AI FUELS SMART MANUFACTURING

As manufacturers move beyond automation, artificial intelligence is emerging as the next driver of operational excellence. However, realizing its full potential requires more than deploying advanced algorithms—it demands trusted data, contextual intelligence, and human oversight.

Over the last few years, I have walked the floors of more than 25 factories and worked with over 80. As an engineer, the fascination never fades. I have watched imbibition water land on a bed of crushed sugarcane, a lab validate a pharmaceutical batch molecule by molecule, an air conditioner take shape across two miles of assembly line, and a semiconductor fab run in air cleaner than an operating theatre.

The Next Frontier

The last decade was the era of automation. The factory built a body. Sensors gave it the ability to feel; automation systems enabled it to act. In most advanced plants, that body is now firmly in place. The next frontier is intelligence—the ability to anticipate what lies ahead, notice signs that something is going wrong, and recommend the next course of action. That, in essence, is smart manufacturing and it rests on three capabilities: forecasting, anomaly detection, and optimization.

Intelligence needs something to think with, and that is where many AI projects falter. Walk into almost any manufacturing plant and a familiar pattern emerges: sensor data exists, but confidence in it is low. Critical readings are still logged by hand, while information remains scattered across SCADA, historians, LIMS, ERP systems, and spreadsheets, leaving no single source of truth. Without a trusted data foundation, even the most advanced AI models cannot deliver reliable outcomes.

Before intelligence comes memory—a governed data layer built through data logistics. It provides a catalog of available data, traces its lineage from the shop floor to the dashboard, and establishes semantic models so that ‘yield’ means the same thing across the enterprise. It is unglamorous work, but it creates the Information Architecture (IA) that makes AI possible. Generative AI, after all, is only one branch of the broader AI ecosystem.

Forecasting with AI

Forecasting is foresight, and it is where AI most clearly beats traditional spreadsheets. On the shop floor, demand, warranty returns, and part failures are often treated as random events to be managed with safety stock. In reality, they are rarely random. A warranty return is often the delayed consequence of a sale made years earlier, while many components follow predictable failure curves three to five years after deployment. When these timelines are mapped against the installed base, what appears to be noise begins to reveal clear patterns.

Working with manufacturers such as Daikin and Bourns taught me that forecasting accuracy depends less on sophisticated algorithms than on the breadth of data—often drawing on more than 500 external variables, from weather to construction—and the discipline of data management. Classical methods typically deliver error rates of 15–40 percent, while multivariate machine learning models routinely reaches 5–15 percent. That gap is the difference between the forecast finance trusts and one it overrides.

Detecting the Unexpected

Anomaly detection is a sharper sense, and perhaps the most undervalued capability of AI. A plant produces far more signals than any control room can monitor, and the anomalies worth catching rarely reveal themselves through a single tag. Instead, they lie in the relationship between a decision and its consequence—a wash time and final-product purity, or a cleaning schedule and steam economy. AI can monitor all of these relationships continuously, keeping the closest watch on assets one cannot afford to lose: mills, turbines, boilers, and centrifugals.

The real craft lies in tuning the alerts with the operators themselves. A tool the shop floor learns to ignore is worse than having no tool at all. In factory operations, that continuous watch transforms reliability from a monthly post-mortem into something managed in real time. We have seen downtime fall from 4.89 percent to 0.03 percent.

Optimization with Human Oversight

Optimization is judgment, and it has to be earned. Once the plant trusts what the system sees, the same models can recommend decisions: a setpoint, a cleaning interval, or a scrap-or-keep call. Optimization that comes before people trust the underlying detection simply gets switched off. The version that lasts keeps a human in the loop: the model recommends, the operator decides and acts, and the manager reviews at regular intervals. Here, intelligence is an advisor with an audit trail.

Building Trust in AI

For Indian manufacturers, the good news is that AI does not require replacing existing control systems or ERP platforms. Instead, these capabilities sit as an intelligent layer above the current infrastructure, drawing data from systems and feeding recommendations back into the manufacturing operations.

Manufacturers should calibrate against their own measurements rather than industry reference values and rely only on numbers validated by their own data. The next wave of smart manufacturing will not belong to those who simply adopt the most advanced AI models. It will belong to those who build the trustworthy data foundation, apply intelligence to the decisions that matter most, and have the patience to earn one proven trust result at a time. The body is built. The interesting work now lies in orchestrating it with intelligence.

Source: Isgec Heavy Engineering Ltd

SAGAR MAHURKAR Vice President Findability Sciences