[Manufacturing AI ②]AI Detects Defects and Operates Equipment in Push Toward Autonomous Factories

Mid-sized manufacturing systems integrators are actively developing and deploying artificial intelligence (AI) models tailored to manufacturing. Their approaches differ, however, reflecting the characteristics of their respective parent groups.

​POSCO DX focuses on equipment operation and control, CJ OliveNetworks on autonomous operation, Kolon Benit on automatically executing optimal production conditions, and Asiana IDT on real-time safety monitoring. The following outlines the manufacturing-specific AI strategies and implementations of the four companies, in alphabetical order by company name.

​◇CJ OliveNetworks Brings Autonomous Operation to K-Food Manufacturing

​CJ OliveNetworks is building its capabilities by combining general-purpose large language models (LLMs) with data and domain knowledge accumulated at manufacturing sites. As general-purpose models improve rapidly, the company believes competitiveness depends less on the models themselves and more on how precisely on-site data, work tools, and verification procedures are connected.

​CJ OliveNetworks stated that it relies on three core pillars — AI agents, predictive and optimal-control AI, and an AI vision inspection platform — which it combines and applies according to business needs. Its AI agents combine LLMs with manufacturing data structures, on-site terminology, approved work tools, and verification procedures so they can be applied directly to field operations.

Predictive and optimal-control AI is a process-specific model that accumulates and analyzes years of equipment and process operating data in a real-time database (RTDB), shifting operations that once relied on workers' experience toward AI-driven autonomous operation. The AI vision inspection platform is an AIOps-based system that allows field users to create vision inspection models themselves and manage retraining and deployment within a single operating framework.

​With K-food gaining ground in global markets, CJ OliveNetworks sees significant growth opportunities in the K-food industry. In the past, products were made largely on the basis of skilled workers' experience, often described as a manufacturer's "hand taste." Today, however, the company believes that collecting diverse data from production sites and converting it into digital assets can enable manufacturers to develop higher-quality food and beverage products while maintaining consistent quality.

​To that end, the company plans to build AI agents for key processes and integrate them to provide differentiated services specialized for manufacturing AI transformation (AX) in the K-food industry. Although it did not disclose specific external customers, CJ OliveNetworks noted that it is continuously expanding the scope of autonomous operation based on predictive and optimal-control AI, drawing on the results of various projects it has secured.

​◇Asiana IDT Bets on Seamless Integration of Legacy Systems and AI

​Asiana IDT plans to expand AI applications by combining its domain expertise in aviation, airports, logistics, manufacturing, finance, and the public sector with its systems integration (SI) capabilities. In manufacturing and logistics, it is applying internally developed solutions — the agentic AI platform ModelWave, AI optical character recognition (OCR) solution Ondocs, AI vision assistant A-VLAN, AI model performance management solution ModelOps AI, and AI coding assistant EasyVibe — to internal projects and customers' AX.

​ModelWave, an agentic AI platform, automates complex workflows through multi-agent collaboration. Asiana IDT explained that it supports conversational knowledge exploration, rapid information searches, and automated analysis of documents in formats including PDF and Hangul, while allowing users to design and edit AI workflows themselves. It also emphasized that flexible engine replacement and robust access controls and monitoring enable reliable services optimized for field operations.

​Ondocs is an AI OCR solution that automatically extracts key data from documents. It supports full-text and table-structure extraction, layout-based structural recognition, and template-area extraction, enabling processing optimized for different document types and work environments. In logistics, it automatically extracts key data such as numbers, dates, and amounts from waybills and invoices, reducing manual entry. In manufacturing, it recognizes data in purchase-order and receiving documents, reducing the amount of information that must be entered into enterprise resource planning (ERP) systems. The solution also contributes to training manufacturing-specific AI models with accurate data.

​The A-VLAN vision AI solution analyzes closed-circuit television (CCTV) footage in real time to detect dangerous situations and provide response procedures. Asiana IDT stated that A-VLAN can transform on-site safety management and operational efficiency by integrating object identification, movement and zone monitoring, and abnormal-event alerts at factories and logistics centers.

Performance management for AI deployed in the field is handled by ModelOps AI. Asiana IDT describes it as an MLOps/model operations solution that detects declines in AI performance and automates the entire process through retraining and deployment. The company noted that it can proactively prevent data changes over time, AI errors, and performance drift. Continuous monitoring and an automated retraining pipeline help maintain the accuracy and stability of AI services operating in the field.

