Sensor-Redundant Control Hybrids

• Sensor-redundant control hybrids are architectures that integrate overlapping sensors with both continuous and discrete mechanisms to ensure robust, failure-tolerant operation.
• They employ methodologies like observer banks, Bayesian networks, and anomaly detection to isolate sensor faults and mitigate adversarial attacks in real time.
• Hybrid switching logic and supervisory automata underpin these systems, ensuring closed-loop stability and seamless adaptation in dynamic, multi-sensor environments.

Sensor-redundant control hybrids refer to a class of control and estimation architectures that integrate multiple, functionally overlapping sensors and exploit that redundancy through both continuous and discrete (hybrid) mechanisms to achieve robust, resilient, and failure-tolerant operation. Such systems are designed to maintain closed-loop stability and performance even under sensor noise, faults, or adversarial attacks, leveraging both analytic redundancy and supervisory logic at the control/estimation layer. Major methodological threads include observer banks with decoder-based attack isolation, ensemble learning with uncertainty-driven mode selection, switching among sensor-pinned controller realizations, and adaptive online fusion leveraging learned inter-sensor mappings.

1. Principles of Redundant Observability and Security

The foundation for sensor-redundant hybrid control lies in the formal notion of $q$-redundant observability for a discrete-time linear time-invariant (LTI) plant: $x(k+1) = Ax(k),\quad \bar{y}(k) = Cx(k) + a(k),$ where $a(k)\in\mathbb{R}^p$ is an additive, potentially adversarial attack vector. The essential property is that the system is $q$-redundant observable if for every subset $\Lambda\subset\{1,\ldots,p\}$ of cardinality $|\Lambda|\geq p-q$, the reduced output map $C_\Lambda$ (with rows in $\Lambda^c$ zeroed) remains observable. Key implications (Lee et al., 2018):

• $q=0$: reduction to standard observability.
• $2q$-redundant observability implies resilience to $x(k+1) = Ax(k),\quad \bar{y}(k) = Cx(k) + a(k),$0 arbitrary sensor losses for detection, $x(k+1) = Ax(k),\quad \bar{y}(k) = Cx(k) + a(k),$1 for correction.
• The dynamic security index $x(k+1) = Ax(k),\quad \bar{y}(k) = Cx(k) + a(k),$2 is defined as the minimal number of sensors whose simultaneous attack remains stealthy, established via the $x(k+1) = Ax(k),\quad \bar{y}(k) = Cx(k) + a(k),$3-stacked cospark of the observability matrix $x(k+1) = Ax(k),\quad \bar{y}(k) = Cx(k) + a(k),$4. For $x(k+1) = Ax(k),\quad \bar{y}(k) = Cx(k) + a(k),$5-redundant observability, $x(k+1) = Ax(k),\quad \bar{y}(k) = Cx(k) + a(k),$6.
• Equivalently, $x(k+1) = Ax(k),\quad \bar{y}(k) = Cx(k) + a(k),$7-attack detectability and $x(k+1) = Ax(k),\quad \bar{y}(k) = Cx(k) + a(k),$8-sparse observability formalize the system's ability to detect and uniquely reconstruct the state under sparse adversarial sensor manipulation.

Methods for verifying redundant observability include checking that for all reductions of the stacked observability matrix $x(k+1) = Ax(k),\quad \bar{y}(k) = Cx(k) + a(k),$9 by zeroing any $a(k)\in\mathbb{R}^p$0 blocks, the remaining block preserves full rank ($a(k)\in\mathbb{R}^p$1), i.e., $a(k)\in\mathbb{R}^p$2, $a(k)\in\mathbb{R}^p$3.

2. Architectures for Sensor-Redundant Hybrid Control

A variety of architectures instantiate sensor redundancy in hybrid control loops:

