Neural Computing and Applications (2021) 33:239–255

https://doi.org/10.1007/s00521-020-04988-3(0123456789().,-volV)(0123456789().
,- volV)

ORIGINAL ARTICLE

Research on tunnel engineering monitoring technology based

on BPNN neural network and MARS machine learning regression
algorithm
Jianbo Fei1,2 • Zezhou Wu2 • Xiaohui Sun1,2 • Dong Su1,2 • Xiaohua Bao1,2

Received: 18 February 2020 / Accepted: 2 May 2020 / Published online: 16 May 2020
Ó Springer-Verlag London Ltd., part of Springer Nature 2020

Abstract
Tunnel engineering is affected by a variety of factors, which results in large detection errors in tunnel engineering. In order
to improve the monitoring effect of tunnel engineering, based on BPNN and MARS machine learning regression algorithm,
this research constructs a tunnel engineering monitoring and prediction model. Moreover, the gray residual BP neural
network designed in this study uses a series combination, and the residuals obtained from the gray model are used as the
input data of the BP neural network, and the output of the combined model is used as the prediction result. By applying the
monitoring data of the convergence of the upper surrounding of the tunnel surface section and deformation of the arch
subsidence, it is verified that the proposed method based on the combined model of BPNN and MASR can predict and
analyze the tunnel deformation monitoring data very well.

Keywords BPNN  MARS  Regression algorithm  Tunnel engineering  Intelligent monitoring

1 Introduction physical information (such as displacement, strain, stress or

load) measured on site and reflecting the mechanical
Underground rock and soil are formed through long-term behavior of the system to calculate each or some initial
geological evolution and have complex structures, and their parameters of the system (such as initial ground stress,
physical and mechanical characteristics are difficult to constitutive model parameters and geometric parameters)
describe accurately. When the engineering design of rock through the inversion model (the physical property model
and soil is required, the core problems of establishing of the system and its mathematical description, such as the
accurate constitutive model of rock and soil and initial relationship between stress and strain, etc.). The purpose is
values of performance parameters must be solved. How- to establish a theoretical prediction model that is close to
ever, geotechnical engineering generally does not have the the actual measured results in the field, so as to correctly
conditions for in situ testing of geotechnical parameters reflect certain mechanical behaviors of predicted geotech-
and in situ testing of initial geo-stress. Therefore, the nical structures [2].
parameters are usually obtained through laboratory simu- After the development of many scholars since the
lation experiments, and the results are usually not satis- 1970 s, the computational model of geotechnical engi-
factory [1]. Therefore, a back analysis method for neering back analysis method has evolved from linear to
geotechnical engineering was adopted. It is based on some nonlinear, from the original analytical method to the
numerical method, from homogeneous to non-homoge-
neous, and form a simple plane problem to a space problem
& Zezhou Wu that needs to consider more factors [3]. On the basis of
[email protected] deterministic back analysis, non-deterministic square
1 analysis is developed. Moreover, with the development of
Underground Polis Academy, Shenzhen University,
Shenzhen 518060, China artificial intelligence technology, the development of
2 geotechnical engineering back analysis has entered a new
College of Civil and Transportation Engineering, Shenzhen
University, Shenzhen 518060, China stage.

