Data-Driven Calibration and Control using Cost-Effective Actuators and Sensors

Precision engineering and control are fundamental to technologies that shape modern society, including satellite communication systems and advanced scientific instrumentation, enabling breakthroughs across disciplines. Making these technologies more broadly accessible stimulates further innovation, as affordable systems are more readily reused, adapted, and integrated into new applications. Achieving this requires scalable production methods that deliver consistent precision across large product volumes while minimizing user intervention, computational resources, and component costs.

Cost-effective components often introduce imperfections, such as torque ripple in actuators or position-dependent measurement errors in sensors. If these flaws are sufficiently repeatable over the product lifetime, they can be compensated through software during deployment. At the same time, cost-saving measures often result in more significant product variation, exacerbating the need for robust calibration solutions that preserve high precision at scale.

This thesis introduces a unifying data-driven framework for large-scale calibration and control of cost-effective actuators and calibration of cost-effective sensors. The main challenge is that cost-effective components introduce consistent but unknown variability, requiring advanced software-based compensation to meet strict precision requirements without expensive hardware upgrades. To address this need, calibration and control methods are developed that combine Gaussian processes, nonlinear system identification, and learning control with optimization-based approaches to deliver practical tools for industrial deployment. All developed methods are experimentally validated on industrially relevant setups, demonstrating their effectiveness in improving performance and reducing costs across both actuation and sensing tasks.

The first part of this thesis develops scalable calibration methods for actuation, with Switched Reluctance Motors (SRMs) as the primary focus. These actuators are mechanically simple and cost-effective but require advanced control strategies to manage their nonlinear dynamics and manufacturing variability. Accurate models of SRM dynamics are obtained from noisy position sensor data, enabling robust control strategies that account for manufacturing-induced product differences without relying on expensive torque sensors.

Beyond this model-based approach, a model-free method based on extremum-seeking control is introduced. This method eliminates the need for a modeling step, adapting control parameters directly during operation. While it requires more experimental data and offers less design flexibility than the model-based method, it provides a viable alternative for large-scale applications where user intervention must be minimized.

In addition to SRMs, control strategies for piezo-stepper actuators are developed. These actuators are critical for nano-positioning stages due to their ability to deliver precise, long-stroke motion. Although typically used in high-end systems, cost remains a concern: mechanical remedies for hardware imperfections are expensive, and adding high-resolution sensors for high-gain feedback control increases costs. The data-driven techniques developed in this thesis compensate for parasitic effects such as hysteresis and mechanical misalignments in software, enabling accurate and flexible positioning without additional hardware.

The second part of this thesis focuses on scalable sensor calibration of cost-effective position sensors. Two complementary approaches are developed to achieve the high degree of automation required in mass production: one relies on external reference encoders, while the other requires no additional sensors.
For calibration using external reference encoders, a dedicated test bench is used to calibrate sensors automatically. Imperfections in this test bench propagate errors to all products it calibrates, leading to a cascaded calibration problem. Gaussian process regression is employed to address this issue, significantly reducing the propagation of errors. Additionally, active alignment of the test bench to the sensor is achieved with minimal modeling effort using data-driven techniques.

Alternatively, repeatable inaccuracies of cost-effective Hall sensors can be calibrated without external reference encoders. A nonlinear system identification technique retrieves sensor inaccuracies by analyzing current and position data while the sensor is attached to a linear actuator. This approach provides an economical calibration solution at the cost of higher computational complexity offline, ensuring high precision with affordable hardware.

Experimental results verify the effectiveness of the developed calibration approaches. For SRMs, robust commutation functions designed using identified models reduce torque ripple and ensure consistent performance despite manufacturing variations, while the model-free method offers flexibility for large-scale applications. For piezo-steppers, data-driven control strategies extend product lifetime and enable task flexibility. In sensing, automated cascaded calibration ensures accurate results across multiple sensors, and the reference-free calibration approach achieves accurate position estimates without external encoders. Together, these contributions demonstrate the potential of data-driven methodologies to address the challenges of precision, scalability, and cost-effectiveness in mass production.