Modern manufacturing enterprises are increasingly moving from experimental AI pilots to industrial-grade implementations. Local computer vision projects on production lines are proving their worth: cameras detect micro-cracks, scratches, and geometric deviations. However, scaling leads to the problem of siloed logs. Information about detected defects often remains trapped in edge devices (Edge AI) and fails to trigger automated response chains within the corporate environment.
The true value of visual AI control systems lies not just in detection, but in the reliable integration of this data into corporate ERP or DMS systems. The main challenge is transforming raw logs into legally binding business events while maintaining an end-to-end audit trail and adhering to industrial cybersecurity requirements.
The "last mile" problem in industrial AI: why precise detection fails without integration
There is a misconception that installing a modern AI camera on a conveyor automatically makes an enterprise Industry 4.0 compliant. In reality, the camera is merely a sensor. If the metadata stream is not integrated into business processes, the enterprise remains dependent on manual labor. An operator must manually transfer defect data from the video analytics interface to accounting systems to initiate write-offs or file claims against suppliers.
It is impossible to completely remove humans from the verification process for critical defects. AI models always have a percentage of false positives. Therefore, the architecture must be based on the Human-in-the-Loop concept, where the system automates evidence collection and routine tasks, while complex cases are passed to an operator for approval, whose decision forms the final document.
The architectural bridge: connecting industrial OPC UA and corporate business logic
The main obstacle to automation is the gap between operational technology (OT) and IT interfaces. Industrial equipment works with specific real-time protocols, while quality management systems (QMS) or electronic document management systems are oriented toward REST API.
To overcome this barrier, the OPC UA protocol is the de facto standard. As defined by the OPC Foundation, OPC UA acts as a platform-independent architecture for secure and reliable interaction in industrial systems. It serves as a data normalization layer before transmission to manufacturing execution systems (MES) or edge analytics. Signals from the AI camera pass through an OPC UA server and are converted into standardized industrial quality events, which the integration layer can directly translate into ERP logic. This allows, for example, a conveyor controller to instantly block the movement of a defective part to the next stage.
Ensuring data immutability: building an audit trail on the UnityBase platform
Transmitted defect data requires reliable storage. Information must not be modified or deleted manually, as it serves as the basis for financial and legal consequences.
Low-code platforms of enterprise level, such as UnityBase (a joint development by companies of the Intecracy Group alliance, where InBase is a key but not the sole developer), are often used as a reliable integration bridge for recording such business events. By using a unified domain metadata model, the platform allows for the rapid generation of REST API to receive quality events and ensure their routing. For critical environments, official documentation recommends Enterprise or Defence editions, which provide built-in mechanisms for maintaining an immutable audit trail, row-level security (RLS), and support for cryptographic signatures.
Securing the OT perimeter: implementing AI systems according to ISA/IEC 62443
Connecting edge AI devices to the corporate network creates new threat vectors for legacy equipment. Designing such systems must rely on the ISA/IEC 62443 standard, developed specifically for the security of industrial automation and control systems (IACS) and applied in over 20 industries.
Furthermore, according to NIST SP 800-82 guidelines, fundamental control measures include IT/OT network segmentation. Following best practices (such as those described in architectural frameworks like AWS Well-Architected IoT Lens regarding the division of data processing between the edge and the cloud), the edge computer vision device must be placed in an isolated demilitarized zone (DMZ). Data exchange is permitted exclusively through secure gateways with mandatory authentication to prevent unauthorized access from the IT network directly to programmable logic controllers (PLC).
Practical case study: the defect journey from AI detection to rejection act in DMS
Let's consider a practical integration process using the technology portfolio of the Intecracy Group alliance.
- Detection (Edge AI): The Ionbond AI Visual Inspection system (from Softengi, which is certified according to the AI management standard ISO/IEC 42001:2023) operates on the quality control line. The camera captures a surface defect on a part with a specific serial number.
- Normalization (IoT/OT layer): Raw metadata with coordinates is read by the AZIOT Platform industrial IoT platform via the MQTT protocol and normalized. Then, the event is transmitted to the integration layer via an OPC UA gateway.
- Business logic (Platform bridge): The UnityBase platform verifies the event and queries the ERP system, blocking the part for further assembly operations.
- Documentation (DMS): Simultaneously, a non-conformance report is automatically created in the document management system (e.g., Megapolis.DocNet or Scriptum.DMS), to which a photo of the defect is attached.
- Approval (HITL): A quality controller verifies the defect and signs the rejection act with a qualified electronic signature (e.g., via the DealsSign module). The business process is fully traceable from the sensor to the electronic archive.
Integration matrix of AI control: from raw data to business event
| Architecture layer | Data and event type | Technology stack / Tools | Business value and security |
|---|---|---|---|
| Data layer (Edge AI) | Raw detection metadata (coordinates, defect class) | Local device log, MQTT/HTTP protocol | No connection to business context, risk of data loss |
| Normalization layer (OPC UA) | Standardized industrial quality events | OPC UA server, UnityBase integration layer | Secure data transit between OT and IT segments |
| Business logic layer (DMS/ERP) | Legally binding document (rejection act, batch block) | Defect card in Megapolis.DocNet with qualified electronic signature | Full traceability (audit trail), business process automation |
FAQ
How to ensure compliance with ISA/IEC 62443 requirements when connecting AI cameras to the corporate network?
A key measure is strict network segmentation. In accordance with ISA/IEC 62443 and NIST SP 800-82 requirements, the edge AI device must be located in an isolated demilitarized zone (DMZ). Data exchange between the corporate network and the industrial circuit (SCADA/PLC) should only be carried out through secure gateways with mandatory authentication.
What is the role of the OPC UA standard in integrating computer vision systems with ERP systems?
OPC UA acts as a platform-independent architectural bridge. It provides a normalization layer, converting raw metadata from AI cameras into standardized industrial quality events that can be securely transmitted to the IT system integration layer (such as MES or ERP) to initiate subsequent business reactions.
How to guarantee that defect data in the document management system will not be deleted or modified manually?
This is solved by using enterprise-level platforms with built-in protection mechanisms. For example, solutions based on the UnityBase platform (specifically Enterprise/Defence editions) use an immutable audit trail, strict access control (RLS), and mandatory cryptographic signing of documents (qualified electronic signature), which guarantees data integrity.