Advances in Data Pre-Processing Methods for Distributed Fiber Optic Strain Sensing

Abstract

Because of their high spatial resolution over extended lengths, distributed fiber optic sensors (DFOS) enable us to monitor a wide range of structural effects and offer great potential for diverse structural health monitoring (SHM) applications. However, even under controlled conditions, the useful signal in distributed strain sensing (DSS) data can be concealed by different types of measurement principle-related disturbances: strain reading anomalies (SRAs), dropouts, and noise. These disturbances can render the extraction of information for SHM difficult or even impossible. Hence, cleaning the raw measurement data in a pre-processing stage is key for successful subsequent data evaluation and damage detection on engineering structures. To improve the capabilities of pre-processing procedures tailored to DSS data, characteristics and common remediation approaches for SRAs, dropouts, and noise are discussed. Four advanced pre-processing algorithms (geometric threshold method (GTM), outlier-specific correction procedure (OSCP), sliding modified z-score (SMZS), and the cluster filter) are presented. An artificial but realistic benchmark data set simulating different measurement scenarios is used to discuss the features of these algorithms. A flexible and modular pre-processing workflow is implemented and made available with the algorithms. Dedicated algorithms should be used to detect and remove SRAs. GTM, OSCP, and SMZS show promising results, and the sliding average is inappropriate for this purpose. The preservation of crack-induced strain peaks' tips is imperative for reliable crack monitoring.

Keywords: algorithm benchmarking; automation; data filtering; data pre-processing; data quality; distributed fiber optic sensing; software development; structural health monitoring.

Figure 1

Schematic of the different measurement disturbances: SRAs, dropouts, and noise. The variation in the signal marked as local effects is not caused by the measurement principle.

Figure 2

The task and workflow duality implemented by the pre-processing module of fosanalysis. (a) Class inheritance hierarchy of task objects. (b) Structure of a pre-processing workflow object. (c) Pre-processing workflow with an exemplary sequence of task objects.

Figure 3

GTM for one-dimensional case, as presented in [19]; additions are highlighted in grey.

Figure 4

OSCP, adapted from [34].

Figure 5

Algorithm of the z-score family.

Figure 6

Cluster filter, according to [42].

Figure 7

Benchmark data scenarios: (a) zero signal; (b) ramps; (c) strain profile with weak DFOS bond; (d) strain profile with medium DFOS bond; (e) strain profile with stiff DFOS bond. The data set is available in [60].

Figure 8

SRA detection accuracy for the algorithms for scenario (a) zero.

Figure 9

SRA detection accuracy for the algorithms for scenario (b) ramps.

Figure 10

SRA detection accuracy for the algorithms for scenario (c) weak bond.

Figure 11

SRA detection accuracy for the algorithms for scenario (d) normal bond.

Figure 12

SRA detection accuracy for the algorithms for scenario (e) stiff bond.

Figure 13

Results of filter benchmarks for case 1: noisy signal without SRAs or dropouts.

Figure 14

Results of filter benchmarks for case 2: noisy signal with SRAs and dropouts.

Figure 15

Section of strain data pre-processed with the filters: sliding average ($r=5$), sliding median ($r=5$), cluster filter ($\alpha =1.3\times {10}^5$). (a) Scenario (b), (b) scenario (d).

Figure 16

Left column: normalized strain peak with applied dropouts; right column: reconstructed strain peak with (i) $t=0.25$, (ii) $t=0.75$, and (iii) $t=1.25$. The right-hand-side plot details the highlighted part in the left-hand-side plot.

Figure 17

Error in crack widths $\mathsf{\Delta }{w}_{\mathrm{cr},\mathrm{rel}}$ for linear and Akima spline interpolation, depending on the amount of dropout t. The right-hand-side plot is the highlighted detail of the left plot.