​EasyVibe is optimized for conversational AI coding and data visualization in the company's closed network environment. It is used to increase developer productivity in source-code creation, quality checks, and data processing while allowing validated AI models to be utilized flexibly without security concerns.

​Asiana IDT positions its ability to seamlessly combine legacy systems and AI, backed by extensive domain knowledge, as a key competitive advantage. The company highlights its years of domain expertise in building and operating enterprise resource planning, manufacturing execution system, and warehouse management system (ERP, WMS, and MES) platforms in aviation, logistics, and manufacturing; its ability to accurately identify on-site pain points and refine unstructured data into high-quality information tailored to operations; and its end-to-end AI integration capabilities connecting existing legacy systems with AI solutions and establishing execution frameworks.

​Although it did not identify the customer, Asiana IDT stated that it applied EasyVibe to a manufacturing customer's production management (MES/POP) system project this year. The company reported that the use of EasyVibe in building the customer's production management and shop-floor control systems improved project development quality and productivity.

​For another manufacturing customer, Asiana IDT conducted a proof of concept (PoC) and deployment of Ondocs. It tested the performance of on-premises Ondocs on invoices and quotation documents for production items before moving to implementation. The process was shifted from manually entering amounts and items to AI automation, significantly reducing working time. Because it was deployed on premises, sensitive data such as invoices could also be kept secure.

​Asiana IDT's AI solutions are also being applied at parent company Asiana Airlines to analyze weather and Notice to Airmen (NOTAM) data and manage ground-handling safety based on CCTV footage. The company stated that the accuracy of flight-safety-related data analysis has improved significantly, while it has conducted a PoC for AI-based turnaround management in a ground-handling safety project. This technology supports AI-based analysis of ground-handling footage as well as verification of safety, compliance with procedures, and operational efficiency at ground-handling sites.

​◇Kolon Benit Offers Solutions Optimized for Customers Using Technology From More Than 100 Members

​Kolon Benit is pursuing a strategy of directly training vision and multimodal AI models and combining validated Korean large language models (LLMs) with retrieval-augmented generation (RAG) to implement AI specialized for manufacturing operations, including production, quality, equipment, energy, and safety. Its solution portfolio includes purpose-specific models, AI Vision Intelligence, r-CoCoAna, and PromptON Pak.

​The company independently trains and provides purpose-specific models for virtual sensing, vision AI quality inspection, deriving and automatically executing optimal production conditions through Golden Recipe and recipe-driven execution (RDE), and predictive maintenance. It is also proposing AI Vision Intelligence as a Korean sovereign AI package. This system combines EXAONE from LG AI Research with ATOM, an AI semiconductor neural processing unit (NPU) from Rebellions. The structure trains models on field data using graphics processing units (GPUs) and then deploys them on NPUs, a setup demonstrated by a five-company consortium comprising Kolon Benit, LG AI Research, Kolon Global, Rebellions, and Wish.

The platform supporting these models is r-CoCoAna, Kolon Benit's proprietary manufacturing platform. It combines a data platform for connecting heterogeneous equipment data, known as a connected data platform (CDP), with manufacturing operations management (MOM). It connects manufacturing execution systems (MES), production planning (SPIC), equipment and data management through BENIT Historian, equipment management through KAMS, and energy management systems (EMS).

​Kolon Benit also combines the technologies and infrastructure of members of its AI Alliance to provide pre-validated specialized AI solutions through PromptON Pak for production, quality, equipment predictive maintenance, energy, and safety, including the prevention of serious accidents. Solutions include Nota's VLM-based video monitoring Pak.Vision, Neurocle's automated deep-learning quality inspection Pak.VI, Infinic's de-identification solution Pak.Private, and RaonPeople's Hi FENN. The models are trained and operated in customers' on-premises environments using their process, quality, and video data.

​Kolon Industries' Gimcheon Plant 2 applied Kolon Benit's virtual sensing, vision AI, Golden Recipe, and RDE to its phenolic resin production line, converting six existing inspection stages to AI-based automation. Additionally, it completed a PoC using Nota's NVA for video monitoring to verify workers' use of protective equipment, abnormal behavior, and compliance with standard operating procedures (SOPs).

