, and NDE data that will be beneficial to all users when reducing the tedious work required on traditional techniques applied in the NDE field.
Intertek has proven experience with advanced analytics using large amounts of data in numerous projects, including power plant monitoring data. Our group has used data analytics and machine learning to show the impact of cycling data on dissimilar metal weld joint failures in recent work.
Intertek's AwareTM platform is an inspection data management software that is currently tracking inspection data for piping, pressure vessels, turbines, boilers, etc. for more than 800 power plant units. This program includes risk-based inspection tools, corrosion monitoring, half-life calculations, inspection scheduling, and turnaround planning widgets to create recommendations and repair tracking systems. Aware has recently merged with Intertek's INGRID, a web-based analytical platform that utilizes the latest in data manipulation technology and features the most recent benchmarking algorithms to provide valuable insights into asset performance, efficiency, and reliability metrics.
Intertek's approach is split in several phases.
- Phase 1 will require the collection of large amounts of NDE data from different sources. The data will require cleaning and will have to be formatted for its use in standardized environments.
- Phase 2 deliverables will include a survey and case studies on several machine learning and artificial intelligence techniques applied to NDE generated data in Phase 1. These techniques will be initially supervised methods and algorithms that comprise feature extraction, outlier detection, dimensionality reduction, and multi-parameter classification on training sets that are labeled from previous works as pretrained models. The objective of this phase is to create an overview of all machine learning vision techniques and artificial intelligence algorithms applicable to NDE testing. Figure 1 shows an example for a Neural Networks classification process for the evaluation of ultrasonic phased array testing (PAUT) indications. This classification scheme can also be applied to other NDE techniques.
- Phase 3 will require adaptation of these techniques and any other existing machine vision techniques to the features generated during Phase 2. During this stage, we will use the output of Phase 2 to improve the detection rates and reduce the noise level with the objective of generating a subset of non-supervised techniques that will allow the expansion of the application of a specific classification method on a specific fault to a broader spectrum of failure types. The proposed algorithms will have to incrementally learn from new information without forgetting previously acquired knowledge, and without requiring access to the original data. A new fault detection model must be capable of using new data that includes examples of previously unseen defects and be able to estimate the confidence in its own classification.
The goal of Asset Integrity Management (AIM) is to effectively manage corporate assets in order to gain maximum value, profitability and returns while safeguarding personnel, the community, and the environment. At Intertek our AIM team delivers trusted and innovative technical solutions that ensure the quality, safety and reliability of our clients' assets.
Dr. Gascón is a data scientist with expertise in evaluating large data sets from fossil and renewable energy power plants. He uses statistical analysis, mathematical modeling and machine learning techniques to generate predictive models to improve plant efficiency. Dr. Gascón earned a Ph.D. in Nuclear Physics developing detectors to study nuclear reactions. During a 2-year postdoc at Stanford and 3-year fellowship at Berkeley National Laboratory, he studied "How to improve the functionality of radiation detectors by applying extreme conditions and using novel techniques and materials," with the objective of increasing the sensitivity of these detectors for national defense applications. Dr. Gascón leverages his broad experience, programming skills and quantitative analytical background to solve a wide variety of problems involving large data sets in the power industry. He is the principal architect of Intertek's data analytics platform Ingrid. His current responsibilities include the development of COSTCOM®, a real-time cost analysis tool for power plant operations, WearMulti and OPCONTM, two multipurpose programs used in the nuclear industry. He also developed Windlife, a fully customizable web-application for wind farms featuring anomaly detection and remaining life estimation. Dr. Gascon has also developed programs and algorithms to predict the impact of the variability in power generation of renewable energy on the wear and tear of fossil fuel plants.