From Video Monitoring to an AI Safety Ecosystem: How to Turn HSE Data into Incident Prevention

The deployment of digital tools in occupational safety is often limited to installing video cameras and basic incident tracking. However, the true potential of data is unlocked when moving from simple violation logging to a proactive ecosystem. In his speech, Alexander Bondarenko, Head of HSE at SSGPO (Iron Division, part of ERG), details the experience of transforming scattered digital solutions into a single AI ecosystem that enabled the enterprise to reach historical highs in operating without fatal injuries.

Evolution of Control: From Drones to Educational Content

Historically, the enterprise experienced an average of five fatalities per year. The turning point was the implementation of adapted global practices and large-scale digitalization. One of the first steps, implemented back in 2016, was the use of drones to monitor high-risk work. In practice, this tool proved its critical value: in 2021, a drone pilot spotted two workers in time resting in the shade beneath a ledge where work was underway. The noise of the propellers prompted them to leave the danger zone just 12 seconds before a rockfall.

Today, technical monitoring encompasses over 1,800 stationary cameras and around 500 personal video recorders. A massive volume of video data does not sit as "dead weight" on servers; instead, it is converted into educational material. Every month, the HSE team edits short video clips demonstrating both mistakes and examples of safe task execution. This content is broadcast on 200 TV panels in pre-shift briefing rooms before every shift, establishing a continuous micro-learning process.

Artificial Intelligence in Behavior Management

The speaker pays special attention to evaluating communication between supervisors and workers. The enterprise has deployed an automated system to analyze safety toolbox talks and behavioral safety dialogues. Supervisors record their briefings on a voice recorder, after which algorithms assess the quality of the conversation against an approved methodology.

Thanks to automation tools, the supervisor receives structured feedback in just 45 seconds: the system highlights the dialogue's strengths and areas for improvement. Data is synchronized with dashboards, creating a unified cross-sectional view across departments.

To prepare young supervisors for safety audits, a specialized electronic coaching simulator was developed. In this system, artificial intelligence plays various worker profiles—ranging from loyal to resistant or fatalistic. The supervisor's task is to structure the dialogue to lead the virtual interlocutor to recognize the risks.

Comprehensive Driver Fatigue Assessment

Worker fatigue remains one of the hidden threats in production that cannot be detected by standard breathalyzers. To address this issue, the team developed a hardware-software module (PAM 60).

The complex assesses a person's reaction time using a gaming steering wheel and pedals, delivering signals over the course of one minute. Simultaneously, a 4K camera equipped with an LED ring performs pupillometry—measuring the speed of pupil constriction and recovery under light stimulation, as well as detecting eye redness. The system's key differentiator is its rejection of averaged standards. The module accumulates statistics for each employee and compares current indicators against their personal baseline. If the reaction speed and pupil parameters deviate from the norm, the employee is not penalized; instead, they are reassigned to auxiliary tasks in the garage, eliminating the risk of a road traffic accident on the line.

Local AI Tools and Data Protection

Deploying generative models often entails risks of confidential data leakage. To avoid this, the enterprise focused on standalone offline browser-based applications and local language models. This allows complex analytics to be automated securely:

• Incident Investigation: The system analyzes audio interviews and case files, automatically builds a cause tree, constructs a bowtie diagram, and prepares presentations.
• Knowledge Verification: A multi-agent AI conducts oral examinations on regulatory requirements, asking questions by voice and evaluating the essence of employee answers in percentages.
• SOP Development: Based on completed process cards, agents generate full-fledged standard operating procedures in strict compliance with methodological requirements.

What You Will Learn from This Speech:

• How to build a pipeline for creating educational video content based on body cam footage?
• How to automate the evaluation of behavioral safety dialogue quality using neural networks?
• How to use pupillometry and reaction tests for personalized driver fatigue monitoring?
• How to apply AI simulators to train supervisors in complex communications with personnel?
• How to deploy autonomous AI tools for incident investigation without data leakage risks?