Machine Learning for NDT in the Finnish Nuclear Industry

Machine learning (ML) has made marked advancements in the recent years. In many industries, machine learning algorithms have reached near-human or superhuman performance in tasks that were long considered impossible to automate. For NDT, the opportunity to improve automation in the tedious and error-prone task of complex data analysis offers several potential advantages. Automated systems would provide significant cost and time savings while at the same time improving the reliability and consistency of inspections.

In this study, the current state of the art of machine learning systems for NDT and, in particular, inspection systems that could be applied to in-service inspections in nuclear power plants, is reviewed based on open literature. The current views and expectations of various stakeholders, including the utilities, NDT vendors, qualification bodies and authorities are surveyed. The necessary steps for bringing a machine learning system to industry use including needs for qualification and vendor participation are evaluated.

The recent literature (including recent work from the authors) show that the ML systems have recently reached a state, where they can show super-human performance in flaw detection tasks, that have traditionally been considered impossible to automate. The key enabling technology is sophisticated data augmentation, that enables generation of sufficient training data-sets from limited flawed data-sets and training mock-ups. The current view among qualification bodies met is, that the ENIQ framework is flexible enough to accommodate qualification of an inspection system including a ML component. The NDT vendors are, in general, interested to make use of ML tools, albeit their development is considered outside their area of expertise.

In summary, the current ML systems are powerful enough to be applied in the demanding environment of nuclear in-service inspection. To enable their use, further work is needed in training these models to wider range of data-sets including realistic plant data. Furthermore development and qualification of ML-powered inspection procedures need to be done in collaboration with NDT vendors.