Robust Optimal Experimental Design Accounting for Sensor Failure

1. Problem Statement

Optimal Experimental Design (OED) is a standard methodology for determining the most informative sensor locations and orientations for calibrating structural dynamics models (e.g., identifying unknown material properties or load forces). However, classical OED formulations assume that all sensors function perfectly. In real-world vibration analysis, sensors are prone to failure due to:

• Physical Detachment: High mechanical accelerations causing sensors to fall off ("dropout").
• Sensor Clipping: Vibrations exceeding the sensor's dynamic range, leading to data distortion or loss.

Existing robust OED methods often rely on reduced design spaces, surrogate models (e.g., Gaussian Processes), or evolutionary algorithms, making them computationally intractable for high-dimensional problems with expensive finite-element models (FEM). Furthermore, many existing approaches do not effectively handle complete binary sensor failure (a sensor works or it doesn't) in high-dimensional spaces without resorting to inefficient post-optimization rounding heuristics.

2. Methodology

The authors propose a framework for Robust OED that integrates sensor failure probabilities directly into the optimization process while maintaining computational efficiency for high-dimensional structural dynamics problems.

A. Mathematical Formulation

• Forward Model: The system is modeled using the discretized Helmholtz equation for linear, time-invariant structural vibrations. The relationship between unknown load parameters ($\theta$) and observable accelerations ($y$) is linear: $y = T\theta + \epsilon$, where $T$ is the frequency response function.
• Inverse Problem: The goal is to estimate $\theta$ using Ordinary Least Squares (OLS). The uncertainty is characterized by the parameter covariance matrix $C(w)$, where $w$ represents the sensor placement design.
• Classical vs. Robust Objective:
  • Classical: Minimizes $\log \det(C(w))$ (D-optimality) assuming all sensors work.
  • Robust: Minimizes the expected information loss accounting for failure. If sensor $i$ fails with probability $q_i$, the covariance is adjusted using a diagonal matrix $R$ where $R_{ii} = 1-q_i$. The objective becomes minimizing $\log \det(C(Rw))$.
  • Scenario Sampling: When failure probabilities are unknown, the authors sample $S$ specific failure scenarios (e.g., "Sensor $j$ fails") and minimize the average log-determinant over these scenarios.

B. Optimization Strategy: Relaxation and Penalty

To avoid the combinatorial explosion of binary optimization ($2^{N}$ possibilities), the authors employ a relaxation-based approach:

1. Fractional Weights: Instead of binary weights ($w_i \in \{0, 1\}$), the design variables are relaxed to continuous values ($w_i \in [0, 1]$). This allows the use of efficient gradient-based optimization.
2. Double-Well Penalty: To recover a binary solution without post-optimization rounding, a novel double-well penalty is added to the objective function:
   $$P(w) = \sum w_i(1 - w_i)$$
   This penalty is minimized when $w_i$ is close to 0 or 1, effectively inducing a sparse binary structure during the optimization process itself.
3. Hyperparameter Tuning: The penalty weight $\gamma$ is swept to find a solution that satisfies the budget constraint while minimizing the OED criterion and driving weights toward binary values.

C. Performance Metrics

The authors evaluate designs not just by the optimization criterion (log-determinant) but also via postprocessing metrics:

• Parameter Mean Squared Error (MSE): Accuracy of the estimated loads.
• Prediction MSE: Accuracy of the predicted structural response.
• Failure Scenario Analysis: Evaluating performance when specific optimal sensors fail, or when failure occurs according to sampled probabilities (Bernoulli trials) or clipping thresholds.

3. Key Contributions

1. General Robust OED Framework: Application of robust OED formulations to high-dimensional structural dynamics problems with expensive FEM models, handling both known failure probabilities and unknown failure scenarios.
2. Double-Well Penalty: Introduction of a binary-inducing penalty term within the OED optimization loop. This eliminates the need for heuristic rounding after optimization, allowing the algorithm to converge directly to binary sensor placements.
3. Comprehensive Evaluation: Moving beyond the optimization criterion to assess designs based on parameter/prediction MSE and their resilience across a wide distribution of failure scenarios (including worst-case and average-case performance).
4. Clipping Mitigation: A specific methodology to handle "sensor clipping" by generating dropout scenarios based on predicted acceleration thresholds and incorporating them into the robust design.

4. Computational Results

The methods were tested on a "Wedding Cake" finite element model (267 candidate sensor locations) with the goal of calibrating base loads.

• Scenario 1: Robustness to One Sensor Failure (Unknown Probabilities)

  • Fractional Designs: When binary constraints were relaxed, the robust design utilized more sensors (less sparse) than the classical design, intuitively spreading risk.
  • Binary Designs: Interestingly, when enforcing binary constraints, the classical and robust optimal locations were often identical. However, the robust design outperformed the classical design when one or two optimal sensors actually failed. The classical design's performance degraded significantly with multiple failures, whereas the robust design maintained stability.
  • Insight: Designing against a single sensor failure provided implicit robustness against multiple failures.

• Scenario 2: Known Probabilities of Failure (PoF)

  • Sensors at the top of the structure (higher acceleration) were assigned higher failure probabilities ($q=0.5$).
  • The robust design shifted sensor locations away from high-risk areas compared to the classical design.
  • Performance: While both designs performed similarly if no sensors failed, the robust design significantly outperformed the classical design in terms of average parameter and prediction MSE across $10^5$ simulated failure scenarios. It effectively avoided rare, catastrophic performance drops.

• Scenario 3: Sensor Clipping

  • Simulating clipping (data loss due to high acceleration) resulted in a robust design that avoided placing sensors on the top plate (highest acceleration).
  • The robust design showed superior performance stability across 100 random force realizations compared to the classical design, which was more susceptible to data loss in high-acceleration zones.

5. Significance and Conclusions

• Practicality: The proposed relaxation + double-well penalty approach makes robust OED computationally tractable for complex engineering problems where previous methods (evolutionary algorithms, surrogates) were too slow or inaccurate.
• Performance vs. Robustness Trade-off: The study demonstrates that while classical and robust designs often share similar optimal locations (suggesting D-optimality has inherent robustness), the robust designs provide a critical safety margin. They ensure that if a sensor fails, the experiment does not yield useless data, whereas classical designs may suffer significant information loss.
• Cost-Effectiveness: When failure probabilities are known, the robust formulation adds no additional computational cost compared to classical OED. When probabilities are unknown (scenario sampling), the cost increases linearly with the number of scenarios but remains feasible compared to the cost of physical experimentation.
• Future Work: The authors suggest extending these methods to multi-objective design problems and nonlinear models.

In summary, this work provides a rigorous, computationally efficient framework for designing sensor networks that are resilient to the inevitable failures encountered in high-acceleration structural testing, ensuring higher reliability in parameter estimation and model calibration.