Factory floors are increasingly running artificial intelligence (AI) inference locally on cameras and edge devices rather than sending all visual data to cloud servers. Electronics manufacturer USI has launched an AI Smart Camera designed to process computer vision workloads directly on the factory floor, allowing real-time image analysis without relying on centralized computing infrastructure.

What Problems Does On-Device Factory AI Actually Solve?

Manufacturing environments face a unique challenge: high-resolution cameras and AI-powered inspection systems generate massive amounts of visual data that can overwhelm production networks. A joint reference architecture from Cisco and Cognex found that many factory networks were originally built for deterministic control traffic and basic automation, not high-bandwidth image streams. Running AI inference close to the image source, rather than sending raw footage to distant servers, reduces latency and limits bandwidth consumption.

Beyond network efficiency, edge inference protects sensitive information. Industrial vision systems can capture proprietary products, manufacturing processes, or customer information. By processing images locally and selectively extracting only relevant data, factories can keep sensitive visual information under local control rather than transmitting raw images across networks.

USI's Smart Camera supports a range of manufacturing inspection tasks, including object detection, defect identification, optical character recognition, component positioning and tracking, assembly verification, and product classification. The system can also detect and analyze operational behavior patterns on the factory floor.

How Does Edge AI Deployment Change Security and Operations?

Edge AI deployment fundamentally shifts how organizations think about security. Instead of protecting a single, well-governed cloud boundary, companies must now secure distributed devices that may be intermittently connected, physically accessible, and harder to monitor. This means the deployment pattern itself becomes part of the security boundary.

The security implications are significant. If a camera, sensor node, or industrial device running local inference is compromised, attackers could affect both the AI workload and the surrounding operational system. Edge designs therefore require stronger assumptions about boot integrity, remote attestation, patching, and fallback behavior than centralized AI services.

Edge deployments also create operational trade-offs. Lower latency and better resilience to network loss are real benefits, but they come with more fragmented fleet management, more software versions in circulation, and more places where models, firmware, or configuration can drift out of sync. Consistency and observability become as important as raw performance.

Steps to Secure and Manage Edge AI Deployments

• Define Trust Boundaries: Clearly determine what must remain local, what can be synchronized with central systems, and what must be blocked from direct exposure. Distinguish whether edge nodes only consume centrally managed models or also store sensitive prompts, telemetry, or inference outputs requiring separate protection.
• Implement Device Identity and Signed Artifacts: Establish clear trust anchors for device identity, use signed software and model artifacts, and deploy secure update channels. This ensures that edge devices can verify they are running authorized versions of software and models.
• Monitor Fleet Consistency: Watch for unmanaged device sprawl, inconsistent model versions, over-permissive local services, and weak telemetry from remote sites. The safest edge deployments are designed for fleet consistency and verifiable updates before they are optimized for raw inference speed.
• Plan for Graceful Degradation: Design systems to function when connectivity is degraded or lost. Edge devices should have fallback behavior and local decision-making capability so that network interruptions do not cascade into operational failures.

USI's Smart Camera hardware includes a high-resolution camera module designed to operate in low-light conditions, paired with an edge computing system. Connectivity options include Ethernet, Power over Ethernet, HDMI, and MicroSD storage. The design is ruggedized for integration with industrial equipment, machine vision systems, and automation platforms.

The company has paired the hardware with embedded software and tools covering data collection, dataset generation, model training, and deployment. USI emphasizes a no-code and low-code AI model training platform intended to reduce dependence on specialized AI expertise when developing and deploying vision applications.

"Today's manufacturers are seeking more than just camera hardware," said Justin Chang, director of the Vertical Mobility Solution Center at USI.

Justin Chang, Director of the Vertical Mobility Solution Center at USI

USI has already deployed AI-assisted visual inspection within its own manufacturing operations. An automated optical inspection and AI defect-detection system introduced at its Jinqiao facility in 2023 and Zhangjiang facility in 2024 uses deep-learning models to identify more than 85% of defect types and improved inspection efficiency by more than 60% compared with processes relying on human visual rechecks.

Edge inference does not mean all computing moves onto the factory floor. Current AWS manufacturing guidance describes machine-learning models running at the edge for real-time defect detection while factory data can still be sent to cloud services for longer-term analysis, model retraining, and historical reporting. Computing resources for model training can reside in the cloud or separate on-premises infrastructure.

The shift toward edge AI in manufacturing reflects a broader recognition that not all AI workloads benefit from centralized processing. By combining local inference with selective cloud synchronization, factories can achieve real-time decision-making, protect proprietary information, reduce network strain, and maintain operational resilience even when connectivity is unreliable.