- Статьи
- Society
- Factory settings: neural networks will start monitoring safety in production facilities
Factory settings: neural networks will start monitoring safety in production facilities
Russia has created an artificial intelligence-based program that identifies actions that can lead to industrial accidents. These include smoking in the workplace, talking on the phone or drinking alcohol. Today, more than a quarter of accidents at work are caused by employee negligence — and the development of scientists from ITMO University is designed to solve this problem. Read the Izvestia article for details on how modern technologies will make Russian production safer.
What is known about the development of ITMO specialists?
Scientists at the St. Petersburg National Research University of Information Technology (ITMO) have created a program that identifies actions that can lead to industrial accidents. These include smoking in the workplace, talking on the phone or drinking alcohol. This was reported to Izvestia by the ITMO press service.
According to the Social Fund of Russia, more than a quarter of accidents at work (27.8%) occur due to employee negligence. Today, video surveillance systems are used in production and in public places to monitor dangerous activities. At the same time, videos are often watched manually, and this is an imperfect method: it is difficult for a person to monitor several screens at once, he may get tired or distracted and miss an important event.
Neural networks help automate the process: they are able to continuously capture events and highlight what is needed in a long stream of video. But each such algorithm has its own limitations.
— For example, the models presented on the Russian market for industrial video surveillance are able to recognize only objects (people, the presence of helmets and masks), but are not able to track actions. Open foreign models do a better job of this, but they still don't capture actions accurately enough: algorithms trained on the ITMO dataset correctly recognized actions only in 24% of cases (VideoMAE) and 48% of cases (Hiera), the university said.
The new Russian ActionFormer algorithm recognizes up to ten actions with an accuracy of about 80%, which is several times higher than that of foreign competitors. Due to its compactness (3.7 million configured parameters versus 22-73 million for analogues), the Russian neural network operates with minimal computational costs.
What are the prospects for the new program?
The development consists of two components: one model identifies key points in people's images, and the second analyzes the actions and tracks the movements of employees. The system records violations: talking on the phone, moving equipment, or interrupting work at an important moment.
Data on violations, depending on the settings, is transmitted to the database or directly to the operator, said Valeria Efimova, project manager, Candidate of Technical Sciences, researcher at the ITMO Computer Technology Laboratory.
— This is how the program allows you to track any actions that may lead to an emergency at work. In addition, the development helps to prevent camera failure if the lens is intentionally dirty or covered to hide prohibited actions. In many industrial enterprises, these actions are considered a violation of safety regulations," Efimova explained.
According to her, the development has already been successfully applied at a large enterprise in the Perm Region. There, she tripled the number of physical inspections and helped prevent incidents, such as equipment repair errors due to a distracted employee. A team of scientists is currently working to expand the program's capabilities. There are plans to adapt neural networks for wearable cameras to use them in mines.
— With the help of neural networks, it will be possible to monitor that the team performs the necessary actions and adheres to safety rules: for example, it uses PPE, carries out work in accordance with the instructions, safely descends the stairs, — Valeria Efimova explained.
In parallel, a version for residential complexes is being developed, trained for 150,000 personnel. She also records violations: alcohol consumption on playgrounds, unauthorized unloading of vans or attempts to enter the entrance. In the future, the system will be able to detect acts of vandalism, such as damage to benches or lawns.
What is unique about the development of ITMO specialists?
As the Director of Business Development at Artezio (part of the LANIT group of companies) explains to Izvestia Denis Kharchenko, attempts to create systems to improve safety in production are being made both in Russia and abroad, but they have a difference in approaches.
Most of the existing solutions for industrial video surveillance on the Russian market operate at the level of object detection, but they are not able to analyze the sequence of actions.
"It's like the difference between photography and film: it's one thing to capture the presence of an object, quite another to understand what that object is doing in time," the expert notes.
If we talk about international experience, there are open models like SlowFast or X3D that can recognize actions, but their accuracy for industrial tasks remains insufficient — it usually ranges from 60-65%. In addition, these solutions require significant computing power, which makes their introduction into production economically impractical.
—— ITMO's development is interesting precisely because of the combination of high accuracy and relative lightness, which is critically important for scaling in large enterprises, where there may be hundreds of cameras. It is a compact solution that can be run even on a relatively simple enterprise server. At the same time, there is no need to rent expensive cloud capacities or buy equipment for millions of rubles," says Denis Kharchenko.
How do IT systems affect the safety of production facilities?
The impact of modern systems on production safety cannot be overestimated. Practice shows that the approach of inspections and manual video viewing cannot ensure continuous monitoring: the inspector is physically unable to be everywhere at the same time, and the operator loses concentration after 20-30 minutes of monitoring dozens of screens.