123
240 Neural Computing and Applications (2021) 33:239–255

With the continuous development of the national econ- cases. At the same time, traditional deformation prediction
omy, China’s infrastructure construction is also increasing methods such as numerical methods, empirical formula
year by year, and more and more tunnels will be con- methods, and numerical estimation methods have been
structed. With the expansion of the scale of tunnel con- derived from tunnel engineering [8]. However, the method
struction, tunnel engineering will encounter greater relying on traditional experience summarization is more
challenges in terms of construction length, difficulty and based on example estimation, does not form a systematic
geological conditions [4]. During the construction process, theory and is too ideal when considering external influence
risks occur between the connection of various processes. factors. Therefore, there is often a large difference between
Therefore, before the next process is carried out, the pos- the calculated result and the actual result. In fact, there are
sible risk points of the next process need to be analyzed in many factors affecting the deformation of tunnel engi-
time according to the construction site conditions, and neering and the project has a certain timeliness. Therefore,
protection works and monitoring preparations must be done it is difficult to make a reasonable analysis of each factor,
before construction to minimize the risk. and if it is required to be detailed to analyze every day, it
On-site deformation monitoring technology plays a will also lose its practical value in engineering.
strong role in the construction of the New Austrian Tunnel. In view of the limitations of traditional monitoring
Deformation monitoring must be performed throughout the methods, many scholars have done a lot of research and
entire tunnel construction process, and the quality of used field monitoring data to build a data model to predict
monitoring is an important basis for tunnel construction the settlement factors. Common models include wavelet
[5]. In order to ensure the safety and stability of the tunnel analysis models, regression analysis models and time series
during the construction and use of the tunnel, deformation models [9]. Wavelet analysis is a practical discipline
monitoring (mainly monitoring the subsidence of the arch developed rapidly based on the research results of Meyer,
and the surrounding convergence) is adopted during the S. Mallat, I. Daubechies [10] and others. It is widely used
construction and use of the tunnel to analyze the factors in the fields of detecting edge features of images and
affecting the deformation, determine the space–time char- determining the accuracy of gravity tide and sand param-
acteristics of the deformation and monitor the safety status. eters. Chenghui Wan [11] used wavelet analysis to process
Deformation monitoring is of great significance for veri- dam deformation monitoring data and realized analysis and
fying the design theory of the tunnel, grasping the defor- prediction of deformation. Hongyan Wen and others stud-
mation law and researching reasonable deformation ied the feasibility and superiority of wavelet noise reduc-
prediction theories and methods, which will improve the tion technology in wavelet analysis application and
safe service life of the tunnel and prevent and reduce dis- deformation analysis data processing. Dezhong Zheng [12]
asters [6]. Tunnel deformation monitoring not only guar- discussed the image denoising method based on node
antees the safety of tunnel construction, but also greatly threshold wavelet packet transform and used soft threshold
improves the economic efficiency of the tunnel. for threshold processing, and spectral entropy method for
noise estimation and verified the significant image
denoising performance. Xianjun Shi [13] and others pro-
2 Related work posed the optimal wavelet packet base denoising method
based on Shannon entropy criterion and achieved good
Foreign experts started earlier in the field of tunnel engi- denoising results. Junquan Yuan also introduced the con-
neering prediction. Terzaghi and Peek [7] and others first cept of average threshold in the process of wavelet packet
studied the tunnel and proposed theoretical methods for denoising and obtained a more significant denoising effect.
estimating the excavation and calculating the load. Eide After the gray system was proposed by Professor Julong
et al. put forward corresponding ideas and analyzed the Deng [14], it has been improved and applied in deforma-
principle and method of tunnel floor bulge. Since the early tion prediction. Among them, Qingguo Hu [15] discussed
1960s, tunnel engineering has introduced scientific instru- the equal-step and non-equal-step GM (1,1) prediction
ments for monitoring. Among them, the more influential models and verified the feasibility and reliability of the
examples are tunnel monitoring in Oslo. Since the 1970s, gray prediction model. Li Li [16] optimized the model
many countries have begun to develop guidelines for tun- from the direction of GM (1,1) original sequence, initial
nel excavation and construction. Since then, more and value, background value, etc., and applied the improved
more scholars have invested in research and made a series model in a tunnel project. Jinrong Yuan proposed that the
of breakthroughs. Since 1990, systematic monitoring of gray model has a certain scope of application when the
tunnel projects has begun in China. Over the years, with the gray model is used to predict the tunnel, and it is not
joint efforts of many researchers and constructors, they suitable for predicting the horizontal displacement of the
have also accumulated a lot of experience in engineering underground continuous wall. In addition, some scholars

123
Neural Computing and Applications (2021) 33:239–255 241

have adopted the iso-dimensional information iterative combination model, etc., which have also been widely
method GM (1,1) model in the tunnel example and proposed and applied.
obtained good accuracy.
In addition, a method based on the human brain neural
network and widely extended, the artificial neural network 3 MARS algorithm
method has been proposed and greatly developed in the
past 20 years. For example, He Keqiang established a 3.1 MARS model
neural network-based tunnel settlement prediction model.
BP neural network [17] (Back-Propagation Network) is a The MARS model can segment large-scale high-dimen-
kind of forward feedback network, which is also the most sional data and obtain excellent regression models, which
used and researched neural network and is especially used is an advantage that many other regression models do not
in tunnel displacement prediction. Goh [18] considered the have. The basis function of the MARS model is a truncated
BP neural network when predicting the maximum dis- spline function or the product of multiple spline functions.
placement of the tunnel retaining wall and achieved results. At the same time, MARS combines the advantages of the
Zheng Cao combined the highly nonlinear mapping capa- projection tracking method for high-dimensional data
bilities of BP neural network to propose a method for analysis and the advantages of recursive auto-regressive
predicting tunnel deformation time series based on BP method for dividing the spatial region and adaptive nodes.
neural network. Yan Li selected a tunnel project in If we set [22]:
Hangzhou as an example and used BP neural network to x ¼ x1 ; x2 ; x3 ; . . .; xp
get a good prediction in the deformation prediction of the  
tunnel envelope. Jiagang Zhou established a construction y ¼ f1 ð xÞ; f2 ð xÞ; f3 ð xÞ; . . .; fq ð xÞ þ e
control prediction model for cable-stayed bridges based on
as the output variable, the model can be established as:
BP neural network theory and verified the feasibility of  
applying BP neural network in construction control of y ¼ f1 ð xÞ; f2 ð xÞ; f3 ð xÞ; . . .; fq ð xÞ þ e:
cable-stayed bridges. In addition, there are researches on
In the formula: e is the error vector of ð1  qÞ, and the
the application of BP neural network in tunnel deformation
exact form of fj ð xÞ is unknown.
prediction analysis and safety assessment during operation.
The goal of regression analysis is to build a prediction
Combined forecasting analysis [19] is a comprehensive
forecasting method that has developed rapidly in recent function f^j ð xÞ on f ð xÞ based on the existing data set [23].
years. The concept of combination refers to the effective The MARS method is to perform regression on local data
combination of different forecasting models or different using a spline function (also known as the basis function
methods to synthesize the advantages and advantages of a Basis Functions, or BF for short), and to fit the relationship
single model or method for forecasting. Lin Gan chose the between the variables, especially the complex nonlinear
gray prediction model and linear regression model as the relationship. It divides the overall data area. After the
basic model and then carried out an equal weighted average division, the truncated spline function is used to fit the data
combination of the two sets of prediction results to estab- to each small region. According to the properties of the
lish a combined prediction model to predict the medium
and long-term power load. Gray system and neural network
25
are more combined in tunnel prediction. Meanwhile, dif-
ferent scholars have proposed combination methods such
as series, parallel, embedding and mixing and have been 20
verified in case analysis. Minggui Liu et al. used a com-
bined gray neural network model to predict the slope dis- 15
placement, indicating that the combined model can
accurately predict the slope displacement with small sam-
ple data and lack of information [20]. Xiuzhen Li estab- 10

