Three electrical-equipment plants belonging to Corporation X run 24/7 with more than 2,800 shift workers. Despite years of investment in safety programs, the plants still recorded an average of 17 lost-time injuries (LTI) per year and more than a thousand near-miss incidents—most of them reported verbally and logged only sketchily. OSHA 300 and ISO 45001 reports were compiled manually and submitted 30–45 days late, leaving top management without timely data for preventive action.
Faced with this situation, the CEO set three clear goals:
Cut LTIs by at least 25 percent in the first year.
Standardize OSHA / ISO 45001 compliance, reducing report latency to fewer than seven days.
Achieve payback within 12 months on every safety-improvement investment.
Current-State Business Analysis
A deep dive into shop-floor operations showed that safety management was heavily “rear-view-mirror” driven. After an incident, a worker filled out a form, a supervisor wrote it in a notebook, and only at the end of the shift was the data keyed into Excel; by the time managers saw the numbers, it was far too late to intervene.
High-risk zones—fork-lift aisles, high-bay pallet racks and welding stations—used analogue CCTV for passive monitoring only, while PLC / SCADA sensor data stayed walled off from the safety platform. The organisation therefore lacked leading indicators such as PPE-violation rates, human-machine density or fork-lift speed, forcing the EHS team to “guess” trends and allocate budgets by gut feel.
Financial consequences were obvious: compensation, line downtime and post-incident clean-ups cost more than ₫ 3.4 billion per year, yet root causes could rarely be traced and eliminated.
Factory accident risk recognition system
(The technical architecture diagram and the five-step operational process remain the same as in the previous summary. They are reproduced below for completeness.)
1. Technical Architecture
| Layer | Component | Core Function |
|---|---|---|
| Edge | • 4 MP IP cameras + NVIDIA Jetson Xavier • 270° LiDAR for fork-lift exclusion zones • UWB tags fixed to safety vests | Detect PPE, count personnel, flag intrusions |
| Gateway | Industrial PC with Docker | Multiplatform ingestion (RTSP, Modbus-TCP, OPC-UA) into unified JSON |
| Time-Series Bus | Apache Kafka | Buffer 50,000 events / s, ensure exactly-once delivery |
| Streaming Analytics | Spark Structured Streaming + MLflow | • YOLOv8-PPE, DETR-Person-Distance models • Anomaly detection (Isolation Forest) |
| Data Lake | Azure Blob (Parquet) | Store three-year video snippets + telemetry for audit |
| Dashboards & Alerts | Grafana + Firebase Cloud Messaging | Hazard heat maps, leading KPIs, push alerts under 30 s |
2. New Operational Workflow
| Step | Actor | System Support | Outcome |
|---|---|---|---|
| 1. Data Capture | Cameras, LiDAR, UWB | Edge inference 20 fps | Events “PPE violation”, “zone breach” |
| 2. Risk Scoring | Spark + ML | 5×5 Risk Matrix | R ≥ 15 → trigger alert |
| 3. On-site Warning | IoT beacons + SMS | MQTT /safety/alert | Immediate worker reaction |
| 4. Safety Dispatch | Android / iOS app | Auto-ticket in Jira Service Desk | Response SLA < 5 min |
| 5. Analysis & Improvement | HSE Manager | Power BI leading-indicator board | Weekly EHS root-cause review |
Deployment & Change Management
To minimise rollout risk, the project was delivered in four well-defined stages. The Proof of Concept ran six weeks on a small assembly line, validating PPE-AI accuracy (≥ 92 percent) and alert latency (< 1.5 seconds). Once KPIs hit their mark, the team launched a Pilot across Workshop A, training 120 workers, integrating Jira Service Desk for automatic ticket creation, and wiring SCADA data so that machine-press speed appeared in real time.
With the Pilot’s success, management green-lit a plant-wide Roll-out: 164 IP cameras, 14 LiDAR scanners and 1,900 UWB tags were deployed across all three factories over the next eight months. Each week a Change Champion group—safety officers, line leaders and HR reps—captured operator feedback, updated quick-start guides and fine-tuned alert thresholds to avoid “notification fatigue.”
Once the entire system stabilised, the Optimisation phase began: AI models were A/B-tested on MLflow, and risk-score thresholds were seasonally tuned (e.g., higher welding-area sensitivity during rainy months when floors are slippery). Thanks to disciplined change management, worker acceptance climbed quickly; operators came to view alerts as a “safety assistant,” not a surveillance tool.
Quantitative Results After 12 Months — Narrative Form
Within a year of full deployment, the system delivered striking gains. Lost-time injuries fell from 17 to 11, a 35 percent improvement.
Automatic logging multiplied the number of recorded near-miss events from 1,250 to 4,600—evidence not of a riskier workplace, but of the platform’s ability to surface hazards that had previously gone unseen. Personal-protective-equipment (PPE) violations dropped by 60 percent, from 0.43 to 0.17 incidents per labor-minute.
Average response time to a safety alert compressed from roughly 25 minutes to just under four, an 85 percent reduction. Altogether, lower injury rates and shorter production stoppages cut compensation and downtime costs from ₫3.4 billion to ₫1.9 billion per year, saving around ₫1.5 billion annually. These savings pushed the project to break even in 14 months—earlier than the CEO’s target.
Key Lessons and Recommendations — Narrative Form
The rollout surfaced four decisive lessons. First, prioritizing leading indicators—such as PPE-violation frequency and human-machine density—delivers far more preventive leverage than merely tallying accidents after the fact; when predictive metrics refresh hourly, the EHS team can intervene before risks escalate.
Second, combining edge AI with streaming analytics keeps alert latency below five seconds, avoiding cloud bottlenecks and ensuring on-site warnings remain actionable. Third, smart change management is critical: clear reward–and-penalty KPIs and a rotating “Safety Champion” on every shift reframed the technology as a helpful assistant rather than a surveillance tool, accelerating workforce adoption.
Finally, tight OT/IT integration—linking SCADA data, Jira Service Desk, and the HR system—eliminated manual re-entry and produced a robust audit trail for ISO 45001. In short, the project proved that technology yields real value only when embedded deeply in safety workflows, turning real-time data into action within seconds; preventive measures will always cost less than post-accident remedies.