Stability, robustness, and systematic parameter selection for the intelligent S-plane controller for micro unmanned aerial vehicles

The intelligent S-plane controller (Chen, Han, Zhang, and Jiang, IEEE Access 2018) drives micro unmanned aerial vehicle (MUAV) motion with a bounded sigmoid feedback law $u = 2/(1+\exp(-k_1 e - k_2 \dot e)) - 1 + \Delta u$ whose gains $k_1, k_2$ are adapted online by an expert parameter-adjustment rule table (Table 1 of the source paper): at every control step the measured error and error rate, partitioned by thresholds $\lambda_1, \lambda_2$, select the adjustment increments $K_1, K_2$ that are added to the gains. The source decouples the six-degree-of-freedom vehicle motion and closes one such loop per degree of freedom. The method was validated only empirically — semi-physical simulation and flight tests against the non-adaptive S-plane controller — leaving four linked gaps: no systematic procedure for choosing the adaptation parameters; no proof of closed-loop stability or quantified robustness margins for the rule-based gain updates under bounded disturbances and measurement noise; no formal study of how the rule set transfers across MUAV platforms, payloads, and maneuvers; and no comprehensive benchmarking against other adaptive controllers. Later work by the same and adjacent groups (H-infinity/S-plane hybrids, neural-network tuning, fractional-order variants, Q-learning tuning, model-reference adaptive S-plane control of surface vessels) consistently replaces or wraps the rule-based adaptation rather than analyzing it, and the 2023 successor papers still explicitly defer Lyapunov stability analysis to future work. The problem is to give this controller a theoretical and methodological foundation: formal stability and robustness guarantees for the adaptive law as published, a principled parameter-selection methodology, and cross-platform and comparative validation.

Background

The S-plane (also called S-surface) controller is a model-light feedback law that originated in autonomous underwater vehicle (AUV) motion control around 2001 (Liu and Xu, The Ocean Engineering) and was developed by the Harbin Engineering University underwater-vehicle community through tuned and hybridized variants (particle-swarm-optimized tuning by Tang et al. 2008, neural training by Li et al. 2009, fuzzy variants by Ma et al. 2011, and expert-controlled variants by Jiang et al. 2015). For one control channel it maps the normalized tracking error $e$ and its rate $\dot e$ to a normalized output through a sigmoid $u = 2/(1+\exp(-k_1 e - k_2 \dot e)) - 1 + \Delta u,$ where $k_1$ and $k_2$ are scalar gains, $\Delta u$ is an additive compensation term for environmental interference, and the actuator command is $U = K\,u$ with output gain $K$. The sigmoid bounds $u$ to $[-1,1]$, so the law behaves like a saturated nonlinear PD controller; its appeal for small vehicles is a simple structure, few parameters, and no need for an accurate dynamics model.

Micro unmanned aerial vehicles (MUAVs) fly a strongly nonlinear, tightly coupled six-degree-of-freedom regime in which precise models are hard to obtain, environmental disturbances are significant, and classical linear controllers (such as gain-scheduled PID) require frequent manual retuning; neural-network, fuzzy, and robust alternatives each carry costs in complexity, stability confidence, or portability across platforms. Chen, Han, Zhang, and Jiang (IEEE Access, 2018), building on earlier underwater-vehicle work by the same first author (Li, Jiang, Ma, and Chen, 2014), ported the S-plane controller to MUAV motion control and added an 'intelligent' adaptation layer: an initialization module, a knowledge base of expert tuning rules, and an inference engine that observes $(e,\dot e)$ online and updates the gains at every control step by $k_1(t+1) = k_1(t) + K_1, \qquad k_2(t+1) = k_2(t) + K_2,$ where the adjustment factors $K_1, K_2$ are discrete increments built from $\delta_1, \delta_2$ and are read off the parameter-adjustment rule table of the paper (Table 1), which partitions the $(e,\dot e)$ plane by thresholds $\lambda_1, \lambda_2$; the source reports $\delta_1 = \delta_2 = 0.1$ generally and $k_1(0) = k_2(0) = 3.0$, and derives the table from four experiential manual-tuning rules (raising $k_1, k_2$ increases sensitivity but causes strong overshoot and oscillation; too-small gains weaken overshoot but make the controller unresponsive; the initial gains are taken as 3.0; if overshoot is too large $k_1$ is gradually reduced while $k_2$ is increased, and the converse if convergence is too slow). Following the source's decoupling of the MUAV motion, one loop of this form is designed per degree of freedom. The controller ran on a PC/104 embedded single-board computer under the VxWorks real-time operating system and was validated in semi-physical simulation (the actual embedded control code exchanging sensor and actuator packets with a vehicle simulator) and in flight tests, improving height, speed, and heading accuracy relative to the non-adaptive S-plane controller.

