Page 92 - Read Online
P. 92
Page 398 Wang et al. Intell Robot 2022;2(4):391-406 https://dx.doi.org/10.20517/ir.2022.25
Table 1. Summarized approaches of feature extraction on bridge cracks
Research Method/approach Advantages Drawbacks
Yu et al. [31]
PCA Ability to remove noise and extract Unable to analyze the real subspace
[32]
Mei et al. effective features structure of data
[34]
Li et al.
Sparse decomposition Ability to separate the signal and noise Mass computation and computational
[35]
Wang et al. complexity
[37]
Nguyen et al.
Wavelet transform Good denoising effect Unable to identify the optimal wavelet
Nigam et al. [38] base
Bilir et al. [43] Adaptive neural fuzzy inference Self-learning ability and fuzzy logic
system reasoning ability
[15]
Guo et al.
Zheng et al. [46] Deep learning Strong expression ability Long calculation time
[47]
Xu et al.
Teng et al. [48]
[49]
Pan et al. Self-Attention Ability to quickly extract important
[50] features of data
Zhao et al.
redundant and related monitoring data of the fracture, the data fusion method can more comprehensively
and accurately evaluate the evolution state of the fracture. Classical information fusion methods include
Bayesian estimation, DS (Dempster-Shafer) evidence theory, and the Kalman filter.
Bayesian estimation is based on the prior probability and posterior probability criteria in probability theory
and uses conditional probability to represent the uncertain information in the monitoring data . Li et al.
[51]
employed a fully CNN and naive Bayesian data fusion model to automatically segment cracks and noise and
fused the extracted multilayer features to obtain significant crack recognition performance . The algorithm
[18]
was verified with 7200 datasets of bridge substructures collected from 20 in-service bridges under various
circumstances. The recognition results showed remarkable performance of the proposed algorithm
compared to other recent algorithms. Bayesian estimation must rely on the distribution of the subjective
prior probability of the data. However, some of the prior probability distribution of the monitoring data in
the actual project is unknown. Therefore, there is a contradiction between this subjectivity and scientific
objectivity.
DS evidence theory uses DS synthesis rules to fuse multiple pieces of evidence (feature information),
eliminate or reduce the complementarity, redundancy and uncertainty among data, and obtain the fusion
judgment or diagnosis of comprehensive features with certain decision rules. Zhao et al. applied DS
evidence theory to perform weighted fusion on structural health monitoring data from multiple sensors of a
two-story concrete frame to provide an accurate and final interpretation of the structural health status .
[52]
Guo et al. applied multiscale space theory and a data fusion method to detect the multiscale damage of
beams and plates in a noisy environment and applied DS evidence theory to fuse and express the multiscale
[53]
damage characteristics in a multiscale space to obtain good anti-noise ability and damage sensitivity . DS
evidence theory is mainly utilized to address the reasoning of uncertain information, but it lacks a certain
theoretical basis and has potential exponential explosion risk in calculation.
According to the statistical principle, the Kalman filter uses the statistical characteristics of monitoring data
and empirical data to perform real-time fusion representation of uncertain and dynamic redundant
monitoring data. Prof. Zhang et al. employed a Kalman filter to fuse the parameters of the residual
generator during the design of two-degrees-of-freedom controllers in a data-driven environment and a
residual generator to explain all the stability . Palanisamy et al. estimated a Kalman state structure model
[54]