lished an optimal weighted combination model of landslide

deformation based on the optimal weighted combination 5
theory and taking deformation monitoring data as an
example. Other combination models include gray time
0
series model [21], wavelet gray combination model, 0 1 2 3 4 5 6 7 8 9 10 11
wavelet neural network model, Kalman filter gray
Fig. 1 MARS model

123
242 Neural Computing and Applications (2021) 33:239–255

truncated spline, each small region is a linear function, as computing power, capable of processing a large amount of
shown in Fig. 1: sample data, and the calculation results can be ANOVA
The MARS model is defined as follows: decomposed, which can well explain the impact of input
X
M variables on output variables.
y^ ¼ f^M ð xÞ ¼ a0 þ am B m ð X Þ The MARS method is to construct a new basis function
m¼1
ð1Þ to fit the data. The final model of MARS is a linear com-
 
X
M Y
km    bination of the basis functions Bm x1 ; . . .; xp :
¼ a0 þ am skm xvðkmÞ  tkm þ
:
m¼1 k¼1   XM  
f^ x1 ; . . .; xp ¼ am Bm x1 ; . . .; xp ð4Þ
In the formula: y^ is the predicted value of the output m¼0
variable, a0 is the parameter, am is the coefficient of the m-
The combination coefficients am of basis functions can
th spline function, Bm ð X Þ is the m-th spline function. The
be obtained using the least squares method:
intersections of the linear regression lines between regions " #2
are called nodes, and KM is the number of nodes. The M
X
N XM  
value of skm is 1 or - 1, which represents the spline fam g0 ¼ arg min yn  am Bm xn1 ; . . .; xnp ð5Þ
n¼1 m¼0
function on the right or left, vðk; mÞ is the identifier of the
independent variable, and tkm is the position of the node. The basis function of the MARS model is the result of
Each basis function represents a given area of the depen- the product of different functions:
dent variable [24]. The spline basis function Bm ð X Þ of
  Ykm   
MARS may be a single spline function or the product of Bm x1 ; . . .; xp ¼ bkm xvðkmÞ jPkm ð6Þ
two or more spline functions. k¼1

   ðx  tkm Þþ x [ tkm bkm is a truncated polynomial function, bkm is related to the
skm xvðkmÞ  tkm þ ¼ ð2Þ
0 otherwise input variable xvðkmÞ and parameter space Pkm . The form of
 bkm is as follows:
   ðtkm  xÞþ tkm [ x
skm xvðkmÞ  tkm þ ¼ ð3Þ
0 other bkm ðxjs; tÞ ¼ ½sðx  tÞþ ð7Þ
In the formula: t is the position of the node. ðx  tkm Þþ In the formula:
and ðtkm  xÞþ are a pair of axisymmetric truncated splines 
z z[0
representing the areas on both sides of the node, as shown ½zþ ¼ ð8Þ
0 otherwise
in Fig. 2.
MARS processes input data similarly to other fitting The MARS algorithm is divided into three steps: for-
methods. It trains the input data ward process, backward pruning process and model
 N selection. The forward process divides the input sample
yi1 ; . . .; yiq ; xi1 ; . . .; xiq i¼1 to get the desired function
^ data, and each divided point is a node. When a spline
fj ð xÞ instead of fj ð xÞ and then analyzes and calculates the
function is fitted to each divided cell, a new basis function
problem and makes a decision. Among the many fitting is obtained. The forward process generates many basis
methods, the MARS model is fast in modeling, strong in functions and results in an overfitted model. The backward
pruning process is a process of removing redundant basis
functions while ensuring the accuracy of the model.
Finally, among the many models, the model with the best
accuracy is selected as the final result.

3.2 Forward process algorithm for MARS model

The first set of basis functions (number of iterations I ¼ 0)

of the MARS model is constant, that is, B0 ð xÞ ¼ 1, and
each basis iteration ðI [ 1Þ produces two basis functions:
B2I1 ð xÞ ¼ Bl ð xÞbðxv jPÞ ð9Þ
 
B2I ð xÞ ¼ Bl ð xÞb xv P ð10Þ

In the formula, Bl ð xÞ is the basis function in the previous

Fig. 2 Truncation function iteration, and xv is the input variable:

123
Neural Computing and Applications (2021) 33:239–255 243

  1 
P ¼ ðs; tÞ P ¼ ðs; tÞ ð11Þ C ðM Þ ¼ trace B BT B BT þ 1 þ dM ð14Þ
In the formula, t is the node of the spline base. During
In the formula, M is the number of basis functions, B is a
each iteration, the position of the node affects the accuracy  
1
of the model construction, and the number of nodes also matrix of M  N, and trace BðBT BÞ BT þ 1 is the
affects the calculation time of the model. The MARS number of effective coefficients in the MARS model,
model does not calculate every input data to see whether which is usually the same as M in numerical terms. d is the
these data can become the node of the new basis function. penalty coefficient of the model, which usually takes values
In order to avoid that the distance between each adjacent between 2 and 4. Among the many models, the model with
node is too small, a minimum step size L can be defined for the smallest GCV value is the optimal MARS model.
each input variable. Its calculation method is: The final form of the MARS model is:
1 X
M Y
km   
LðaÞ ¼  log2  Inð1  aÞ 2:5 ð12Þ f^ð xÞ ¼ a0 þ am skm xvðkmÞ  tkm þ : ð15Þ
PN
m¼1 k¼1
In the formula, 0:01  a  0:05. In this way, the selec-
Since each basis function is a product form of a trun-
tion of nodes will not affect the accuracy of the model and
cated spline function of one or more variables, in order to
can reduce the search area of several points. The important
facilitate the analysis of the model, Eq. (15) can be
thing is that it can reduce the calculation time and greatly
decomposed by ANOVA. In this way, the interactive
increase the speed of model building. The basis function of
influence of input variables and the subordinate relation-
the MARS model is the product of one or more spline
ship between the basis functions can be intuitively seen:
functions. The algorithm must also consider the interaction X X  
of different functions to improve the accuracy of the f^ð xÞ ¼ a0 þ f i ðx i Þ þ fij xi ; xj
model. For the number of interaction basis functions, it can Km ¼1 Km ¼2
X   ð16Þ
freely limit the maximum number of interactions in the þ fijk xi ; xj ; xk þ   
MARS model (for example, if the upper limit of the Km ¼3
number of interactions is 3, the interaction degree of each P
The second term Km ¼1 fi ðxi Þ in Eq. (16) indicates that
basis function of the model does not exceed 3). In the P
there is only one input variable fi ðxi Þ ¼ a B ðx Þ in
process of model establishment, the number of basis P  m m i
each basis function, the third term Km ¼2 fij xi ; xj indi-
functions continues to increase until it reaches the maxi-  
mum value ðMmax Þ of the specified number. In general, the cates that there are two input variables fij xi ; xj ¼
P  
number of Mmax is twice the size of the optimal model M  am Bm xi ; xj in each basis function, and the fourth term
P  
or twice the number of model input variables. It is worth Km ¼3 fijk xi ; xj ; xk means that each basis function con-
noting that the forward process of the MARS model will tains three input variables
generate a large number of basis functions, resulting in   P  
fijk xi ; xj ; xk ¼ am Bm xi ; xj ; xk , and so on. In this way,
over-fitting. the influence of input variables on the model can be ana-
lyzed intuitively.
3.3 Backward process algorithm for MARS
3.4 MARS model post-processing
Because the forward process generates many pairs of
truncated function groups, these functions will be added to In the processing of the model, since the basis function is
the model, causing the model to overfit and increasing the the product of the truncated form spline, the truncated
complexity of the model. Therefore, the forward iteration spline does not have a continuous derivative at the trun-
process of MARS allows the construction of a large cation. To make the model smoother, each truncated spline
number of basis functions. Therefore, the backward prun- should be expressed as a truncated cube at the truncation:
ing process plays an important role in the construction of
the MARS model. This process deletes the basis functions C ðxjs ¼ þ1; t ; tþ Þ
8
in the forward process and uses the generalized cross-val- >
<0 xt
ð17Þ
idation (GCV) criterion. ¼ pþ ðx  t Þ2 þrþ ðx  t Þ3 t  x  tþ
Pn 2
>
:
1 m¼1 ðyi  y^i Þ
xt x  tþ
CGVðM Þ ¼ ð13Þ
n nð1  C ðM Þ=nÞ2

C ðM Þ is a penalty function for model complexity, which is

defined as follows:

123
244 Neural Computing and Applications (2021) 33:239–255

C ðxjs ¼ 1; t; tþ Þ The basis function of the BMARS model is a B-spline

8
< ðx  tÞ
> x  tþ
ð18Þ
basis function instead of a truncated power spline, that is:
¼ p ðx  tþ Þ2 þr ðx  tþ Þ3 t  x  tþ   Ykm  
>
: Tm x1 ; . . .; xp ¼ Bsðk;mÞ xvðk;mÞ : ð24Þ
0 xt
k¼1

