AI-Based System for Automated Digitalization of Relay Protection Settings Forms

Operator of the Unified Power System is introducing technologies for automated processing of non-digital settings for RZA devices into a digital format for their subsequent use in the business process of issuing settings to electric power entities and use in other automated systems

(AS) for data storage and calculation complexes of the relay protection and automation (RZA) service. These systems require as input the actual values of RZA settings, which until recently were stored in the form of analog (paper and scanned) weakly structured documents. Manually converting such forms into digital form takes a significant amount of time, which depends on the availability of templates, the number of settings, and the degree of formalization of the form.

Existing systems using machine learning for document recognition are not suitable for this task due to the specifics of the data contained in the RZA forms.

The purposes of the article is to describe the entire process of creating a system for automated recognition of attributes and values of settings from non-digital forms using machine learning methods, which was named "MaLena".

Data Science and machine learning methods were applied to automate the digitization process.

During system development, requirements for architecture, operating mode, categorization, scaling, software and hardware, interaction with infrastructure systems, high performance, reliability, fault tolerance, and information security were applied. The article will discuss the architectural solutions and key problems encountered when developing the solution. The main results are a fine-tuned model adapted to input data in the form of RZA setting forms, a status model defining the need for manual control of the system by a person, and a business process for interaction between employees of branches using the "MaLena" system, which made it possible to reduce the time for completing the task of digitizing RZA setting forms severalfold.