Consistency Check of EMAT-C Data Evaluation Supported by Deep Learning
Complex pattern recognition using artificial intelligence
Highly trained and experienced data analysts conduct the evaluation of the different complex data sets gathered during EMAT-C inspections. Quality Assessment and Quality Control (QA/QC) is supported with use of machine learning by the novel “consistency check”.
The challenge
The evaluation of EMAT in-line inspection data is a challenging task that requires experienced data analysts and extended training. Multiple information channels need to be combined and analyzed simultaneously. A significant part of the evaluation process is the analysis and interpretation of data patterns, which results in detection, classification and sizing of pipeline features. Therefore, the interpretation of EMAT inspection data can be viewed as a complex pattern recognition problem statement. It is of utmost importance that the overall analysis process conducted by the data analyst achieves the highest possible levels of quality and repeatability.
Our solution
In recent years, supervised machine learning, and especially deep learning has emerged as the dominating technique to tackle such pattern recognition tasks computationally. During the analysis of the data of an EMAT inspection, it is common that tens of thousands of signal indications have to be analyzed. Due to the challenging nature of the analysis process, some of these indications can be misclassified. It is self-evident, that the risk of such errors increases with the overall feature count.
Furthermore, an analysis of an inspection by more than one data analyst could result in an inconsistent evaluation. These risks are well-known and are commonly addressed by quality checks (QA/QC). However, which and how many anomalies shall be re-checked is often not easily definable and transparent.
To address this challenge, ROSEN developed a unique Consistency Check Tool. The Consistency Check is a deep learning based process, trained on high-quality labelled data. It is able to identify anomalies, which are close to a decision boundary, e.g. between crack or non-crack. After the process is finished, each feature has a crack score, i.e. a score how crack-like the deep learning model predicts a feature to be, based on the other features it has seen during training. Then, to minimize the risk for false positive calls, all features, which were labelled as cracks but received a low crack score are reviewed. Similarly, all features, which received high crack scores, but were not labelled as cracks are re-evaluated to avoid the risk of false negative calls.
Your benefit
The implementation of the deep-learning based consistency checker addresses the risk of miss-classification during the conduction of the EMAT-C data evaluation. It is not replacing the data analyst, but supporting during QA/QC to achieve the highest possible level of consistency and repeatability of the reported results. Application of the consistency check supports to avoid both, false positives and false negatives. Thus, a higher quality and consistency of the evaluation result is achieved avoiding cost due to unnecessary digs and further reduce the risk of missed anomalies with improved POX.