Reframing Process Performance
Manufacturing performance has traditionally been measured through outcomes: output, downtime, yield, scrap, energy consumption, cycle time, quality defects and Overall Equipment Effectiveness. These metrics remain important for review, benchmarking and accountability. But they often describe what happened after the process has already run.
This limitation is especially visible in batch and process industries such as pharmaceuticals, chemicals, food and beverage, specialty materials, metals, cement, refining and energy-intensive manufacturing. In these environments, equipment may continue running while the process quietly drifts away from its ideal condition. A reactor may remain operational, a furnace may complete its cycle, a fermenter may finish its batch and a line may report acceptable availability. Yet the process may have spent too much time outside the preferred operating envelope.
The result may be lower yield, inconsistent quality, excessive energy consumption, rework, shortened shelf life or avoidable process variation.
This is where Time-in-State becomes powerful. Instead of asking only whether the asset was running or whether the final batch passed, Time-in-State asks a more operationally useful question:
How much time did the process spend in ideal, acceptable, degraded, critical or fault conditions?
However, Time-in-State should not be positioned as the only leading indicator. It is better understood as an integrating managed metric. It connects classical control, statistical monitoring, predictive analytics and AI-enabled action into a common performance language.
The broader journey can be seen as:
Closed-loop control → Statistical Process Control → Time-in-State → Machine Learning → Digital Twins / Model Predictive Control → Agentic AI
Each stage adds a different capability. Closed-loop control corrects error. SPC detects abnormal variation. Time-in-State converts process behaviour into operating-state health. Machine learning predicts future risk. Digital twins and MPC optimize future trajectories. Agentic AI coordinates diagnosis, recommendation and action.
1. Closed-Loop Control: The Engineering Foundation
Katsuhiko Ogata’s work on modern control engineering provides the classical foundation for understanding closed loop process control. In a closed-loop system, a desired setpoint is compared with the measured process variable. The difference between the two is the error signal. The controller acts on this error and adjusts the process through actuators so that the measured variable moves closer to the setpoint.
In simplified terms:
Error = Setpoint – Process Variable
The purpose of the controller is to reduce this error over time. Concepts such as transient response, overshoot, settling time, stability and steady-state error help engineers assess whether the control system is behaving properly.
This remains essential. Without closed-loop control, modern process manufacturing would be unstable, inconsistent and dependent on manual intervention. Temperature, pressure, flow, level, humidity, speed, torque, pH and many other process variables depend on disciplined control loops.
But closed-loop control mainly answers an engineering question:
Is the process variable close to the setpoint, and is the control system correcting deviations effectively?
This is necessary, but not sufficient for business performance. A loop may be stable, but the process may still spend too much cumulative time near the edge of acceptable performance. A controller may correct deviations, but the batch may remain exposed to quality risk. A line may remain available, but energy intensity may quietly rise.
Closed loop control stabilizes the process. It does not, by itself, provide a complete managerial language for process health, quality exposure, energy loss or value risk.
2. Statistical Process Control: Detecting Abnormal Variation
Statistical Process Control (SPC) adds a different lens. While closed-loop control focuses on correcting deviations from setpoint, SPC focuses on whether the process is statistically stable.
SPC asks:
Is the process behaving normally, or is there evidence of abnormal variation?
The typical tools include control charts, upper and lower control limits, process capability indices such as Cp and Cpk, run rules, trend patterns and special-cause signals. SPC is valuable because it can detect deterioration before the final defect is visible.
For example, in a machining process, part dimensions may still be within specification, but a control chart may show a sustained upward trend. This may indicate tool wear. If the operator waits for parts to go out of specification, scrap has already occurred.
In a chemical reactor, temperature may not breach the alarm limit, but repeated operation above the process centerline may indicate a cooling, valve or heat-transfer issue.
SPC is therefore a concurrent and sometimes leading indicator. It is more useful for intervention than final yield or scrap. But SPC often remains variable-centric. It may not fully represent the combined operating state of a complex process where several variables interact.
