PaperlayerView exact revision on arXiv

Source-linked AI summary

Robust Bayesian compressive sensing for signals in structural health monitoring

Yong Huang, James L. Beck, Stephen Wu, Hui Li

arXiv:1412.4383v1stat.AP

TL;DR

Structural health monitoring produces large sensor datasets that motivate compression, while current Bayesian compressive sensing can yield large reconstruction errors at high compression. This paper improves BCS reconstruction and reports better performance on synthetic and bridge acceleration data.

  • Problem

    SHM produces substantial sensor data requiring compression, while current BCS can yield very large errors when measurements are far below the signal degrees of freedom.

    Source: §1 · p. 2

  • Method

    The paper uses stochastic optimization and successive relaxation to optimize sparse Bayesian learning hyperparameters and reduce suboptimal signal reconstructions.

    Source: §3.3 · p. 11, §5 · p. 16

  • Results

    The proposed BCS algorithm performs better overall than state-of-the-art BCS methods on synthetic and bridge acceleration data, with improved reconstruction robustness and uncertainty quantification.

    Source: §1 · p. 2, §5 · p. 17

  • Takeaways & Limitations

    Posterior error bars quantify reconstruction confidence, and the modified BCS method exhibits strong noise robustness.

    Source: §5 · p. 17

    The discussed sparse Bayesian pruning can delete basis vectors when individual iterated α_n values become numerically large, an imprecise way to induce sparsity.

    Source: §2.1 · p. 4

Abstract

from arXiv · show

In structural health monitoring (SHM) systems, massive amounts of data are often generated that need data compression techniques to reduce the cost of signal transfer and storage. Compressive sensing (CS) is a novel data acquisition method whereby the compression is done in a sensor simultaneously with the sampling. If the original sensed signal is sufficiently sparse in terms of some basis, the decompression can be done essentially perfectly up to some critical compression ratio. In this article, a Bayesian compressive sensing (BCS) method is investigated that uses sparse Bayesian learning to reconstruct signals from a compressive sensor. By explicitly quantifying the uncertainty in the reconstructed signal, the BCS technique exhibits an obvious benefit over existing regularized norm-minimization CS methods that provide a single signal estimate. However, current BCS algorithms suffer from a robustness problem: sometimes the reconstruction errors are very large when the number of measurements is a lot less than the number of signal degrees of freedom that are needed to capture the signal accurately in a directly sampled form. In this paper, we present improvements to the BCS reconstruction method to enhance its robustness so that even higher compression ratios can be used and we examine the tradeoff between efficiently compressing data and accurately decompressing it. Synthetic data and actual acceleration data collected from a bridge SHM system are used as examples. Compared with the state-of-the-art BCS reconstruction algorithms, the improved BCS algorithm demonstrates superior performance. With the same reconstruction error, the proposed BCS algorithm works with relatively large compression ratios and it can achieve perfect lossless compression performance with quite high compression ratios. Furthermore, the error bars for the signal reconstruction are also quantified effectively.

1 INTRODUCTION

Large sensor datasets in structural health monitoring motivate compression; this paper addresses robustness problems in Bayesian compressive sensing reconstruction.

1 INTRODUCTION

  • SHM systems generate large amounts of sensor data, making compression important for reducing signal-transfer and storage costs.The data volume is especially substantial in continuous monitoring of large structures such as long-span bridges.Source: §1 · p. 2
  • Compressive sensing projects N sampled signal points to K ≪ N measurements and can reconstruct accurately when signals are sparse in a suitable basis.Unlike conventional Nyquist-Shannon sampling, CS combines sampling and compression in a specially designed sensor.Source: §1 · p. 2
  • BCS uses posterior probability distributions over basis coefficients to quantify uncertainty in reconstructed signals, unlike methods that provide only point estimates.The introduction contrasts BCS with original CS reconstruction algorithms that estimate basis coefficients as point values.Source: §1 · p. 2
  • Current BCS algorithms can converge to suboptimal solutions with very large reconstruction errors when compression ratios are high.The paper motivates a more robust reconstruction algorithm that permits higher compression ratios.Source: §1 · p. 2
  • The improved BCS algorithm outperforms state-of-the-art BCS methods on synthetic and bridge-monitoring data in reconstruction robustness and uncertainty quantification.It is designed to reduce suboptimal reconstructions and produce smaller reconstruction errors more robustly.Source: §1 · p. 2

2 BAYESIAN COMPRESSIVE SENSING

This section frames Bayesian compressive sensing as sparse-signal recovery from compressed measurements with uncertainty estimates, examining its fragility under scarce or noisy measurements and how prediction-error variance relates to reconstruction error, model sparsity, and information extracted from measurements.

