Summary
The controller was right on the day it was commissioned
Almost every industrial machine is controlled by a strategy that was tuned once. The gains were set on a test rig or at commissioning, against the load, the temperature and the supply that happened to be present, and then left. The machine then spends years meeting loads, temperatures, supplies and ageing effects that nobody tuned for.
The result is not failure. It is drift. Efficiency slides, ripple grows, protection margins shrink, and the people who notice are the maintenance team, late. This article sets out what adaptation actually means in a drive or a converter, why it has historically stayed in research, and what an architecture that adapts safely looks like.
Why control is fixed
Tuning rules were designed for a world without compute
The tuning rules most controllers still rest on were published by Ziegler and Nichols in 1942, for pneumatic controllers in process plants. They are a procedure for picking fixed gains from a step response. They are excellent at what they do. They assume the plant does not change.
Adaptive control, which does not assume that, is nearly as old. Åström and Wittenmark's standard text traces it to autopilot work in the 1950s. The theory of self-tuning regulators and model-reference adaptive control has been mature for decades. What kept it out of most industrial equipment was not theory. It was the cost of the sensing, the arithmetic and the assurance argument, against a commercial reality in which a fixed controller that works adequately is cheaper to certify and cheaper to support.
Two things have changed. Compute that can run a model of the machine every control period now fits inside the drive. And the pressure on efficiency has become regulatory rather than optional.
The pressure
Efficiency is now a compliance question, and the component ratings have been collected
The IEA 4E Electric Motor Systems Platform estimates that electric motor systems accounted for 53 percent of global electricity consumption in 2023. That makes the efficiency of drives and the machines they control one of the largest single levers on industrial energy use.
Regulators have responded by raising the minimum efficiency of the components. In the EU, Regulation 2019/1781 has required IE3 for most three-phase motors since July 2021 and IE4 for motors between 75 kW and 200 kW since July 2023, and sets minimum efficiency for variable speed drives. IEC 61800-9-2 defines how to classify the efficiency of the whole power drive system.
The component ratings are the easy part. Once the motor and the drive are both at their rated class, the remaining losses sit in how they are operated: the switching strategy, the control strategy and how well either suits the condition the machine is actually in. That is where adaptation lives, and it is not a component you can buy.
What actually changes
Four things move after commissioning
Temperature
Winding resistance rises with temperature, and permanent magnet flux falls: roughly 0.1 percent of remanence per degree for NdFeB magnets. A controller that assumes the cold parameters produces the wrong torque for the current it commands.
Load
Inertia, friction and the load cycle of a real installation are rarely the ones on the rig. Gains that were well damped on the bench are sluggish or oscillatory in the field.
Supply
A varying DC bus, from a battery discharging or a weak grid, changes the voltage the modulator has to work with and moves the optimum switching strategy.
Ageing
Bearings wear, capacitors lose capacitance, magnets weaken, insulation degrades. Each shifts the plant away from the model the controller was built on, slowly, until something trips.
What adaptation means
Four levels, and most machines need the second
Estimate
Track the parameters that drift: resistance, flux, inertia, bus voltage. Use the estimates in the existing control law. This is the least glamorous and most valuable form of adaptation, and it changes nothing about the control structure.
Re-tune within bounds
Adjust gains, dead time, modulation strategy or operating point within limits set at design time, based on the estimated state. The controller structure and its safety functions are unchanged; the numbers inside it move.
Suggest
A separate layer, with its own model of the machine, evaluates the state and proposes changes to the primary controller, which accepts or rejects them against fixed rules. The primary controller keeps its authority and its certification argument.
Learn
A learned component changes the control law itself. Powerful, and the hardest to assure. Outside research and a narrow set of non-safety functions, this is not where industrial machines are going next.
Most of the efficiency and most of the robustness is in the first three. The fourth is where the hard assurance problems live, and it is rarely needed.
The architecture
Adapt alongside the control, not inside it
The architecture that lets an existing machine become adaptive without rebuilding its control is a supplementary layer. The primary controller, deterministic and tuned and certified, keeps doing what it does. Beside it runs a model of the machine fed by the same sensors. The model estimates the state, compares it against expectation and produces bounded suggestions. The primary controller applies the ones that fall within rules fixed at design time.
This has three consequences that matter to an integrator. The safety argument for the primary loop does not change, because its behaviour under any input from the adaptive layer is bounded by its own rules. The adaptive layer can be developed, validated and updated on its own schedule. And the integration surface is an interface specification rather than a rewrite.
It also has a hardware consequence. The model has to run every control period, with its inputs sampled at the right moment, inside the drive's power and thermal budget. In practice that puts the estimator in FPGA fabric or on a dedicated real-time core beside the control, with acquisition timing shared between them. The companion article on digital twins at the edge covers what that demands.
What adaptation does not do
It cannot fix a power stage that was designed without the control in mind
Adaptation recovers the performance that a fixed controller leaves on the table. It does not add performance the hardware cannot deliver. A power stage with poor layout, an under-specified sensing chain or a thermal design with no headroom will be a mediocre adaptive drive just as it was a mediocre fixed one.
That is why we argue for designing the control, the sensing, the power stage and the adaptive layer as one system. The remaining efficiency in a drive sits at the boundary between switching strategy and control strategy, and that boundary is only visible to a team that owns both sides of it.
We will not quote an efficiency or ripple improvement for a machine we have not instrumented.
Next step
Is your drive efficient on paper and disappointing on the bench?
Talk to an engineer
Start a conversation
Send the electrical basics: motor type, DC bus, current, speed, torque and switching frequency. An engineer assesses architecture fit before anyone offers an opinion on cause.
Read next
Safe integration of adaptive intelligence into industrial control
The supplementary-layer pattern and what has to be tested.