pþ ¼ ð2tþ þ tþ  3tÞ=ðtþ  tþ Þ2 ð19Þ Similarly, the first basis function set of the BMARS
rþ ¼ ð2t þ tþ  tÞ=ðtþ  tþ Þ 3
ð20Þ model is constant, that is:
T0 ð xÞ ¼ 1: ð25Þ
p ¼ ð3t  2t  tþ Þ=ðt  tþ Þ2 ð21Þ
The second basis function is the product of the resulting
r ¼ ðt þ tþ  tÞ=ðtþ  t Þ3 ð22Þ B-spline function and the first basis function, that is:
Three nodes t; tþ ; t can write a cubic truncation T1 ð xÞ ¼ T0 ð xÞBs ½xv : ð26Þ
expression. Because each variable forms multiple basis
functions, there will be multiple nodes. These nodes are The BMARS backward pruning process is similar to the
projected into the corresponding variable space. tþ ; t , MARS model. At each iteration, it deletes a basis function
respectively, representing the midpoints of adjacent nodes. that does not significantly affect the model and finally uses
Although the calculation speed of this model will increase, the GCV standard to select the best one among the many
the increase is not large and the connection between models as the final result. The characteristic of BMARS is
regions will be smooth on the graphic display. to use the B-spline function as the basis function of the
model, making the model smoother.
3.5 BMARS algorithm
3.6 Bernstein-MARS algorithm
Sergey Bakin et al. proposed the multivariate adaptive
regression B-spline (BMARS) method in 1998. BMARS In image signal processing, noise reduction is the key issue.
inherits many advantages of MARS and can adaptively The commonly used methods are Fourier transform and
process high-dimensional data. At the same time, B-splines wavelet analysis. In this section, a method of noise
are smoother than truncated splines. Sergey Bakin et al. reduction using Bernstein basis functions is given. At the
used a parallel algorithm when constructing the BMARS same time, Bernstein basis functions are introduced into the
model, which speeded up the construction of the model. MARS method.
BMARS can be constructed with B-spline functions of The Bezier curve fitted by the Bernstein basis function
any order. Generally, second-order B-splines are used. has the form of the prediction curve xðtÞ under ideal
Moreover, second-order B-splines have a computational conditions:
advantage over other functions as a basis function. The xðtÞ ¼ b0 Bn0 ðtÞ þ    þ bn Bnn ðtÞ ð27Þ
second-order B-spline has three local supports, and the size
Bni ðtÞ in the formula is a Bernstein basis function, so we get
of the support interval is controlled by the nodes. In this
the following form:
way, BMARS can construct the basis function of the model
by selecting the support. b0 Bn0 ðt0 Þ þ    þ bn Bnn ðt0 Þ ¼ p0
The modeling process of BMARS and MARS is similar     ð28Þ
b0 Bn0 tp þ    þ bn Bnn tp ¼ pp :
and includes three steps: the forward process generates a
lot of basis functions and simultaneously generates an It can be written as the following matrix:
over-fitting model, the pruning process eliminates redun- 2 32 3 2 3
Bn0 ðt0 Þ    Bnn ðt0 Þ b0 p0
dant basis functions and guarantees the model’s fitting 6 .. .. .. 76 .. 7 6 .. 7
4 . ¼
54 . 5 4 . 5: ð29Þ
accuracy, and at the end, the best model is selected. n
  . n
. 
BMARS uses different sizes of B-splines to divide the B0 tp    Bn tp bn pp
input data in the forward process, which is a significant It can be abbreviated as:
difference from the MARS forward process. Similar to the
MB ¼ P: ð30Þ
MARS model, the basis function of the BMARS model is
also the result of a different function tensor, that is: When the number of data points p is greater than the
  X
M   degree n of the curve, this equation is an overdetermined
f^ x1 ; . . .; xp ¼ am Tm x1 ; . . .; xp : ð23Þ equation. By left multiplying the matrix M T at both ends of
m¼0 the equation, the following formula can be obtained.

