Distributed optic fiber sensor (DOFS) is widely used for status monitoring and damage detection of aerospace vehicles due to its ability to achieve large-area and high-density sensing of structures. However, in the face of uncertainties caused by the harsh service environment of aerospace, a phenomenon of strain reading anomalies (SRAs) occurs in DOFS measurements. These SRAs result in significant strain peaks occurring in localized regions or at specific moments in time, thereby posing challenges for DOFS to accurately measure physical quantities and making it even more difficult to interpret these measurements. To minimize the negative effects of SRAs, some researchers have adopted a series of data processing methods, such as polynomial fitting method, spectral shift quality (SSQ) method, geometrical threshold method (GTM), and polynomial interpolation comparison method (PICM). Although these data processing methods are effective in reducing random errors in measurement data, they fail to completely remove the phenomenon of SRAs, and there is still a risk of removing highly reliable measurement readings. Meanwhile, the above methods still use the fixed threshold method to detect and determine the anomalies, and the determination of the fixed threshold relies on manual experience, which has low detection efficiency and a high false alarm rate, thus limiting its application in complex service environments. Therefore, we propose an intelligent adaptive post-processing method for detecting and quickly removing SRAs from DOFS.

The proposed algorithm, namely the adaptive geometrical threshold offset method (AGTOM), adopts the K-means clustering method to adaptively determine thresholds for distinguishing differences of thresholds caused by various structural features and service conditions. A continuous geometric correction is implemented on the distorted strain curves to effectively eliminate SRAs. To verify the effectiveness of the proposed method, a case study is conducted on the processing of DOFS measurement data collected during the pressure cycling test of a fuel tank. The Pearson correlation coefficient (PCC) is utilized to evaluate the correlation between the post-processing curves and normal strain curves. Besides, a comparison is conducted with other post-processing algorithms (GTM and PICM) to highlight the advantages of the proposed method.

Based on their different response characteristics, SRAs can be classified into two categories: harmless strain reading anomalies (HL-SRAs) and harmful strain reading anomalies (HF-SRAs). For the HL-SRAs, AGTOM consistently yields optimal post-processing results with PCC values not less than 0.965. It is followed by GTM, whose PCC values are all not less than 0.798. However, GTM interferes when HL-SRAs are coupled with NaN values. In addition, PICM achieves promising processing results only in the first typical case (i.e., sparsely distributed HL-SRAs). In the remaining three typical cases, PICM still produces distortions with a PCC value not greater than 0.512. Importantly, both GTM and PICM exhibit distorted post-processing curves when HL-SRAs are coupled with NaN values. For HF-SRAs, AGTOM also yields the highest post-processing results, with no PCC value lower than 0.917. The susceptibility of PICM to curve distortion accurately reflects the difference between HF-SRAs and HL-SRAs because the main characteristic of the former is that strain values follow an erroneous strain response or frequent sudden changes. It is difficult to determine the change in strain increment using PICM because it detects and removes SRAs by comparing the fitted value with the original value. Compared with PICM, GTM takes into account the sudden changes of the strain increment, resulting in improved post-processing results when HF-SRAs consist of densely changed SRAs. However, similar to HL-SRAs, both GTM and PICM show worsened post-processing results when NaN values interfere with HF-SRAs, indicating lower algorithmic robustness for GTM and PICM compared to AGTOM

The proposed algorithm AGTOM is able to distinguish the differences in thresholds due to different structural characteristics and service environments. The K-mean clustering algorithm uses an internal evaluation metric, namely Davies-Bouldin index (DBI), to characterize the clustering effect of strain increments. The threshold is determined by obtaining the optimal k value. For the HL-SRAs, both GTM and AGTOM methods can achieve satisfactory processing results. However, PICM is susceptible to interference when facing densely distributed and coupled HL-SRAs, leading to serious distortions in its post-processing curves. For HF-SRA, the post-processing curves of the other two algorithms are distorted to varying degrees, except for AGTOM, which exhibits the highest PCC compared to the normal strain curve. For both HL-SRAs and HF-SRAs, GTM and PICM are interfered with when SRAs are coupled with NaN, indicating that the algorithmic robustness of both is lower than that of AGTOM. To further validate the effectiveness of AGTOM, it will still be necessary to test AGTOM by applying it to different experimental scenarios in the future.