That is where Time-in-State becomes useful.
3. Time-in-State: From Variable Control to Process-State Health
Time-in-State does not replace closed-loop control or SPC. It sits above them.
A control loop may regulate temperature. Another may regulate pressure. Another may regulate flow. SPC may monitor whether each variable is statistically stable. But the true health of the process often depends on whether these variables are simultaneously within a desired operating envelope.
Time-in-State converts continuous process behaviour into interpretable operating states:
Ideal: The process is within the preferred operating envelope.
Acceptable: The process is safe and within tolerance, but not optimal.
Degraded: The process is drifting and may require intervention.
Critical: The process is close to quality loss, safety risk, energy waste or failure.
Fault: The process is outside acceptable limits.
A batch may therefore be described as having spent 72% of its duration in Ideal state, 18% in Acceptable, 8% in Degraded and 2% in Critical. This is a different performance language from saying that the batch passed, the line ran or the equipment had acceptable OEE.
The advantage is that Time-in-State adds temporal context. A deviation that lasts for ten seconds may not matter. A mild deviation that lasts for three hours may matter significantly. The state, severity and duration all become visible.
For example, in a pharmaceutical batch, temperature, pH, agitation speed and residence time may each remain individually within broad limits. But if the combined process spends too long in a degraded state, the product may face yield or quality risk.
Time-in-State therefore moves the discussion from single-variable control to multivariable process-state management.
4. Outcome Metrics: Necessary but Mostly Lagging
OEE, yield, scrap, downtime, rework, customer complaints and final quality results remain important. They are not wrong metrics. They are necessary for management review and financial accountability.
But they are mostly lagging indicators.
OEE asks whether equipment was available, performing and producing good output. This works well in high-speed discrete manufacturing, where the dominant question is often whether the machine is running and producing acceptable units.
But in process industries, the equipment may continue running while the process deteriorates. The machine has not stopped; the process has drifted.
A reactor may complete the batch, but with lower yield. A furnace may complete the cycle, but with excessive energy consumption. A packaging line may remain available, but seal quality may deteriorate because of temperature, pressure, film or speed variation.
Outcome metrics tell the organization what happened. Time-in-State and related leading indicators help the organization intervene before the outcome is locked in.
5. Multivariate Analytics and Machine Learning: Predicting Process Risk
Many real processes cannot be understood one variable at a time. Temperature, pressure, pH, viscosity, flow rate, humidity, residence time, agitation, raw-material variation and equipment condition may interact. A variable may look acceptable in isolation but problematic in combination with others.
Multivariate analytics addresses this problem by detecting patterns across variables. Techniques such as Principal Component Analysis (PCA), Partial Least Squares (PLS), correlation analysis, multivariate control charts and residual monitoring can identify abnormal process behaviour before any individual variable crosses its limit.
In a refinery unit, pressure and temperature may each remain within limits, but their combined trajectory may indicate fouling or catalyst deterioration.
Machine learning (ML) extends this journey from detection to prediction. Instead of asking only whether the process is currently abnormal, ML asks:
What is likely to happen next?
ML models can predict final yield, quality failure, energy consumption, equipment failure, remaining useful life, anomaly probability or likely movement from one process state to another.
For example, an ML model may predict that a batch currently in Acceptable state has a high probability of moving into Degraded state within the next 30 minutes. A predictive-maintenance model may detect that vibration, current draw and temperature patterns suggest an impending pump failure. An energy model may predict that a furnace will exceed target specific energy consumption unless the operating profile is corrected.
Time-in-State is useful here because it provides a meaningful target for analytics. Instead of training models only to predict final failure, organizations can train them to predict movement from Ideal to Acceptable, Acceptable to Degraded, or Degraded to Critical.
In this sense, Time-in-State is not the only leading indicator. It is also a practical target, feature and explanatory framework for machine learning.