2 BAYESIAN COMPRESSIVE SENSING

  • The sensor takes K ≪ N random linear projections of a signal represented by sparse basis coefficients, leaving an ill-posed recovery problem.Source: §2 · p. 2

2.1 Sparse Bayesian learning of linear regression model for compressive sensing

  • Sparse Bayesian learning uses coefficient-wise Gaussian ARD priors and MAP hyperparameter selection to induce sparsity while yielding a Gaussian posterior for signal weights and reconstructed signals.The Bayesian formulation also estimates penalty parameters and quantifies reconstruction uncertainty.Source: §2.1 · p. 3
  • Top-down pruning can be slow and numerically imprecise, while Bottom-up adds relevant terms sequentially and is more efficient; local evidence maxima can trap both algorithms and impair robustness.Using the Woodbury identity reduces Top-down matrix inversion to O(K^3) when K < N.Source: §2.1 · pp. 4–5

2.3 The robustness problem of BCS reconstruction

  • Exact recovery required K > 140 (uniform)/100 (non-uniform) for Bottom-up and >130/120 for Top-down; lower K caused unstable local-maxima failures.The thresholds are measurements K for the two 20-spike signals of length N=512.Source: §2.3 · p. 5, Table 1 · p. 6
  • Below K=60/40 (uniform/non-uniform) for Bottom-up and K=50/40 for Top-down, evidence differences became negligible, making correct recovery by evidence maximization infeasible.Source: §2.3 · p. 6
  • With 5% measurement noise, reconstruction error increased, log evidence decreased for smaller optimal models, and most runs selected models larger than 20 terms.Source: Figure 3 · p. 7, §2.3 · pp. 7–8
  • The initial prediction-error variance can significantly affect the iterative reconstruction, and its value controls the trade-off between data fit and signal-model sparsity.Source: §2.3 · p. 8
  • As σ² increases, the posterior mean log-likelihood data-fit generally decreases and the KL information between converged posterior and prior PDFs shrinks, except for a significant data-fit increase at the optimal signal model.Larger σ² allows more data misfit to be treated as error, so less information is extracted from measurements.Source: §2.3 · p. 8
  • For uniform and non-uniform spikes, evidence-profile optima occur at σ²=0.64×var(y) and 0.04×var(y), respectively, and agree with the lowest average reconstruction error.The uniform-spike result is described as the optimal trade-off between data-fitting and model complexity; the reported measurement-noise standard deviation is RMS(y)×10^-5.Source: §2.3 · pp. 8–9, Figure 4 · p. 8
  • Away from the optimum, smaller σ² yields an under-sparse signal model, whereas larger σ² yields an over-sparse model; generally, larger σ² also produces fewer model terms.Figure 4 compares uniform and non-uniform spikes and notes different y-axis ranges for cases (iv) and (v).Source: §2.3 · pp. 8–9, Figure 4 · p. 8

3 PROPOSED BCS RECONSTRUCTION METHOD

The proposed BCS reconstruction method targets robustness by combining stochastic optimization over prior variances with successive relaxation over hyperparameters. It trades some computational efficiency for a greater chance of avoiding suboptimal local maxima.

3 PROPOSED BCS RECONSTRUCTION METHOD

  • The improved method combines stochastic optimization and successive relaxation to seek more robust BCS reconstruction with a reasonable computational-effort tradeoff.The stochastic procedure is less efficient than deterministic Bottom-up optimization.Source: §3 · p. 9, §3.3 · pp. 10–11

3.1 Stochastic optimization over the prior variances

  • Injected uncertainty broadens the search over prior variances and increases the probability of finding the global evidence maximum despite multiple local maxima.The deterministic search can be sensitive to measurement settings when measurements fall below a threshold.Source: §3.1 · p. 9

3.2 Optimization over all hyperparameters by successive relaxation

  • Successive relaxation alternates optimizing prior variances with fixed prediction-error variance and optimizing that variance with prior variances fixed, repeating until convergence.The algorithm initializes from a single selected basis term, then updates the prediction-error variance and prior variances in turn.Source: §3.2 · p. 10

3.3 BCS reconstruction algorithm using stochastic optimization and successive relaxation

  • The complete algorithm uses inner-loop stochastic optimization and outer-loop prediction-error-variance updates, while retaining the original BCS posterior mean and covariance updates.The outer loop stops when changes in reconstructed signal models are sufficiently small.Source: §3.3 · pp. 10–11

4 EXAMPLE RESULTS

Across the tested settings, BCS-SO* has the best overall reconstruction, sparsity, and uncertainty performance when measurements suffice. Results also show limitations for non-sparse bridge signals and a computational cost premium.

