Page 94 - Read Online

Page 400 Wang et al. Intell Robot 2022;2(4):391-406 https://dx.doi.org/10.20517/ir.2022.25

Table 2. Multisource heterogeneous data fusion representation methods
Research Method/approach Advantage Drawback
Li et al. [18] Bayesian estimation Simple operation Reliance on the distribution of the subjective
prior probability of data
[52]
Zhao et al.
DS evidence theory Applies reasoning to address uncertain Exponential explosion risk
[53]
Guo et al. information
[54]
Zhang et al.
Kalman filter Sufficient processing of data subject to Inadequate processing of non-Gaussian data
Palanisamy et al. [55] Gaussian distribution
[56]
Li et al.
Li et al. [57] Deep learning Strong expression ability Long calculation
time
[58]
Chen et al.
Yang et al. [59]
[60]
Li et al. Granular computing Sufficient processing of multi-granularity
spatiotemporal data
models and deep network models.

The shallow network model adopts a neural network with one single hidden layer and mines the
characteristic information in the nonlinear data to realize the intelligent identification and diagnosis of
scientific problems. Classic shallow diagnosis algorithms mainly include extreme learning machines, BP
neural networks, and SVMs. Wang et al. proposed a multiview multitask crack detector to calculate various
visual features (such as texture and edge) of the image area, suppress various background noise, and
emphasize the separability of crack region features and complex background features . The experimental
[61]
results with 350 crack images showed that the proposed method improves the training efficiency with a
precision of 92.3% and a recall of 89.7%. Yan et al. used a simple supported beam with a single crack and
double cracks as an example to identify local cracks and proposed a damage identification method for beam
[62]
structures based on a BP neural network and SVM . The results showed that the strain mode difference
curve at the damaged part undergoes considerable changes, and better identification accuracy is obtained,
where the recognition efficiency for a single crack is 99.9% and 100% for a double crack with the BP neural
network and 99.6% for a single crack and 99.9% for a double crack with the SVM. Liu et al. selected the
cantilever beam as the research object and proposed a crack damage detection method with a BP neural
network based on the curvature modulus ratio, the natural frequency, which uses the parameters of
frequency relative attenuation and the first-order maximum curvature modulus ratio as the input, and the
parameters of crack position and damage degree as the output and achieves a good crack identification
effect in terms of a relative error of 1.7 on crack images with a size of 248.3 mm and an injury degree of
[63]
84.6% . The shallow intelligent diagnosis algorithm has a hidden layer, which can mine complex feature
information and be applied for intelligent diagnosis in many fields. However, when the data present massive
growth and have multilevel high-dimensional features, the shallow intelligent diagnosis method is limited.

The broad learning system (BLS) method [64,65] is a kind of neural network that does not depend on the
structure depth and can realize the diagnosis of research problems by widening the width of the network for
information mining. Guo et al. adopted a network structure that combines deep learning and BLS on 40,000
concrete crack images, intelligently trains the network with the original image via linear and nonlinear
mapping processes with dynamic updating of the weights, and performs binary classification of concrete
surface cracks . The results proved that the accuracy of the presented method achieves 98.55% and 96.12%
[66]
and that the training time is 59.8 s and 95.76 s with two different datasets. Chen et al. and Xu et al. proposed
a recursive BLS, which uses recursive connections at the enhancement nodes of the network to capture the
dynamic characteristics of time series and shows excellent performance on the chaotic time series [67,68] . Their