123
Neural Computing and Applications (2021) 33:239–255 245

M T MB ¼ M T P ð31Þ Similar to the MARS model, the basis functions of the

Bezier-MARS model are also tensor products of different
This is a system of n þ 1  n þ 1 linear equations. The functions, that is:
coefficient matrix M T M is a symmetric invertible real-
valued solution. In this way, the coefficient bi can be   XM  
f^ x1 ; . . .; xp ¼ am Fm x1 ; . . .; xp : ð36Þ
obtained from Eq. (31). The predicted Bezier curve can be m¼0
obtained by bringing bi into Eq. (27).
The basis function of the BMARS model is a B-spline
This method does not care about the smoothness of the
basis function instead of a spline in the form of a power
obtained curve, and the obtained curve may have a large
cut, that is:
swing. Moreover, in some cases, the wobble does not
correspond to reality. In addition, sometimes the statistical   Ykm  
error of individual data will cause the curve of the pre- Fm x1 ; . . .; xp ¼ Pi Binsðk;mÞ xvðk;mÞ : ð37Þ
k¼1
diction to change greatly. In order to reduce the impact, the
curve must be modified, so a simpler method is chosen Similarly, the first basis function set of the BMARS
here. When the second-order difference quotient D2 bi of model is constant, that is:
the curve coefficient is relatively small, the curve will not F0 ð xÞ ¼ 1: ð38Þ
swing particularly strongly. Therefore, in addition to for-
mula (31), the following formulas can be obtained: The second basis function is the product of the resulting
Bezier curve and the first basis function, that is:
b0  2b1 þ b2 ¼ 0
.. F1 ð xÞ ¼ F0 ð xÞBs ½xv : ð39Þ
. ð32Þ
bn2  2bn1 þ bn ¼ 0 The Bezier-MARS backward pruning process is similar
to the MARS model. That is, each iteration, a basis func-
It can be abbreviated as: tion with a low contribution to the model is deleted, and the
SB ¼ 0: ð33Þ GCV standard is used to select the best one among the
many models as the final result. Bezier-MARS is charac-
Bringing Eq. (33) into Eq. (30), we can get a new
terized by using Bezier curves as the basis function of the
overdetermined equation:
model, which makes the model smoother.
M P
B¼ ð34Þ
S 0
4 BP neural network prediction model
In the formula (34), the shape control equation of the
prediction curve is not processed. However, we can exer-
Artificial neural network is to organize a large number of
cise weighted control over two different equations.
simple functional neurons through a certain topology to
ð1  aÞM ð1  aÞP form a computing structure for group parallel processing.
B¼ ð35Þ
aS 0 Among them, the BP network is a multi-layer forward type,
and the transmission of neurons usually borrows S-type
At this time, a 2 ½0; 1. When a ¼ 0, the equation
functions, and the output is a continuous quantity of 0–1,
returns to Eq. (30), when a ¼ 0:5, the equation returns to
and it achieves the purpose of arbitrary nonlinear mapping
Eq. (34), and when a ¼ 1, the equation returns to Eq. (33).
from input to output. Because its weight is adjusted by the
According to experience, the larger the fluctuation, the
back-propagation learning algorithm, it is also called BP
closer the value of a is to 1. If the degree of volatility is
network. In practical applications of artificial neural net-
uncertain, a better result will be obtained when a ¼ 0:1.
works, BP networks and their variations are most widely
The Bezier curve constructed by the Bernstein basis
used.
function is smoother than the truncated spline. In this way,
Artificial neural network, also called neural network, is a
the Bezier curve can be constructed for each small region
mathematical model that simulates biological neural net-
by stepping through the fixed node selection.
works for information processing. Artificial neural net-
Similarly, Bezier-MARS’s modeling process is similarly
works are developed from biological neural networks. A
composed of 3 steps. The forward process generates a lot of
neuron is a nerve cell, and about 10 billion neurons appear
basis functions and generates an over-fitting model, and the
in the human cerebral cortex, and it has 60 trillion synapses
pruning process deletes redundant basis functions and
and their connections. When designing a neural network,
guarantees the model’s fitting accuracy, and finally selects
the following three aspects are mainly considered: the
the best model.

123
246 Neural Computing and Applications (2021) 33:239–255

functional function of the neuron, the topology of the only receives information feedback from the output
neural network, and the learning of the neural network. layer and other neurons in the same layer, but also
receives internal feedback information from itself.
1. Neuronal function
(1) Learning method The biggest significance of neural
Neurons are the basic unit of neural network for informa- networks is the ability to learn from the environment
tion processing. The artificial neuron model is composed of and improve their ability through repeated learning.
three basic elements: According to different learning environments, neural
networks can be divided into supervised learning and
(1) A group of connections: the connection strength is
unsupervised learning. Supervised learning is to
represented by the value on each connection, the
input training sample data to the input layer and
weights have positive and negative values, positive
compare the corresponding expected output with the
indicates activation, and negative indicates
network output to obtain error information, to control
inhibition;
the connection strength adjustment of the weights,
(2) An adder: it is used to find the sum of the
and to converge to a certain weight after repeated
corresponding synaptic weights of the input signals
training simulations. The characteristic of this
to the neurons;
method is that when the sample changes, the weights
(3) An activation function: Its function is to limit the
can be adjusted to adapt to the new situation after
input signal to a certain value of the allowable range,
network training. Unsupervised learning does not
and the normal amplitude range of the output can be
give samples, and the network will be placed directly
written as the unit closed interval ½0; 1 or ½1; þ1.
in the environment, and their parameters or struc-
At present, neural networks have nearly 40 models such tures will be adjusted according to the statistical
as back-propagation network, perceptron and self-orga- characteristics of the data, and the weights and
nizing map. According to the different topology of the thresholds will be adjusted through the input, and
connection, it can be divided into forward network and there is no target output.
feedback network: (2) Learning Algorithm The learning algorithm is a
learning problem with clear rules. Different learning
(1) Forward network The principle is that each neuron
algorithms have different expressions for neuron
receives the input from the previous level and
weights, and different neural networks need to
outputs it to the next level. Each layer only has
choose a suitable learning algorithm. The commonly
input connections to the adjacent previous layer and
used learning algorithms include Hebb learning
follows the rule from the back. It can be represented
rules, learning rules, random learning algorithms
by a directed acyclic graph, and the forward network
and competitive learning algorithms.
structure is shown in Fig. 3.
(2) Feedback network The difference between this The neural network consists of an input layer, an output
neuron and the forward-type network is that it not layer and a hidden layer. The layers are completely con-
nected. The BP neural network value uses an error back-
propagation algorithm. The BP neural network diagram is

