On the factory floor, large models are starting to take over from experienced workers.

```

Everyone thought that large models would first replace people sitting in offices.

Writing copy, writing code, reviewing contracts, making reports... these jobs are inherently screen-based, with data naturally in text form, and seem most suited to be handled by AI.

But the first people to actually be learned by large models were a different group.

They stand in front of steel furnaces, judging the color of flames; crouch next to compressors, listening to machine sounds; watch gauges and material changes, using decades of experience to decide the next move. This craftsmanship has never been written into SOPs, nor is it easy to express in words; it only exists in the eyes, ears, and intuition of veteran workers.

Logically, this should be the hardest ability for AI to replicate.

The result is just the opposite.

After the intelligent steelmaking large model went online at Yongyang Special Steel in Hebei, it learned from 200,000 heats of historical data, processed signals from tens of thousands of sensors per second, and raised the end-point carbon pass rate from 75% to 97%; Shandong Haihua turned the veterans’ “listening to sounds to diagnose faults” skills into an intelligent control large model with over 95% accuracy in equipment fault prediction; Guangxi Huasheng shortened originally six-hour manual laboratory tests to three-minute predictions; Zhongtian Technologies compressed the original six-hour manual fiber optic path planning by masters down to forty minutes.

It looks like four industries, four systems.

In reality, they are doing the same thing: turning the veterans’ decades of tacit knowledge into data capabilities that large models can learn, copy, and use for real-time decisions.

This is the most counterintuitive aspect of this round of industrial AI.

When the veteran retires, the experience leaves with them

Dai Ziwei, Director of Yongyang Special Steel Enterprise Management Department, said: “In the past, steelmaking relied entirely on the veterans’ experience.”

Converter steelmaking is a black box process—blowing high-pressure oxygen into hot molten iron, completing decarburization, heating, and impurity removal in five minutes. Each batch of molten iron's composition is different, and even subtle changes in oxygen flow and lance position affect the final carbon content. Veterans judge by the color and sound of the flames—their skill level determines the pass rate.

A 75% pass rate means that one in every four heats fails. Failing means rework, loss of heat, waste of materials. All the experience accumulated by seniors over a lifetime leaves with them when they retire. Newcomers have to start from scratch, and the pass rate drops again for years.

Shandong Haihua faces the same problem. Chlorine compressors run 24 hours in highly corrosive environments—the 'heart' of the chlor-alkali industry, expensive, and with no backup. Minor vibration anomalies are hard to detect; once the compressor makes unusual noises, senior workers have to use a stethoscope against the machine to listen, guessing the problem by experience—sometimes repairs take three or four days.

Yan Guohui, General Manager of Process and Digitization at Shandong Haihua, said the company always struggled between “over-maintenance” and “unplanned downtime”. A single chlor-alkali production line going down for an hour loses over 30,000 RMB directly, and every downtime takes at least 8 hours to recover. To be safe, the company adopts a “replace early rather than late” conservative policy: minor repairs every year, major repairs every two years. Annual maintenance costs are 800,000 RMB.

Huasheng Aluminum’s pain point is even more direct. Alumina production is complex, with key indicators having poor timeliness and operational adjustments relying on experience. Manual tests for a set of indices take six hours for results. By then, the operating conditions have already changed; the data at hand is "outdated".

Zhongtian's optical cable fiber allocation is also entirely manual. Workers plan optical fiber routes using drawings, which is time-consuming, labor-intensive, and error-prone. Daily fiber assignment takes six hours. Over 50 million RMB is long tied up in raw materials because they don’t know exactly what will be needed and when, so must overstock.

The common feature in all four industry scenarios: the most critical knowledge is inside people’s heads; when the person leaves, the knowledge leaves with them.

From watching flames to reading data

The converter model at Yongyang Special Steel learned from 200,000 heats of reliable, historical smelting data. Each heat’s iron composition, temperature, oxygen flow, lance height, end-point carbon content—they're all recorded. The large model digests this data, finds the rules for “what parameters to use under what conditions”, and then gives the optimal control plan before each smelt.

The results are directly reflected in the numbers. Pass rate rose from 75% to 97%, material consumption per ton of steel dropped by 4.5 kg, smelting cycle time shortened from 35 to 30 minutes, and cost per ton fell by more than 20 RMB.