• Bank of Partial Observers + Redundant Decoders: Each sensor is associated with a partial observer built via Kalman decomposition, monitoring only the observable subspace with respect to that sensor. Stacked partial estimates are fused via a redundancy-exploiting decoder, typically using residual tests to isolate attacked sensors and reconstruct the state from healthy subsets (Lee et al., 2018).
• Ensemble Bayesian Networks: Multiple Bayesian neural networks (BNNs), each trained on a distinct sensor modality or channel, generate both control outputs and explicit uncertainty estimates. At runtime, the controller selects the BNN with minimal predicted uncertainty, realizing a mode-switching hybrid automaton (Lee et al., 2018).
• Dual-Estimator Networks with Online Anomaly Detection: Parallel estimators—e.g., one using vision+proprioception, another using proprioception only—are trained jointly. Online anomaly detectors monitor sensor health (e.g., via autoencoder residuals), and a confidence-driven mixing or hard switching between estimators provides seamless transitions under sensor failure (Zhang et al., 11 Sep 2025).
• Controller Copies Pinned to Sensor Subsets: Multiple controller realizations, each constructed around a disjoint group of sensors but dynamically equivalent in nominal operation, allow the use of majority vote or direct state comparison to detect and isolate faulty or attacked sensors, with hybrid switching logic selecting the healthy realization (Huisman et al., 8 Apr 2025).

These architectures may include explicit hybrid automata, with discrete state transductions (mode switches) triggered by residuals, majority voting, uncertainty thresholds, or anomaly detection.

3. Fault, Attack, and Anomaly Detection Mechanisms

Robust sensor-redundant hybrids require mechanisms for identifying faulty or attacked sensors in real time. Strategies include:

• Residual Testing and Decoding: After stacking partial state estimates, projection onto the nullspace of the fusion matrix $a(k)\in\mathbb{R}^p$4 yields a residual signal, which, under $a(k)\in\mathbb{R}^p$5-error detectability, serves as a basis for isolation—blocks with residuals exceeding a threshold are flagged as suspect (Lee et al., 2018).
• Majority Voting and Signal Comparison: For systems with triplicated or redundant measurements, majority voting on core signals (e.g., distances, speeds) resolves the most likely correct value, while direct comparison of control outputs across sensor-combination–pinned controller realizations exposes anomalies (Huisman et al., 8 Apr 2025).
• Uncertainty-Driven Selection: In Bayesian ensemble approaches, each network outputs both a predicted value and associated variance. The system dynamically selects the mode with minimal predictive variance, implicitly discarding unreliable sensors (Lee et al., 2018).
• Anomaly Detection Autoencoders: CNN autoencoders reconstruct recent sensory input; large reconstruction error (exceeding a threshold $a(k)\in\mathbb{R}^p$6) signals sensor failure, prompting confidence weighting or switching (Zhang et al., 11 Sep 2025).
• Statistical Outlier Detection: In musculoskeletal humanoids, Mahalanobis distance between predicted and observed intersensory signals (e.g., tension and length) over sliding windows is used to trigger hypothesis tests for rupture/failure detection (Kawaharazuka et al., 2024).

Table: Major Detection Mechanisms and Associated Architectures

Detection Approach	Core Mechanism	Example Reference
Residual/nullspace test	Decoder residual thresholding	(Lee et al., 2018)
Majority voting	Measurement triplication	(Huisman et al., 8 Apr 2025)
Uncertainty-driven switch	Bayesian variance minimization	(Lee et al., 2018)
Anomaly autoencoding	Depth frame reconstruction error	(Zhang et al., 11 Sep 2025)
Statistical outlier test	Mahalanobis distance (muscle)	(Kawaharazuka et al., 2024)

4. Hybrid Switching Logic and Stability Guarantees

Embedding these detection mechanisms in a hybrid control framework is essential for ensuring closed-loop performance:

• Supervisory Automata: Continuous plant dynamics are augmented with discrete states encoding the current sensor subset or estimator in use, with transitions governed by fault/isolation outcomes. Dwell-time or hysteresis on switching prevents chattering (Lee et al., 2018).
• Controller State Reset: On transitioning to a new healthy realization, internal controller/observer states are reinitialized (or “reset”) to avoid transient mismatch or integration error (Huisman et al., 8 Apr 2025).
• Stability Analysis: Common Lyapunov functions, often quadratic, can be constructed for all modes (sensor subsets) due to their shared plant structure and gain matrices. This ensures global exponential convergence under arbitrary admissible switching (Lee et al., 2018, Huisman et al., 8 Apr 2025). In deep policy-based cases, boundedness of the estimation error and stability under hybrid blending or soft-switching is supported by empirical results and stability arguments based on small-gain reasoning (Zhang et al., 11 Sep 2025).
• Seamless Degradation: Architectures trained jointly with both (or all) sensors learn to provide smooth transitions between estimation modes; abrupt performance drops are avoided under the designed fusion/switching policy (Zhang et al., 11 Sep 2025).