6. Digital Twins and Model Predictive Control: Optimizing the Future Trajectory
Digital twins and Model Predictive Control (MPC) take the next step. They do not merely detect or predict deterioration. They simulate future trajectories and identify actions that can improve the outcome.
The question becomes:
What action now will improve the future state of the process?
A digital twin can estimate how the process will behave under different operating choices. Model Predictive Control can use these estimates to optimize control actions while respecting constraints.
For example, in a chemical reactor, MPC may adjust temperature, feed rate or pressure to keep the process within an optimal trajectory. In a kiln, it may optimize fuel, airflow and feed characteristics to improve thermal stability and energy efficiency. In continuous manufacturing, it may balance throughput, quality and energy consumption dynamically.
Time-in-State can become part of the optimization objective. Instead of merely minimizing instantaneous error, the system can seek to maximize future time in Ideal state, reduce time in Degraded state and avoid Critical or Fault states.
This is a major shift.
Classical feedback control reacts to current error. Predictive control anticipates future behaviour. Time-in-State provides a business-relevant way to evaluate whether that future behaviour is desirable.
7. Agentic AI: From Insight to Coordinated Action
Agentic AI adds another layer. It should not be seen as a replacement for controllers, safety systems, operators or engineers. In industrial environments, this distinction is critical. Safety-critical control must remain governed by validated control systems, operating procedures and human oversight.
The role of agentic AI is better framed as decision orchestration.
Agentic AI can monitor process states, interpret deviations, investigate probable causes, recommend interventions, retrieve standard operating procedures, raise alerts, prepare work orders, coordinate with maintenance systems and escalate decisions to the right human owner.
The question shifts again:
What should be diagnosed, recommended, escalated or coordinated?
For example, if a reactor repeatedly moves from Ideal to Degraded state during a specific phase of the batch, an agentic system could compare prior batches, examine historian data, check raw-material lots, review recent maintenance activity and suggest likely causes.
If a furnace spends excessive time in an energy-inefficient Acceptable state, the agent could compare shift patterns, fuel parameters, loading conditions and maintenance history, then recommend corrective action.
The value of agentic AI lies in connecting detection, diagnosis, decision and workflow. But it must be carefully governed. It should support operators and engineers, not bypass them. It should explain recommendations, respect safety constraints, maintain audit trails and operate within approved decision rights.
8. Comparative View of Process-Performance Intelligence
| Stage | Core Question | Typical Signal | Nature of Indicator | Primary Value |
|---|---|---|---|---|
| Closed-loop control | Is the process variable close to setpoint? | Error, overshoot, settling time, stability, steady-state error | Real-time / concurrent | Stabilizes the process |
| SPC | Is the process statistically stable? | Control charts, control limits, Cp, Cpk, special-cause signals | Concurrent to leading | Detects abnormal variation |
| Outcome metrics | Did the process deliver the expected result? | OEE, yield, scrap, downtime, rework, defects | Mostly lagging | Supports accountability |
| Time-in-State | How much time was spent in desired or undesired states? | % Ideal, Acceptable, Degraded, Critical, Fault | Concurrent and leading | Translates process behaviour into operational health |
| Multivariate analytics | Are combinations of variables drifting? | PCA/PLS scores, residuals, correlation patterns | Leading | Detects interaction effects |
| Machine learning | What is likely to happen next? | Failure probability, predicted yield, anomaly score, state-transition risk | Predictive / leading | Anticipates future risk |
| Digital twin / MPC | What action improves the future trajectory? | Simulated trajectory, optimized control path, avoided constraint violations | Predictive and prescriptive | Optimizes future performance |
| Agentic AI | What should be diagnosed, recommended or escalated? | Recommendations, root-cause hypotheses, work orders, SOP guidance | Prescriptive / action-oriented | Converts insight into governed action |
9. Two Illustrative Examples
Batch Reactor
In a specialty chemicals reactor, closed-loop control maintains temperature and pressure. SPC shows that temperature variation is increasing across recent batches. Time-in-State shows that the current batch spent only 64% of its duration in Ideal state and 12% in Degraded state. ML predicts a higher probability of yield loss if the trajectory continues. A digital twin evaluates alternate feed and heating profiles. Agentic AI summarizes the issue, compares prior batches, identifies likely causes and alerts the process engineer.