Dai Ziwei did the math: every percentage point increase in carbon-qualified rate means less heat and material loss. “Digital-intelligent transformation isn’t the old money-burning road, it’s a new profitable path.”

Hebei’s steel industry has already begun replicating these capabilities to more factories. HBIS Tangsteel used an integrated scheduling model to shorten raw material turnover from 10 days to 5 days, generating annual profit of over 10 million RMB. HBIS Handan Steel’s quality model improved key product pass rate by 8%, reducing losses by 12.6 million RMB per year. Shougang Qian’an’s AI production control model saves 70 million RMB a year and reduces CO2 emissions by 40,200 tons.

Data from the Hebei Provincial Industry and Information Technology Department show that all steel companies in the province have applied AI large models to varying degrees, with over half deepening applications in intelligent industrial control. From January to November 2025, Hebei’s steel industry profits reached 28.135 billion RMB, a year-on-year increase of 16.8 times, and profit per ton 26.25% higher than the national average.

Translating “listening for faults” into technical parameters

Shandong Haihua’s transformation began in February 2025. In partnership with Inspur Digital Enterprise, and based on the Inspur Haiyue Model Ch1, they built a smart control model for salt chemistry, creating three agents: predictive maintenance, process optimization, and smart inspection.

The deployment process for the predictive maintenance agent was itself one of “experience extraction.” At first deployment in May 2025, accuracy was only 60% to 70%. Vice General Manager Wu Mingfu of Inspur’s Digital Enterprise Platform Product Division said, engineers worked repeatedly with senior workers, turning "listening for faults" and "judging wear by vibration" into technical parameters, then optimized through combining large and small models. By August, accuracy improved to 90%, and later exceeded 95%.

The veterans’ stethoscopes retired. The large model now monitors equipment in real time, precisely predicts maintenance periods, and fault recognition accuracy exceeds 95%. Chlorine compressor unplanned shutdowns dropped from four times in 2024 to zero in 2025. Maintenance costs fell from 800,000 to 200,000 RMB annually.

The process optimization agent addresses another challenge. Electrolysis processes are affected by multiple coupled parameters like current, temperature, acid addition—manual calculation can’t achieve optimal balance across indices. The model tracks data in real time, achieving minute-level autonomous process optimization. The electrolysis cell process model helps the plant save 4.5 million kWh per year, extends membrane life from 4 to 5 years, and generates nearly 10 million RMB in total benefits.

Yan Guohui’s ledger shows: phase I investment of 32 million RMB, expected returns of 23 million RMB in 2025; process stability rate up 55%, automated control rate of key indicators up 61%, manual operation frequency down 68%. Business processes streamlined from 1,803 items to 400, a 77.8% reduction.

6 Hours to 3 Minutes

Guangxi Huasheng’s alumina large model, “Zhisheng”, is developed with Chinalco Group’s “Kunan” model, partners including Chinalco Smart and Central South University. Using a “four-horizontal, four-vertical” architecture, it optimizes from data sensing to intelligent decision-making.

The most direct result is testing timeliness. What used to take six hours with manual testing, the large model now predicts within three minutes. Key index prediction accuracy exceeds 80%, and dissolving aK accuracy breaks 90%. Main control workload falls 85%, sampling and testing efficiency rises 30%.

Single plant annual benefits are in the tens of millions RMB. Full Chinalco Group rollout is expected to save 130 million RMB a year. The project was also selected as a 2025 MIIT (Ministry of Industry and Information Technology) typical case of manufacturing digital transformation, the only such project in Guangxi.

Zhongtian’s “Tianji” model took another route. In 2025, they set up an AI research center, developed their own industry-specific model. After the “smart fiber planning” model went online at the main optical cable plant, only one person is needed to operate; daily fiber planning shrank from six hours to 40 minutes. Raw material reserves dropped from over 50 million to less than 10 million RMB. After expansion to transformer divisions, strip material stock dropped from 120 tons to 75 tons, with the turnover period shortened from 45 to 20 days.