​At Kolon Industries' Gumi plant aramid production line, the company collected, connected, and visualized process data to shorten inspection times and improve quality, productivity, and profitability. At Kolon Global construction sites, AI Vision Intelligence is utilized to detect workers without safety helmets, entry into hazardous areas, the absence of signalers, and the approach of heavy equipment in real time. Cognitive Autonomous Manufacturing technology has also been gradually applied over the past three years across manufacturing affiliates, including Kolon Industries, Kolon Life Science, and Kolon Pharmaceutical.

​Outside its captive market of group affiliates, Kolon Benit cited SeAH Special Steel as a customer. At SeAH Special Steel's Wonju plant, the company established equipment and energy infrastructure using BENIT Historian and EMS in the second half of last year and began building an integrated manufacturing operations platform in February this year.

Kolon Benit has designated this year as the starting point for the transition to Cognitive Autonomous Manufacturing. The company plans to integrate not only process data but also enterprise resource and supply chain management information from ERP and supply chain management (SCM) systems onto a single platform, advancing toward a structure in which AI assesses process conditions and controls operations. In the long term, it aims to realize a dark factory with minimal human intervention.

​To achieve this, Kolon Benit plans to expand the application of RDE based on r-CoCoAna at actual production sites and build a digital twin within the year to expand into remote integrated monitoring. PromptON Pak will also be expanded from quality inspection into safety monitoring and equipment predictive maintenance. Through its newly established AX Center and manufacturing digital transformation consulting team, the company plans to support the entire process from diagnosing manufacturing sites to implementation, operation, and expansion.

​Kolon Benit highlights three areas as its differentiated competitive strengths: operational technology (OT) capabilities accumulated in continuous-process manufacturing fields including chemicals, materials, films, biotechnology, and mobility; the ability to standardize and connect heterogeneous data, including data from aging equipment; and the execution capabilities required to bring multiple technology companies together and deploy their solutions in actual operations. The company particularly emphasizes technology neutrality—diagnosing operations and defining problems before forming a consortium of companies possessing relevant technologies and references. Its combination of Korean LLMs and NPUs for sovereign, on-premises AI is an extension of this strategy.

​◇POSCO DX Emphasizes Full-Stack AI, IT, and OT Capabilities

​POSCO DX stated that it develops and utilizes domain-specific AI models by fine-tuning open-source foundation models to reflect domain data and work characteristics at manufacturing sites. These models are primarily applied to simple, repetitive, or dangerous tasks, such as equipment operation and control and support for production operations.

The company also selectively applies models of different sizes depending on site requirements. POSCO DX emphasized that it focuses on ensuring accuracy and applicability by training models on relevant documents, process and equipment data, and work-knowledge data, followed by fine-tuning them to meet the characteristics and performance standards demanded by operations.

​POSCO DX describes itself as a full-stack company with integrated capabilities spanning on-site equipment and operational technology (OT) through management systems and information technology (IT). It plans to combine OT, IT, and AI capabilities to evolve into an "AI-native company" capable of delivering AI services optimized for industrial and office environments, including physical AI and agentic AI services.

​◇Domain- and Task-Specific Language Models Will Drive the Enterprise Generative AI Market

​The market for domain- and task-specific language models (DSLMs) is still in its early stages. According to a Gartner forecast report, the DSLMs and specialized generative AI models segment represents the smallest share of the market but exhibits the highest growth rate, at 210%.

​The report estimates that worldwide end-user spending on AI models and platforms will reach $64 billion, up 63.4% from $39 billion in the previous year. Gartner divides the market into four segments: foundation generative AI models; DSLMs and specialized GenAI models; AI application development platforms (AI ADP); and AI platforms for data science and machine learning, with the latter being the largest segment.

​Previously, Gartner projected that DSLMs would account for just 1% of enterprise AI models used in 2024, but revised that estimate to more than half by 2027, and subsequently to more than 60% by 2028.

​DSLMs are not restricted to specific industries; models specialized for functions such as security, marketing, and sales also fall into this category. Their primary role is to replace broad-topic LLMs that often fail to provide sufficiently accurate information for niche industries and specific tasks. According to Gartner research, DSLMs can significantly reduce costs and deployment times while enhancing reliability, relevance, and performance. This explains why the manufacturing sector is closely monitoring domain-specific models and DSLMs as key enablers of the real-time responsiveness and site-specific decision-making required by physical AI and autonomous manufacturing.

​Ultimately, the manufacturing AI strategies of mid-sized Korean companies converge on a single objective: automation through AI. Although different terminology is used, their overarching direction is remarkably similar—transitioning AI from data analysis to task execution, and ultimately to decision-making and control across manufacturing processes.