Artificial intelligence is already working for Slovenske elektrarne

Možnosti zdieľania:
Data scientists and engineers from Slovenske elektrarne managed to develop a predictive model under their own roof in six months, which they put into production after several months of testing. It works with Slovak language and a specific technical language used within the power plants.

Notification General (NG) is a type of reporting in the SAP information system that is used by the employees of Slovenske elektrarne to report faults, problems, suggestions for improvement, for evaluation activities with self-assessment and benchmarking tools, and to consider requirements for the issuance, revision or cancellation of a document. The registered NG reports are assigned a severity level and different trending codes every working day. This part of the process is referred to as screening and is handled by a separate group of staff. Now, the pre-screening committee will be assisted by artificial intelligence, which will suggest the coding as soon as the NG report is entered into the SAP system.

A similar approach is being applied at a US power plant, where an intelligent system of automatic code assignment based on recognition of text and technical locations from reports has been in operation for three years. Following the data science strategy at the Slovenske elektrarne, the idea of trying to apply such a method in our environment was born.

Data scientists and engineers took on the task and in agile collaboration developed a predictive model internally in half a year, which was tested in practice for another three months. The benefits of the project are consistency in the assessment of NG reports over time, between plants and people, automation and acceleration of operations, increased awareness and data accessibility for better alignment of internal processes.

What lies inside AI

The predictive model is created using NLP (Natural Language Processing) methods, in which the input data - the report name, description, technical location and its designation are parsed (separated) into individual words, endings are trimmed, grouped by word base and standardized into a computer-processable binary form. This produces more than 15,000 attributes (flags) from a single report, which enter the predictive model.

At the heart of the AI are 3 multi-layer neural networks with a total of over 70 million estimated parameters, resulting in a determination of whether a report has an impact on nuclear safety and "multilabel" coding for 52 causes and 13 types of impacts of a report.

Proces spracovania dát, tvorba modelu a následná interpretácia bola realizovaná pomocou služieb Azure Machine Learning Studio a Databricks.
The data processing, model building and subsequent interpretation was performed using Azure Machine Learning Studio and Databricks services.

How "intelligent" is this artificial intelligence?

The best NLP models in the world, used by e.g. Facebook or Amazon, achieve an accuracy of around 92% on standard spoken English. Our models achieve over 80% accuracy on Slovak and the specific technical language used in Slovenske elektrarne. Consultation with an independent firm has evaluated our models as above-average accurate, but we will continue to work on improving them. Consistent and accurate recording of NG reports and feedback from staff and the SNAP process will help us greatly in this regard.

"The quality of NG input significantly affects model accuracy, and the motto: garbage in-garbage out applies here as well. If we put low-quality inputs into the machine, we will get nicely transformed but still low-quality outputs out of the machine."

- Peter Kertys,
Data Scientist, Slovenske elektrarne

We work at the highest level

A few years ago, calculating such models was very time-consuming, but nowadays we use the latest Cloud technologies in power plants, such as Databricks, ML Studio, Azure Synapse and so on. Powerful computers with the latest graphics cards optimized for neural networks help us with the calculations, and we only pay for the computation time. For comparison, a similar model would take hours to compute in the past, today we have it in minutes, with computations running in parallel. In particular, for example, our model was created by many variations and combinations - hyperoptimization of more than 1,000 different models.

"Thanks to modern technology, we are able to process large amounts of data more efficiently and look for hidden connections in it."

- Maroš Bubán,
ICT BI Architect, Slovenske elektrarne
Návrh architektúry neurónovej siete: 42 481 717 odhadovaných parametrov.
Architecture design of one neural network with 42,481,717 estimated parameters.

What's next?

We have a number of other ideas in which the potential of machine learning could be used in the process of optimizing production resources, streamlining processes, better production planning, and so on. If you want to be part of our team, keep an eye on the offers posted on Nalgoo, Profesia or on our social networks.