Towards Autonomous Commissioning of Industrial Drives via Multi-Objective Bayesian Optimization

Technical Summary: Towards Autonomous Commissioning of Industrial Drives via Multi-Objective Bayesian Optimization

Problem Statement
The commissioning of industrial electric drives remains a critical bottleneck in automation workflows, heavily reliant on manual, iterative tuning of cascaded control loops (current, velocity, position). This process requires significant expert knowledge, is difficult to standardize, and often yields suboptimal or oscillatory responses in complex systems. While data-driven and optimization-based methods have been proposed, many existing approaches rely on simulations, simplified models, or digital twins. Direct validation on real industrial hardware under practical constraints—such as discrete parameter spaces, noisy measurements, communication latency, and strict limits on the number of physical trials—remains limited. Furthermore, traditional Gaussian Process (GP)-based Bayesian Optimization (BO) often struggles in this setting due to the discrete nature of parameters and the high cost of evaluations.

Methodology
The authors propose a fully automated, model-free framework for tuning the current control loop of industrial drives directly on real hardware. The approach treats the drive as a black-box system, requiring no system identification, plant models, or firmware modifications. Interaction is achieved via a standard communication interface (Modbus), with controller parameters updated remotely.

The core technical contributions and methodological steps include:

1. Multi-Objective Formulation: Instead of scalarizing objectives, the tuning problem is formulated as a multi-objective optimization task minimizing four complementary metrics simultaneously:

   • IAE (Integral of Absolute Error): Average tracking accuracy.
   • ITAE (Integral of Time-weighted Absolute Error): Penalizes errors persisting later in the response.
   • OS (Overshoot): Maximum directional deviation during transitions.
   • OSC (Oscillation): Normalized RMS of residual ripple during steady-state segments.
     The goal is to approximate the Pareto-optimal set of controller configurations $(K_p, K_i)$ representing trade-offs between these metrics.

2. Algorithm Selection (Multivariate TPE): The authors select the Tree-structured Parzen Estimator (TPE) as the underlying BO strategy, specifically its multivariate extension. This choice is driven by the need to handle:

   • Discrete Parameters: Industrial hardware often requires integer-valued gains.
   • Noisy Evaluations: Physical experiments introduce measurement noise.
   • Parameter Coupling: The multivariate variant jointly models the distribution of $K_p$ and $K_i$, capturing their interdependencies, unlike standard TPE which models parameters independently.
   • Limited Budget: TPE avoids the computational overhead and smoothness assumptions of GP-based BO, offering faster convergence in low-budget regimes (tens of trials).

3. Experimental Loop: The optimization runs on an external PC. Each trial involves updating parameters, applying a predefined rectangular pulse reference signal (60% of rated peak current), acquiring data via the drive's internal oscilloscope, and computing metrics offline.

4. Post-Selection Strategy: Since the output is a Pareto front, a final controller is selected via a two-step process:

   • Hard Constraints: Filtering out configurations violating specific overshoot or oscillation limits.
   • Weighted Distance: Selecting the candidate closest to an ideal point using a weighted $\ell_2$ distance. Predefined strategies (Balanced, Fast, Smooth) allow users to prioritize specific metrics without re-running optimization.

Key Results
Experimental validation was conducted on a real industrial motor drive (230 V, 1.8 A) under no-load conditions, comparing TPE-based BO against Random Search (RS) and GP-based BO across trial budgets of 15, 20, 30, 50, and 100.

• Convergence Speed: TPE-based BO demonstrated faster growth in normalized hypervolume compared to both GP-based BO and RS, particularly immediately after the initial random exploration phase. This indicates more efficient exploration of the parameter space under tight evaluation budgets.
• Pareto Front Quality: TPE consistently identified larger and richer Pareto fronts than RS and, in many cases, GP-based BO. This suggests TPE provides a better empirical approximation of the trade-off surface, offering more diverse high-quality options for deployment.
• Performance vs. Expert Tuning: The method achieved performance comparable to, and in some metrics superior to, expert manual tuning within a few minutes (e.g., ~30% improvement in IAE and 27% in ITAE at 30 trials compared to the expert baseline). The automated solution matched the expert's performance in tracking accuracy while reducing oscillatory behavior, though the expert baseline achieved slightly lower overshoot in the specific test case.
• Robustness: TPE-based BO showed significantly lower variability across independent seeds compared to RS. While GP-based BO offered high consistency in final metrics at higher budgets, TPE maintained robustness in both final solution quality and the consistency of the Pareto front size across runs.
• Computational Efficiency: TPE-based BO incurred negligible computational overhead compared to GP-based BO, which suffered from increased tuning times due to surrogate model fitting and acquisition optimization.

Significance and Claims
The paper claims to present a practical step toward fully autonomous commissioning of industrial drive systems. Its significance lies in:

• Real-World Applicability: The method operates under realistic industrial constraints (no models, discrete parameters, noisy data, limited trials) without requiring firmware changes.
• Algorithm Suitability: It establishes that for this specific class of industrial problems, multivariate TPE is better aligned with the constraints than traditional GP-based BO, offering a superior balance of convergence speed, Pareto front richness, and computational efficiency.
• Autonomy: The framework demonstrates that high-quality controller configurations can be identified autonomously in minutes, reducing reliance on expert heuristics and standardizing the commissioning process.

The authors remain modest in their claims, noting that the validation is currently restricted to the no-load current loop of a single drive setup. They acknowledge that broader validation under load and across different drive configurations is necessary for future work. They do not claim uniform superiority of TPE over GP on every single metric but argue for its overall suitability in this specific multi-objective, constrained industrial setting.