Fig. 3 Forward network Fig. 4 Schematic diagram of the BP neural network

123
Neural Computing and Applications (2021) 33:239–255 247

shown in Fig. 4. BP algorithm learning refers to the error (2) Its susceptibility to local minima: The BP algorithm
back-propagation algorithm using the weights and thresh- uses the gradient descent method. The starting point
olds of the network. The BP neural network diagram is of training is to meet the requirement of minimum
shown in Fig. 4. In the figure, x1 ; x2 ; . . .; xn is the input error value according to the slope surface of the error
layer and y1 ; y2 ; . . .; yn is the output layer. function. However, for complex networks, the error
The BP network is divided into two stages: the first stage function is a multi-dimensional surface, which may
is the input of known learning samples, and the output of fall into a small area and produce a local minimum.
each neuron is recalculated from the first layer of the net- (3) The number of hidden layer neurons is not specified:
work through the network structure and the weights and The network training scale is determined by the
thresholds set in the first iteration. The second stage is that number of hidden layers and the number of nodes,
the weights and thresholds of the network are modified which has a lot to do with network performance.
from the previous calculation of the influence of the However, the determination of the number of hidden
weights and thresholds on the total error, and then the layers and the number of nodes does not have a
weights and thresholds are modified. We set the number of specific theory and can only be obtained based on
samples to k ¼ 1; 2; . . .; m. Then, the error function can be experience or repeated experiments.
expressed as: (4) Online learning and memory are unstable: The newly
X
k added samples for training need to restart the trained
1
e¼ ðTk ðkÞ  y0 ðkÞÞ2 : network, and the network cannot remember the
2 O¼1 previously trained weights and thresholds.
The learning steps of a three-layer BP network are as In view of the shortcomings of the standard BP algo-
follows: rithm, many optimization algorithms have appeared:
(1) Each connection weight of the network is initialized (1) Additional momentum method This algorithm adds a
to a random number between 0 and 1; momentum term to the BP algorithm to fine-tune the
(2) The training samples and their corresponding correction value of the weight to achieve the purpose
expected outputs are input to the network; of preventing learning from falling into a local
(3) The input and output of each neuron in the middle minimum.
layer and the error e are calculated; (2) Gradient descent method with adaptive learning rate
(4) The calculation error e and the total system error The adaptive learning rate can adjust the learning
E is, respectively: rate according to the local error surface, which can
improve the model’s convergence speed.
(3) Elastic gradient descent method Its advantage is to
1Xm1
e¼ ð dk  y k Þ 2 ð40Þ eliminate the adverse effect of the gradient ampli-
2 k¼1
tude. When the training oscillates, the weight change
will decrease.
! (4) Levenberg–Marquardt algorithm The LM algorithm
X
p
1 XX
p m1
is to avoid calculating Hessian matrices when
E¼ EI ¼ ðdik  yik Þ2 ð41Þ
i¼0
2P i¼0 k¼1 modifying the approximate second-order training
rate. Moreover, the LM algorithm is also the fastest
(5) The weights of the output layer and the hidden layer algorithm proposed for training medium-scale feed-
are calculated and adjusted until the set requirements forward neural networks.
are met; (5) Combining with other models The neural network
(6) n ¼ n þ 1 is set and the algorithm returns (3) until model can be combined with other methods to
the predicted value is output. realize the advantages and disadvantages and opti-
The BP network algorithm has a wide range of appli- mize the model.
cations due to its own advantages, but it also has some This paper uses a three-layer neural network with a
shortcomings, including the following aspects: hidden layer and uses MATLAB to design the network.
(1) The convergence speed is slow: The traditional BP (1) Input layer design For the prediction of deep
neural network uses gradient descent to learn. This foundation pit deformation, the structure of the input
learning method requires relatively low learning rate, variables depends on the number of main factors
which makes the network training time long and the affecting the deformation of the foundation pit.
convergence speed slow. However, there are many factors that affect the

123
248 Neural Computing and Applications (2021) 33:239–255

deformation of the foundation pit, such as the depth

of the foundation pit, the type of support and changes
in water pressure, etc., and the foundation pit
engineering is a dynamic process. If the influencing
factors are used as the input variables of the network,
the network structure will be complicated, and the
factors affecting the deformation of the foundation
pit cannot be accurately obtained. Therefore, when a
neural network is used to predict the deformation of
a deep foundation pit, the existing foundation pit
deformation monitoring values that can comprehen-
sively reflect the above factors are usually used as
input layer data. That is, the existing displacement
monitoring value is used to predict the displacement
value of the same position within a certain time. A
large number of studies have shown that it is more
appropriate to use the actual monitoring data of the
day before yesterday, yesterday and today to predict
the deformation of tomorrow, that is, the 3 ? 1
input–output structure.
(2) Hidden layer design The network model can be used
for one or more hidden layers. However, research
shows that a three-layer BP neural network can
achieve any nonlinear mapping from input to output. Fig. 5 Flowchart of BP neural network prediction model
The number of nodes in the hidden layer is
determined by the complexity and model of the 5 Gray residual BP neural network model
actual problem, and there is no clear theoretical
formula. According to the empirical formula, the With the development of the disciplines of system science
number of nodes in the hidden layer is and systems engineering, the technology of system
pffiffiffiffiffiffiffiffiffiffiffi assessment and prediction has been continuously devel-
m ¼ n þ 1 þ o, n, l are the number of input and
output layer nodes, and o is a constant between 1 and oped. Generally, each method and model have its advan-
10. tages and disadvantages. In practical applications, several
(3) Output layer design In order to achieve the purpose methods or models are usually combined to form a com-
of enabling the network to output arbitrary values, bination. The goal is to combine and absorb the advantages
the output layer uses pureline as the transfer of different methods and overcome the limitations of a
function, and the number of output layer nodes is single model. Among them, the gray system theory is
selected as one. The calculation flowchart of the strongly complementary to many other traditional tech-
whole model of BP neural network is shown in nologies and soft computing calculations. The neural net-
Fig. 5. work has the characteristics of strong error correction
ability, adaptive ability and nonlinear ability. However, the
The operation mechanism of BP neural network is as neural network is difficult to grasp the development trend
follows: The original data sequence is X ¼ fxi g; i ¼ of the system when there are insufficient samples, and the
1; 2;    ; n and the input sample is: qualified sample training will cause poor training results
 