The insight is not merely that the reactor ran. The insight is that the product did not spend enough time in value-protecting conditions. The risk is visible before the customer complaint occurs.
Oil and Gas Processing Unit
Consider a crude distillation unit, gas processing plant, compressor train or refinery hydrotreater.
At the closed-loop control layer, temperature, pressure, flow, level and composition-related variables are continuously regulated. Controllers keep process variables close to setpoints and respond to disturbances such as feed variability, ambient conditions or equipment behaviour.
At the SPC layer, control charts may reveal that pressure variation, column temperature profile, compressor vibration or product-quality parameters are becoming less stable over time, even though they remain within operating limits.
At the Time-in-State layer, the unit may show that it spent 70% of the shift in Ideal operating state, 20% in Acceptable state, 8% in Degraded state and 2% in Critical state. This may indicate that the unit continued running, but not always in the most energy-efficient, quality-supporting or reliability-protecting condition.
At the machine-learning layer, models may predict higher risk of off-spec product, compressor trip, fouling, catalyst deactivation, excessive energy consumption or unplanned maintenance based on current trajectory.
At the digital twin or MPC layer, alternate operating strategies can be simulated to optimize throughput, product quality, energy intensity and constraint management.
At the agentic AI layer, the system could summarize the deviation, compare it with past operating patterns, check recent maintenance history, examine feed-quality changes, retrieve relevant operating procedures and recommend escalation to the process engineer, reliability engineer or shift supervisor.
The insight is not merely whether the oil and gas unit remained online. The deeper question is how much time it spent in operating conditions that protected yield, product quality, energy efficiency, asset integrity, safety and margin.
10. From Error Minimization to Value Maximization
The broader shift can be expressed simply as follows:
Ogata gives the engineering logic of correction.
SPC gives the statistical logic of stability.
Time-in-State gives the operational logic of process health.
Machine learning gives the predictive logic of future risk.
Digital twins and MPC give the optimization logic of future trajectory.
Agentic AI gives the decision and action logic for smart manufacturing.
The development of manufacturing performance management will require an architecture that connects them.
Closed-loop control remains the foundation. SPC adds statistical discipline. Time-in-State converts process behaviour into operating-state language. Multivariate analytics and machine learning identify complex patterns and predict future risk. Digital twins and MPC optimize future actions. Agentic AI coordinates diagnosis, recommendation and workflow.
The management challenge is to avoid treating these as disconnected initiatives. If each layer remains isolated, plants may end up with more dashboards, more alerts and more complexity, but not necessarily better decisions.
The opportunity is to connect them through a managed performance architecture.
Time-in-State as the Integrating Managed Metric
Time-in-State is not the only leading indicator. SPC, multivariate analytics, machine learning, digital twins and MPC can all provide leading signals. But Time-in-State may be one of the most useful managerial abstractions for process industries because it converts technical process behaviour into a language that operators, engineers, quality teams, plant managers, data scientists and business leaders can share.
Closed-loop control is focused on: Is the process variable close to the setpoint?
SPC is focused on: Is the process statistically stable?
Time-in-State is focused on: How much time did the process spend in conditions that protect quality, yield, safety, energy efficiency and business value?
Machine learning is focused on: What is likely to happen next?
Digital twins and MPC are focused on: What action improves the future trajectory?
Agentic AI is focused on: What should be diagnosed, recommended, escalated or coordinated?
The first shift was from measuring whether the machine ran, to understanding how well the process lived while it ran.
The next shift is from understanding process health to acting intelligently, safely and economically before value is lost.