4.1 Synthetic sparse spike signals

  • At CR 5.69 for uniform spikes and 8.53 for non-uniform spikes, BCS-SO and BCS-SO* produce almost perfect reconstructions; suboptimal solutions can have underestimated error bars.The uncertainty underestimate occurs when the algorithm reaches a local rather than global evidence maximum.Source: Figure 5 · p. 11, §4.1 · p. 12
  • At 0.001% noise, BCS-SO and BCS-SO* outperform all other methods, especially at RE threshold 0.01; at 5% noise, none achieves RE<0.01 for CR>4.The study compares six algorithms across compression ratios and acceptable-error thresholds.Source: §4.1 · p. 12
  • For a given compression ratio and error threshold, acceptable-performance rates are higher for non-uniform than uniform spikes, including at higher noise levels.The authors relate this pattern to greater diversity in coefficient magnitudes.Source: §4.1 · p. 13
  • BCS-SO and BCS-SO* require about 6–18 times more computational effort than BCS-B-F, though reconstruction can keep up online below 350 Hz if good results are consistent.The online-operation statement is an inference for the tested PC.Source: §4.1 · p. 13

4.2 Application to bridge accelerometer data

  • For sparse Case 2, BCS-SO and BCS-SO* achieve RE<0.01 at CR<2.2; in Case 1, no algorithm meets the 0.01 or 0.1 error thresholds.Case 1 is the real, non-sparse signal; its small wavelet spikes act as modeling error for sparse reconstruction.Source: Figure 12 · p. 15, §4.2 · p. 16
  • The five BCS algorithms produce reconstructed coefficients sparser than BP and with sparsity comparable to or better than the de-noised signal.BCS automatically determines the signal model's effective dimensionality by retaining coefficients during Bayesian inference.Source: Figure 13 · p. 16, §4.2 · p. 16
  • For BCS-SO*, larger error bars track suboptimal reconstructions in Case 2 but not Case 1; BCS-SO's correspondence is less clear and can give false confidence.The Case 1 exception is associated with larger termination-criterion values used to avoid excessive iterations.Source: §4.2 · p. 16
  • When K is large enough for good reconstructions, BCS-SO* provides the best overall performance in reconstruction error, sparsity, and uncertainty quantification.The comparison covers all methods evaluated in the bridge-data examples.Source: §4.2 · p. 16

5. CONCLUDING REMARKS

The conclusion presents an improved Bayesian compressive sensing method for sparse-signal reconstruction, validated on synthetic signals and bridge acceleration data. It reports better performance than existing BCS algorithms and identifies further work on diagnosing and correcting suboptimal reconstructions.

5. CONCLUDING REMARKS

  • When measurements are much fewer than the discrete-time signal length, BCS reconstruction lacks robustness.The section identifies this as the motivation for improving the reconstruction method.Source: §5 · p. 16
  • Stochastic optimization and successive relaxation optimize sparse Bayesian learning hyperparameters to reduce the chance of suboptimal signal reconstructions.The method is evaluated using synthetic signals and real bridge SHM acceleration data.Source: §5 · pp. 16–17
  • Across synthetic signals and real bridge acceleration data, the proposed BCS algorithm outperforms state-of-the-art BCS methods in reconstruction robustness and uncertainty quantification.Figure 14 compares reconstruction error and average error bars for six algorithms on the first 50 real-data segments at K=200.Source: §5 · p. 17, Figure 14 · p. 17
  • Future work could use Bayesian statistics to diagnose suboptimal reconstructions and develop algorithms to “heal” them for SHM signals.Source: §5 · p. 17

APPENDIX A. PREDICTION ERROR VARIANCE OPTIMIZATION

Appendix A derives re-estimation equations for prediction error variance and hyperparameters in the sparse Bayesian learning optimization. It uses the log evidence and a Woodbury-identity simplification, and fixes a parameter whose optimization cannot converge to a finite value.

APPENDIX A. PREDICTION ERROR VARIANCE OPTIMIZATION

  • The appendix derives prediction-error-variance updates by differentiating the objective with respect to σ^2 and recommends applying updates (32) and (34) successively each iteration.The stated procedure iterates over all hyperparameters.Source: Appendix A · pp. 18–19
  • The Woodbury identity reduces the quadratic measurement term to σ^2 times the squared residual between measurements and their model prediction.The displayed derivation ends with σ^2‖y−Φμ‖^2.Source: Appendix A · p. 18
  • The log evidence objective is expressed using the determinant of the covariance and a quadratic measurement term, with a prior contribution involving hyperparameters.The derivation uses a determinant identity and the Woodbury inversion identity to simplify the objective.Source: Appendix A · p. 18
  • Because optimization of a cannot converge to a finite value, the authors fix it at 1 and give a numerical re-estimation equation for a.The appendix also presents re-estimation steps for b.Source: Appendix A · pp. 18–19
Loading 1412.4383v1…