Center: AI-LiquidOil

Main article:

• Video analytics (terms, applications, technologies)
• Computer vision (machine vision)
• Pattern recognition

2022: AI-LiquidOil Announcement

On June 27, 2022, the Center company announced the development of the AI-LiquidOil solution, which automatically processes video images to detect a spill on the earth's surface and leak of liquid in production using artificial intelligence.

According to the company, the original video/photo data can be obtained from satellites, unmanned aerial vehicles, stationary cameras. Accuracy is ensured by the use of artificial intelligence technologies and can reach 98%. The software product is cross-platform, can be installed on a physical and virtual computer has the ability to work with various sources, data optimization of calculations can be provided by using graphics cards with CUDA support. To determine the exact signs of spills, different data are used:

• video images of industrial facilities;
• synthesized data obtained from processing the provided data by changing viewing angle, rotation, cropping, changing image proportions;
• data obtained by rendering an image of the created 3d model.

This approach helps to expand the input datasets to model the exact signs of spills. For different situations, "AI-LiquidOil" has three modes of use with different confidence levels:

• The basic mode is responsible for detecting possible liquid spills in the visible range with capturing the video image of the location, the time of detection and the data source;
• IR mode in addition to the basic mode, the analysis uses data from cameras with infrared illumination, which makes it possible to optimize the probability of recognizing the spill of liquids from cameras, for example, on a well or separation unit;
• The thermal mode in addition to the basic mode, the analysis uses data from a thermal imaging camera, which will optimize the probability of recognizing spills of oil-containing liquids and distinguish it from a puddle, wet asphalt.

To periodically check the correctness of the process described by the algorithm, the training data is interleaved. Neural network training was carried out on the basis of 200 video data files. The trained detection accuracy was 90%. The predicted accuracy of training on 10 thousand files is estimated at 98%.

System Operation Result

To identify more accurate results, it was necessary to enable the system to accurately display the main signs of a spill. To do this, the volume (up to 2 thousand) and the quality of the training material were increased due to the application of transformations to the source data, the standardization of the training material was applied.

System Operation Result

During work, most of the time is spent training a large stream of data. According to the asynchronousness of multithreaded computing, it is possible to optimize video processing or process several in parallel. If you are working with multiple instances within a single thread, only a sequential processing type is selected. This limitation is imposed by the architecture for working on a multithreaded system. Preference is given to parallelizing calculations for high processing speed. This approach reduces the processing time by Nx0.7 times, where N is the number of streams. This dependence is explained by the need to perform procedures that are not related to the calculations of the neural network.

Improve algorithms the normalization of heterogeneous (different format) input information to optimize the processing speed and quality of spill detection.

To optimize recognition accuracy:

• A rating system is being developed in which a specialized specialist will evaluate the results of the algorithm and make adjustments to train the system, watch controversial, implicit "spills" and make an additional assessment.
• The initial data collection database is increased due to satellite data and the amount of data from the Customers.
• An algorithm competition system is optimized to identify the most effective recognition algorithm method for different types of data.

To create artificial training data, digital 3D models of spills are created.

The system can be integrated with the enterprise IS, as well as operate locally and independently. The local-independent system uses a client application with a simple intuitive graphical interface. If the system is deployed on the server side, graphical and console interfaces can be used.

Client Application GUI

The main emphasis in the AI-LiquidOil solution is aimed at integration with the Customer's existing IC.