Eliminating blind spots: closing the data gaps in advanced process control

What is self-adaptive control?

Most advanced process control systems in cement rely on fixed process models. These are mathematical representations of how the process responds to control actions, typically built from historical data at a specific point in time. Model predictive control (MPC) is the most widely used example. Such models perform well when operating conditions remain close to those under which they were developed.

However, as conditions drift through raw material variation, equipment wear, fuel changes, or seasonal effects, the model becomes less representative of the process, predictions become less accurate, and control performance declines. Retuning can temporarily restore performance, but the cycle typically repeats.

Self-adaptive control takes a different approach. Rather than relying on a fixed model, it continuously evaluates real-time process data to estimate the direction and rate of change of key variables and adjusts control actions accordingly. This allows the controller to respond proactively, stabilising conditions before deviations escalate rather than reacting after the fact.

Because it adapts continuously to evolving process conditions, the need for periodic retuning is reduced. A single controller can manage the process across a wide range of operating states (stable, transitional and upset), helping to maintain consistent performance over time, even as underlying conditions change.

A new control paradigm

The shift towards self-adaptive control marks a fundamental change in how advanced process control systems are designed and applied. Unlike fixed-model approaches, which rely on predefined process models and periodic retuning, self-adaptive control continuously evaluates real-time process data, predicts how key variables are likely to change, and adjusts control actions to stabilise conditions before they escalate. In practice, this allows a single controller to manage the full process circuit across a wide range of operating conditions, reducing the need for scenario-specific tuning that conventional systems require.

GCC's experience at its Chihuahua plant in Mexico illustrates this shift in approach. The plant initially installed Fuller Technologies’ ECS/ProcessExpert® (PXP) 8.5 in early 2024. Early gains included a 2% increase in productivity. This led to an upgrade to PXP 9.1, which incorporates self-adaptive control, following its release later that year.

Compared to manual operation, PXP 9.1 delivered a 4.2% increase in kiln production, a 3.2% reduction in specific heat consumption, and a 30.8% improvement in free lime standard deviation (Figure 1). The plant also increased alternative fuel use while controlling CO and sulphur spikes that had previously limited higher thermal substitution rates (TSR). Process upsets became less frequent, and ring and coating formation were reduced. Plant Production Manager Ing. Carlos Guerrero Anaya Regalado described the outcome as a “total success” for this critical project.

Across more than 50 installations globally, similar patterns have emerged. While conventional advanced process control systems typically deliver improvements of 2–5% over manual operation, PXP 9.1 demonstrates gains of 4–7%, alongside improved stability and utilisation. At CEMEX Croatia, for example, self-adaptive control delivered a further 3–4% improvement in mill throughput and power consumption over the existing APC baseline, resulting in cumulative gains of 7–8% over manual operation.

Sustained utilisation is central to these outcomes. Because the controller continuously adapts to changing conditions, it reduces the performance drift that often leads to operator disengagement or system bypass. In doing so, it helps convert theoretical optimisation potential into consistent, long-term plant performance.

The remaining blind spots

A self-adaptive controller, however capable, can only optimise what it can see. In practice, there remain critical parameters that even well-instrumented cement plants cannot reliably observe in real time. These constraints reflect fundamental limitations of what physical measurement can deliver in a cement plant environment. Even the most advanced control system is therefore subject to blind spots: periods where critical process and quality information is unavailable or delayed.

AI-based soft sensors offer a way to address these limitations. Integrated into PXP 9.1 in partnership with Imubit, an industrial AI specialist with more than 100 closed-loop applications across refining, calcining and cement, they supplement existing instrumentation by providing continuous predictions in the intervals where physical sensors cannot reach.

How soft sensors work: the GCPV Monjos trial

While physical sensors measure process conditions directly, soft sensors infer them. Using deep learning models trained on a plant’s own operating and laboratory data, they generate continuous predictions of parameters that cannot be measured reliably in real time. Once deployed, these models run continuously, typically producing updated predictions every few minutes. New laboratory results are used to validate performance and adjust model bias automatically, maintaining accuracy over the years in evolving process conditions.

In PXP 9.1, these predictions are used as live input variables within the control strategy. This enables the controller to update targets far more frequently than a sample-driven approach allows, shifting from multi-hour adjustment cycles to minute intervals. The result is a more responsive control loop that can act on predicted quality changes rather than waiting for confirmation from laboratory analysis. Initial deployment focuses on free lime, Blaine fineness, and kiln inlet oxygen (Table 1).

The impact of this approach was demonstrated in a closed-loop trial at GCPV Monjos in Spain. Integrating a free lime soft sensor into its PXP system delivered predictions every five minutes, compared to the two-hour cycle of laboratory measurements. Where the controller had previously been forced to maintain a conservative operating position between laboratory samples, it could instead respond much more regularly, adjusting kiln torque targets every fifteen minutes: an eightfold increase in responsiveness compared to the standard laboratory cycle (Figure 2).

By tightening the control window around free lime, the system reduced both over-burning and under-burning, highlighting how closing data gaps with soft sensors translates directly into operational performance. The result was a 25% reduction in off-spec clinker, alongside a measurable 3.2% improvement in energy efficiency.

The optimization stack

The performance gains delivered by self-adaptive control and AI-based soft sensors are cumulative and complementary. The self-adaptive controller provides the foundation, maintaining stability and sustained utilisation across changing process conditions. Soft sensors build on that foundation by extending the controller’s visibility into parameters beyond the reach of conventional instrumentation.

At GCC Chihuahua, the impact of self-adaptive control demonstrated what can be achieved when advanced process control maintains performance over time. At GCPV Monjos, soft sensors showed what changes when the controller no longer has to wait for laboratory results. Together, they represent a shift to continuous, data-driven control.

This layered approach reflects a broader evolution in cement process optimisation. For decades, the focus has been on improving control algorithms and tuning strategies. Increasingly, the constraint is no longer how well the controller acts, but how much it can see, and how quickly it receives that information. By closing these data gaps, soft sensors extend the reach of advanced process control into areas that were previously beyond its limits.

Advanced process control has long proven it can deliver improvements at commissioning. The lingering question has always been whether those improvements can be sustained and extended over time. By combining self-adaptive control with continuous, AI-driven predictions from integrated soft sensors, the answer is, increasingly, yes.