Everything in this validation is empirical. The rule-table parameters are chosen from tuning experience; no theorem guarantees that the closed loop — a saturated sigmoid law whose gains are switched at every step by the very error signals that are corrupted by disturbances and measurement noise — remains stable, and no quantified disturbance-rejection margins exist. The evidence covers one self-built platform, a limited maneuver set (level acceleration, climb, circular flight), and comparison only against the non-adaptive S-plane baseline.

Several further features of the published object are specified only partially in the source, and this is part of what an analysis has to face. The cell-by-cell content of Table 1 appears in the published paper as a table image, so the surrounding text supplies the parameter structure ($\lambda_1, \lambda_2$ as parameter-adjustment thresholds set according to $e$ and $\dot e$; $K_1, K_2$ as the adjustment factors of $k_1, k_2$; $\delta_1 = \delta_2 = 0.1$ generally; the gains "can be adjusted adaptively using values in TABLE I") but not the mapping in words, and no successor paper reproduces it. The compensation term $\Delta u$ is described only verbally, as the adjustment parameter that eliminates environmental interference; no formula, bound, or measurement source is given, and the family has no fixed convention (the fractional-order successor of Li et al. 2023 sets $\Delta u = 0$, the neural successor of Chen et al. 2021 introduces its own compensation module, and the earlier expert S-plane variant of Jiang et al. 2015 uses an integral term $\Delta u = k_3\int e\,dt$). The source gives no sampling period, no hold model, and no normalization scales for $e$ and $\dot e$, although the increments are applied at every control step of the embedded implementation; and it prints no plant equations, referring instead to a MUAV model taken from the flight-control literature (its ref. [20]).

Problem Statement

The source paper leaves the intelligent S-plane controller without a theoretical or systematic-methodological foundation. The problem is to supply it, in four linked parts whose common objective is to turn an empirically validated, experience-tuned rule-based adaptive controller into a sound and transferable design method:

(1) Formal stability and robustness analysis. For the intelligent S-plane closed loop as published — the sigmoid law $u = 2/(1+\exp(-k_1 e - k_2 \dot e)) - 1 + \Delta u$, $U = K\,u$, coupled with the discrete rule-based gain adaptation $k_1(t+1)=k_1(t)+K_1$, $k_2(t+1)=k_2(t)+K_2$ whose increments are read off the Table 1 rule mapping, one loop per degree of freedom as in the source — under MUAV motion models subject to bounded environmental disturbances and measurement noise, either prove closed-loop stability in an appropriate precise sense (e.g., Lyapunov stability, input-to-state stability, or uniform ultimate boundedness of the tracking errors) together with quantified robustness margins, or refute the implicit stability claim by exhibiting a plant, parameter set, and disturbance/noise realization for which the published rule-driven adaptation destabilizes the loop (for example unbounded growth of the adapted gains or chattering between adjacent rule cells). Since the source fixes neither the plant equations, nor the disturbance and noise bounds, nor the sampling and normalization conventions, nor a formula for $\Delta u$, whichever branch is pursued must declare the admissible plant class, the disturbance/noise assumptions, the sampled-data conventions, and the $\Delta u$ convention it treats, and the conclusion must hold for that declared class; that declared class must be nondegenerate and must leave the question nontrivial — it must be a stated canonical MUAV model class representative of the source's platform, its disturbance and noise bounds must be nonzero, and the constant-gain (non-adaptive) S-plane loop must be stable on it, so that any destabilization found is attributable to the published Table 1 rule-driven adaptation rather than to the chosen plant; in the refutation branch the adaptation must be the published Table 1 mapping with its reported parameter values, not a rule table chosen by the analyst.

(2) Systematic parameter selection. Provide a principled procedure, with design guidelines and sensitivity analysis, for choosing the thresholds $\lambda_1,\lambda_2$, the increments $\delta_1,\delta_2$, the initial gains, and the rule-table mapping for new MUAV platforms and different flight regimes, replacing selection by experience.

(3) Generalization and scalability. Quantify, in a formal study, how the rule set and update law perform across different MUAV configurations, payloads, and more extreme maneuvers than the tested scenarios.

(4) Comparative benchmarking. Provide quantitative comparisons against other adaptive control approaches — for example adaptive neural methods, robust controllers, or online-optimization schemes — over a wider set of operating conditions than the single non-adaptive baseline used in the source paper.