Tianji model achieved results in other areas too: bobbin residue detection responds in 100 milliseconds, saving over 80% labor; mandrel length precision improved from 2 mm to 0.5 mm, surpassing imported Japanese technology. Ultra-complex image-text parsing processes thousands of pages per second, reducing technical staff by 25% and error rates by 74%. After AI was introduced to the RF plant, labor costs dropped 80%, quality increased 28%.

Chairman Xue Chi of Zhongtian commented: “AI is not some mystical concept; it is a core tool for deeply cultivating our main business and improving quality and efficiency.”

Four Steps to Put Experience Into Models

If you break down these four companies’ approaches, the paths are surprisingly similar.

Step 1: Collect historical data. Yongyang Special Steel collected 200,000 furnace records. Shandong Haihua worked with seniors to turn experience into technical parameters. Huasheng Aluminum expanded testing parameter sets for evaporation, dissolution, precipitation. Zhongtian Technology mapped 50 typical scenarios across nine workflow stages. Where does the data come from? From decades of sensor recordings, lab reports, and operation logs on the production line. Without these historical data, large models can’t learn.

Step 2: Train an industrial large model. Generic models cannot understand industrial data. Yongyang and Northeastern University jointly developed a steelmaking-specific large model; Shandong Haihua used Inspur’s Haiyue model; Huasheng used Chinalco’s Kunan model; Zhongtian built its own Tianji model. The methods are the same: use generic models as the base, import industry data and process knowledge for fine-tuning, letting the models learn industry language.

Step 3: Embed into the production line for real-time decisions. The model doesn’t just run tests in labs—it is connected to control systems on the line. Yongyang’s model processes tens of thousands of sensor signals per second, adjusting blow curves in real time. Haihua’s agents monitor vibration data and auto-flag anomalies. Huasheng’s model provides predictions in three minutes for real-time parameter adjustment. Zhongtian’s planning model outputs routing solutions directly.

Step 4: Veterans shift from operators to annotators. Haihua’s case is clearest: engineers worked repeatedly with seniors, turning “listening for faults” and “judging wear by vibration” into technical parameters. The old skills weren’t thrown away, but translated into numbers models can understand. The process of moving from 60% to 95% accuracy is one of continually extracting and calibrating expert experience.

Crucially: the veterans’ expertise never needed to be written in words, just converted to data. Thirty years of experience in judging flames can’t be written into an SOP, but 200,000 heats’ sensor data already fully record the causal link of "under what conditions to stop blowing". Text is the easiest format to feed into a model, but it’s not the only one. Temperature curves, vibration waveforms, chemical components, current parameters—these structured signals coming from the line are more precise and even easier for models than text. That is why crafts like “watching the flame and listening to sounds”, which seems least likely to be optimized by AI, actually proved feasible first: it never lacked data, only lacked someone to feed the mapping between data and results to the model.

From resistance to reliance

The reaction of the veterans can’t be ignored. After working thirty years at a craft, to have it suddenly replaced by a machine will understandably cause feelings.

The deployment timeline at Haihua gives some clues. Launched in February 2025, the first deployment in May had only 60%-70% accuracy—at this stage, the seniors would have been most skeptical. If the model is less accurate than people, why trust it? By August, accuracy had climbed to 90%, and the model began outperforming people. When it later exceeded 95%, the seniors’ attitudes changed.

There is no shortcut. When accuracy is below 60%, the model is only a reference. At 90%, veterans start comparing its results to their own judgments. Above 95%, they may take the initiative to see what the model says. Huasheng Aluminum’s prediction accuracy is over 80%, dissolving aK accuracy broke 90%—manual lab work takes six hours for results, while the model predicts in three minutes. Precision is still climbing, but timeliness is already overwhelming.

Zhongtian’s approach is “pilot-verify-rollout”. Once a factory worked, they quickly expanded to others after evaluation. Zhang Xiangen said: “We encourage each plant to actively explore pain points.”

This means that every plant’s veterans can participate in tuning and verifying the models, rather than passively accepting a solution parachuted in.

Source: AI Original Lab

Risk Reminder and DisclaimerThe market carries risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account individual users’ special investment objectives, financial situations, or needs. Users should assess whether any opinions, views, or conclusions in this article apply to their own circumstances. Any investment made accordingly is at your own risk. ```