— Automated action recognition systems work 24/7 without loss of attention and are able to instantly respond to potentially dangerous situations. Imagine: the system sees that an employee entered a dangerous area without a helmet or started smoking near flammable materials, and immediately sends a signal to the dispatcher. This is no longer an after—the-fact debriefing after an incident, but the prevention of the incident itself," explains Denis Kharchenko.
According to him, the application of such algorithms is in full swing, and the results are impressive. In the metallurgical industry, computer vision (CC) systems are being actively implemented to control hot workshops, where the temperature does not allow a human controller to be constantly present. The system monitors the correct use of thermal protective suits, maintains a safe distance from molten metal, and monitors that workers are not under the trajectory of the melt buckets.
"At one of the plants in the Chelyabinsk region, after the introduction of such a system, the number of thermal injuries decreased by 40% in the first year," Kharchenko notes.
Stanislav Yezhov, Director of AI at Astra Group, calls the introduction of such systems an urgent process for most modern enterprises. AI—based systems are capable of detecting up to 350 critical events in one shift, such as the appearance of an intruder in a protected area, a violation of the regime of wearing overalls, an attempt to enter or a fall of a person. Such a burden is simply too much for a person.
What is the potential of AI in the field of security?
According to experts, AI not only improves safety at work, but fundamentally changes its very paradigm — companies are moving from a reactive approach to a proactive one. Previously, security was based on analyzing incidents that had already occurred and trying to prevent their recurrence. AI now allows you to predict dangerous situations before they occur. Modern algorithms analyze patterns of behavior and identify anomalies that the human eye is unable to detect in the flow of routine operations, explains Denis Kharchenko.
— Let's take a specific example: the system may notice that an employee has started moving more slowly than usual, stops more often, and his posture changes. For a human observer, these are imperceptible changes, but the algorithm already understands that an employee may be tired or unwell, and the risk of error increases dramatically. The system may recommend an additional break or a change of tasks before fatigue leads to injury and other serious consequences," the expert notes.
In his opinion, a real breakthrough can be expected in the security sector in the next 2-3 years. The systems will become multimodal — they will analyze not only video and images, but also sound, vibrations, and temperature maps.
— Imagine: the system hears the uncharacteristic sound of the machine and compares it with the micro-movements of the operator, realizes that the equipment is working hard, and warns of the need for maintenance before breakage. This is the next level of predictive security," says the expert.
There is also great potential for integration with wearable devices. Smart helmets and wristbands will be able to transmit data on an employee's pulse, body temperature, and stress level, and a computer vision system will compare this data with human behavior. If the pulse rate is through the roof and the person is working with dangerous equipment, the system may temporarily block the machine and suggest taking a break.
"We are moving towards the concept of a digital twin for each employee, where AI will know the individual characteristics and risks of a particular person," the expert concludes.
How is industrial video analytics developing in Russia?
According to Stanislav Yezhov, today about 80% of solutions in the field of industrial video analytics in Russia are domestic developments. This means that most of the systems that analyze video from cameras in factories, power plants, pipelines and other important facilities are created by Russian companies.
— The predominance of domestic equipment is not an accident. This became possible thanks to the support of import substitution, the development of our own technologies and strict safety requirements," says the expert.
According to him, AI-based systems are used in critical sectors of the economy. At nuclear power plants, they monitor compliance with safety regulations: for example, they automatically check whether employees are wearing all 26 elements of personal protective equipment (PPE) — from helmets and glasses to special shoes and gloves. This helps to avoid mistakes and violations that can lead to accidents. At metallurgical plants, AI analyzes images from cameras in order to identify defects on the metal surface in time: cracks, inclusions, and irregularities. Such systems work in real time and allow you to detect defects even before the product is released, saving time and resources.
According to experts, the volume of the Russian industrial video analytics market is growing rapidly and will reach 5.6 billion rubles by 2030. This suggests that companies are increasingly adopting smart technologies instead of manual controls. At the same time, Stanislav Yezhov notes, several important trends are already noticeable.
— Firstly, the systems are moving from a single general algorithm to multi-agent solutions, that is, they use several specialized AI models at once. Secondly, more and more calculations are transferred directly to the camera or the nearest equipment — this is called edge computing. Thirdly, as 5G networks develop, it becomes possible to transmit large amounts of video in real time," the expert concludes.
However, as scale increases, it is critically important that the implementation of AI is combined with compliance with all safety regulations.
Переведено сервисом «Яндекс Переводчик»
