AI Safety & Industrial Safety OS

Overview

Artificial intelligence is increasingly moving from advisory analytics into systems that perceive the physical world, estimate risk, recommend actions and interact with industrial equipment.

This creates a fundamental engineering question:

How can AI improve industrial safety without becoming an uncontrolled source of safety risk itself?

My research in AI Safety and Industrial Safety OS focuses on system-level architectures that combine conventional functional safety principles with AI perception, spatial intelligence, uncertainty estimation, risk reasoning, deterministic supervision and evidence-driven lifecycle engineering.

Research Philosophy

The central principle is:

A safety-oriented AI system must know when its information is insufficient to support a safe decision.

AI should therefore not only output:

It should also provide information about:

When confidence is insufficient, the system should transition toward a safer operating state rather than continue making increasingly uncertain decisions.

Industrial Safety OS

Safety OS

Safety OS is an industrial spatial safety architecture for people, vehicles, robots, machines and materials.

Its architecture can be represented as:

Physical World -> 3D / Sensor / AGV / Robot / PLC -> Device Adapter Layer -> Unified Industrial Data Model -> Spatial Modeling -> AI Perception & Fusion -> AI Risk Engine -> Safety Supervisor -> Command / PLC / Robot / AGV -> Safe Action -> Event Trace / Recorder / Safety Case

Core Architecture

Safety OS is organized around six primary layers.

1. Safety Perception

Inputs include:

The perception layer produces a unified representation of the physical environment.

2. Spatial Intelligence

The spatial layer maintains:

This transforms isolated sensor measurements into an explicit representation of industrial space.

3. AI Risk Engine

The AI Risk Engine estimates operational risk using information such as:

AI is used as an enhancement mechanism rather than as the sole functional-safety mechanism.

Safety Supervisor

The Safety Supervisor provides deterministic orchestration between AI risk estimation and physical control.

Typical states include:

Normal -> Warning -> Slowdown -> Pause -> E-Stop Request

The complete runtime also includes:

Degraded
Recovery Check
Fault

The core design rules include:

  1. safer actions have higher priority;
  2. loss of critical information triggers degradation;
  3. communication timeout must be explicitly handled;
  4. failed commands must produce deterministic fallback behavior;
  5. recovery requires verification rather than automatic reset.

Safety Quality Frame

A safety-oriented spatial sensor should provide more than depth data.

A Safety Quality Frame may include:

Depth
Depth Confidence
Data Validity
Occlusion Ratio
Multipath Interference
Ambient-Light Interference
Time Synchronization Status
Calibration Status
Data Integrity
Safety Quality Level
Recommended Degraded Action

The key idea is:

The sensing system should report not only what it sees, but whether the current information is reliable enough to support a safety decision.

Functional Safety + AI

Safety OS follows a layered safety strategy:

L0  Inherent Safety -> L1  Functional Safety -> L2  Process / Machine Control -> L3  AI-enhanced Safety -> L4  Governance & Evidence

AI should enhance — not silently replace — the deterministic safety mechanisms required by the system risk assessment.

Relevant engineering frameworks include:

Safety Case & V&V

A core Safety OS lifecycle is:

Hazard -> Safety Goal -> Safety Requirement -> Design Control -> Test Case -> Evidence -> Residual Risk

Important engineering artifacts include:

Example Industrial Applications

Current application directions include:

Battery Pack Production Lines

Industrial Machinery

Safety OS concepts can also be applied to:

Research Roadmap

Industrial Safety -> Spatial Safety -> AI Safety -> Safety OS -> Safety Agents -> Safety Foundation Models ->Autonomous Safety Engineering

The long-term goal is a safety intelligence infrastructure capable of supporting risk discovery, system design, verification, evidence generation and runtime safety operation across industrial systems.