A complete answer must address the stability and selection questions for the controller as published — the Table 1 rule-based adaptation with its stated parameter structure, applied per degree of freedom as in the source — and not for a simplified or restricted substitute (e.g., frozen gains or an imposed dwell time between rule firings) unless such a restriction is itself derived and justified as part of the analysis. Parts (1) and (2) are the theoretical and methodological core of the problem; parts (3) and (4) are the validation objectives that this core is meant to support, and no particular experimental infrastructure or protocol is prescribed for them. A complete answer requires parts (1) and (2); parts (3) and (4) are optional objectives that strengthen the core but are not required to resolve the task.

References

1. Chen P., Han Y., Zhang P., Jiang Y. (2018). Intelligent S-Plane Controller for Micro Unmanned Aerial Vehicle. IEEE Access 6:68096–68103. DOI 10.1109/ACCESS.2018.2879063. https://doi.org/10.1109/access.2018.2879063
2. Liu X. M., Xu Y. R. (2001). S control of automatic underwater vehicles. The Ocean Engineering (海洋工程) 19(3):81–84. No DOI (Chinese journal).
3. Tang X., Pang Y., Li Y., Wan L. (2008). Improved PSO-based S Plane Controller for motion control of underwater vehicle. 2008 27th Chinese Control Conference. DOI 10.1109/CHICC.2008.4604970. https://doi.org/10.1109/chicc.2008.4604970
4. Li Y., Tang X., Pang Y. (2009). S Plane Control of underwater vehicle and its hybrid training algorithm. 2009 Chinese Control and Decision Conference. DOI 10.1109/CCDC.2009.5192715. https://doi.org/10.1109/ccdc.2009.5192715
5. Ma S., Pang Y., Zhang T., Lv C. (2011). Fuzzy S plane controller for motion control of underwater vehicles. 2011 6th IEEE Conference on Industrial Electronics and Applications. DOI 10.1109/ICIEA.2011.5975853. https://doi.org/10.1109/iciea.2011.5975853
6. Jiang C., Wan L., Sun Y. (2015). Expert S-plane Control of Underwater Vehicle. Proceedings of the 2015 International Conference on Automation, Mechanical Control and Computational Engineering, Advances in Intelligent Systems Research. DOI 10.2991/amcce-15.2015.341. https://doi.org/10.2991/amcce-15.2015.341
7. Li Y., Jiang Y., Ma S., Chen P., Li Y. (2014). Inverse speed analysis and low speed control of underwater vehicle. Journal of Central South University 21(7):2652–2659. DOI 10.1007/s11771-014-2226-7. https://doi.org/10.1007/s11771-014-2226-7
8. Zhao X., Yuan M., Cheng P., Xin L., Yao L. (2019). Robust H∞/S-plane Controller of Longitudinal Control for UAVs. IEEE Access 7:91367–91374. DOI 10.1109/ACCESS.2019.2927000. https://doi.org/10.1109/access.2019.2927000
9. Chen P., Zhang G., Guan T., Yuan M., Shen J. (2021). The Motion Controller Based on Neural Network S-Plane Model for Fixed-Wing UAVs. IEEE Access 9:93927–93936. DOI 10.1109/ACCESS.2021.3093768. https://doi.org/10.1109/access.2021.3093768
10. Wu G., Luo W., Guo J., Zhang J. (2022). A Sigmoid-plane adaptive control algorithm for unmanned surface vessel considering marine environment interference. Transactions of the Institute of Measurement and Control 44(10):2076–2090. DOI 10.1177/01423312211072658. https://doi.org/10.1177/01423312211072658
11. Li J., Chen P., Chang Z., Zhang G., Guo L., Zhao C. (2023). Trajectory Tracking Control of Quadrotor Based on Fractional-Order S-Plane Model. Machines 11(7):672. DOI 10.3390/machines11070672. https://doi.org/10.3390/machines11070672
12. Chen P., Zhang G., Li J., Chang Z., Yan Q. (2023). Path-Following Control of Small Fixed-Wing UAVs under Wind Disturbance. Drones 7(4):253. DOI 10.3390/drones7040253. https://doi.org/10.3390/drones7040253
13. Zhang P., Zhu S., Xu H., Zhang X., Chen C., Zhang X. (2023). A Machine Learning-Based Parameter Adaptive S-Plane Controller for Motion Control of Unmanned Underwater Vehicles. 2023 China Automation Congress (CAC). DOI 10.1109/CAC59555.2023.10450718. https://doi.org/10.1109/cac59555.2023.10450718
14. Rairán Antolines J. D. (2022). Performance evaluation of a Time Scale Controller. DYNA 89(220):178–186. DOI 10.15446/DYNA.V89N220.96502. https://doi.org/10.15446/dyna.v89n220.96502