5. Practical Applications and Case Studies

Sensor-redundant control hybrids have been applied across domains:

• Three-Inertia System: Demonstrated as a canonical example for observer banks and residual-based attack correction; full reconstruction achieved despite a large injected sensor bias, with the controller maintaining reference tracking (Lee et al., 2018).
• Cooperative Adaptive Cruise Control (CACC): Multiple realization controllers, each using different sensor triplets, efficiently detect and isolate false data injection (FDI) attacks, with near-zero detection latency and no loss of tracking accuracy, as validated in switching-attack simulations (Huisman et al., 8 Apr 2025).
• Quadruped Locomotion under Visual Collapse: RENet’s dual-estimator network automatically falls back to a proprioception-only estimator when vision degrades, achieving near-oracle performance and stability in real-world long-distance and high-noise outdoor tests (Zhang et al., 11 Sep 2025).
• Musculoskeletal Humanoids under Muscle Rupture: Online-learning intersensory networks detect ruptures and modulate torque commands via masked correction, allowing continued motion even with actuator loss (Kawaharazuka et al., 2024).
• Autonomous Driving with Redundant Perceptual Inputs: Ensemble BNNs using GPS and stereo cameras, switching by uncertainty, achieve zero-control failures on aggressive racing tracks in the presence of severe sensor failures, whereas single-modality controllers fail catastrophically (Lee et al., 2018).

6. Computational Complexity and Scalability

Sensor-redundant architectures entail significant efficiency improvements over naïve combinatorial observers:

• Observer Complexity: Instead of running $a(k)\in\mathbb{R}^p$7 full-order observers for resilience to $a(k)\in\mathbb{R}^p$8 sensor faults, bank-of-partial-observer schemes require $a(k)\in\mathbb{R}^p$9 low-order observers and a decoder, reducing both memory and computation by orders of magnitude for $q$0 (Lee et al., 2018).
• Switching/Fusion Overhead: In ensemble deep architectures, runtime cost arises from parallel network inference and uncertainty sampling; e.g., 3 BNNs × 10 Monte Carlo samples × 20 Hz (Lee et al., 2018). Event-driven hybrid logic incurs minimal additional latency.
• Scalability: Scaling to more modalities or sensors is principled: new observer/estimator modules are added alongside expanded fusion logic; majority-voting and decoder-based algorithms generalize to larger $q$1 with computational feasibility preserved (Lee et al., 2018, Lee et al., 2018). Adaptive-hybrid fusion approaches, as in RENet, operate efficiently due to single-stage network training with joint estimator policies (Zhang et al., 11 Sep 2025).

7. Future Directions and Limitations

Contemporary sensor-redundant control hybrids face open challenges and present avenues for innovation:

• Chattering in Mode Switching: Some architectures report potential for high-frequency mode switches under repeated transient sensor faults; smoothing or filtered switching logic is required (Lee et al., 2018, Zhang et al., 11 Sep 2025).
• Generalization to Heterogeneous Sensing: While linear/observer-bank–based schemes are well understood for structured sensing, deep ensemble and learned-joint mapping approaches enable extension to modalities with complex or nonlinear observation functions (Kawaharazuka et al., 2024, Zhang et al., 11 Sep 2025).
• Online Adaptation: Adaptive schemes that mask failed sensors and adjust both policy/gain and estimation network weights online enable graceful degradation and continuity post-failure (Kawaharazuka et al., 2024).
• Computational Overhead: Parallel estimator policies and real-time uncertainty/fault analysis impose increased requirements for embedded control hardware, though continued optimization and hardware acceleration alleviate such concerns, especially for perception-driven robotics (Lee et al., 2018, Zhang et al., 11 Sep 2025).

Sensor-redundant hybrid control thus synthesizes analytic redundancy, adaptive learning, and discrete-mode supervisory logic to provide robust, fault-resilient, and attack-tolerant operation over a wide variety of engineering platforms, with a rich methodology comprising rigorous algebraic guarantees, online adaptation, and scalable architectural constructs (Lee et al., 2018, Huisman et al., 8 Apr 2025, Kawaharazuka et al., 2024, Lee et al., 2018, Zhang et al., 11 Sep 2025).