fx1 ; x2 ; x3 g; fx3 ; x4 ; x5 g; xðn6Þ ; xðn5Þ ; xðn4Þ : due to the existence of some points with large fluctuations.
  Moreover, the gray system theory requires a small number
The output sample is x4 ; x5 ; xðn3Þ . This is an equal of samples, which can compensate the neural network’s
dimensional information BP network model, which pro- dependence on a large number of samples and can better
cesses the data according to the metabolic principle in the process small sample data in engineering. By using the
gray theory, so that the model learns according to the latest advantages of the two models, a non-equidistant interval
information and improves the prediction accuracy. The gray GMð1; 1Þ model and a BP neural network model are
purpose of learning is to use the network’s error to reverse combined to form a gray residual BP model to improve
the weights, and then use the trained network model to prediction accuracy.
predict the n  2, n  1, and n displacements.

123
Neural Computing and Applications (2021) 33:239–255 249

The gray residual BP neural network designed in this

paper uses a series combination, that is, the residual
obtained by the gray model is used as the input data of the
BP neural network, and the final output of the combined
model is the prediction result. When predicting the hori-
zontal displacement of a deep foundation pit project, we
take the error sequence of the gray model as the input data
of the neural network and take the measured value of the
output data to construct a BP neural network for training. A
hidden layer is set in the neural network in the gray residual
BP neural network combination model. The method
adopted is to use the BP neural network to correct the
residual of the gray model. The principle is to first establish
a gray GMð1; 1Þ prediction model and obtain the simula-
tion values to obtain the residual sequence
 
eðiÞ ; i ¼ 1; 2; . . .; n. After that, the residual sequence fe1 ;
 
e2 ; e3 g; fe2 ; e3 ; e4 g; . . . eðn6Þ ; eðn5Þ ; eðn4Þ is used as the
input sample of the network, and the output sample is
fe4 ; e5 ; en g, which is the BP network model of equal
dimension information. The BP network model of equal
dimension information processes the sequence according to
the metabolism in the gray theory and uses the BP network
 
model to predict the residual sequence eðiÞ to obtain a Fig. 6 Flowchart of gray residual BP neural network prediction
n o
model
new sequence e0ðiÞ , and then to obtain a new predicted
ð0Þ  
value yðiÞ ¼ xðiÞ þ e0ðiÞ . In summary, yðiÞ is the predicted P ¼ Xðnþ1Þ ¼ iðnþ1Þ jX0 ¼ i0 ; Xi ¼ i; . . .; Xn ¼ in
value of the gray BP neural network. The calculation   ð42Þ
¼ P Xðn1Þ ¼ iðn1Þ jXn ¼ in
flowchart of the complete gray residual BP neural network
model is shown in Fig. 6. Then, we call fXn ; n 2 T g as Markov chain.
The gray residual BP neural network model does not For any integer n 2 T and state i; j 2 I
 
have high prediction accuracy for abrupt data or data with Pi;j ðnÞ ¼ P Xðnþ1Þ ¼ jjXn ¼ i ð43Þ
large random volatility. Therefore, the Markov chain,
Markov, is introduced to modify the gray residual BP is the transition probability of the Markov chain.
neural network model, in order to be able to make more 2 3
  P11 P12    P1n   
accurate predictions of volatile data appearing in the hor- P ¼ Pij ¼ 4 P21 P22    P2n    5 ð44Þ
izontal displacement of the deep foundation pit press.     
The research object of Markov theory is mainly the state
and transition of sequences, and the process is a random P is the transition probability matrix of the system state,
h i
ðnÞ
process. We assume a sequence has several states, the and pðnÞ ¼ pij is the n-step transition probability matrix.
sequence is now in one of these states, and the next
moment may transition to other states. By studying the
initial probability of each state and the transition proba-
6 Markov-Gray residual BP neural network
bility between states, the state trend of the next moment in
combination model
the sequence is determined, and the prediction is per-
formed. The random process under this discrete time is the
The Markov-Gray residual BP neural network combination
Markov process.
model is a new method that combines the advantages of the
Markov chain is the simplest Markov process. We
gray GMð1; 1Þ model, BP neural network and Markov. The
assume that fXn ; n 2 T g is a random process. For any
gray GMð1; 1Þ model is suitable for the prediction of sit-
integer n 2 T and any state i0 ; i1 ; iðnþ1Þ 2 I, the conditional
uations with few samples and poor information, but the
probability satisfies. prediction accuracy for mutation data is not high, and it is
not suitable for long-term prediction. BP neural network

123
250 Neural Computing and Applications (2021) 33:239–255

has strong fault tolerance and learning ability, can predict

long-term data and has high prediction accuracy. Because
the output of the neural network is gradually approaching a
fixed value with a certain accuracy, and the output result is
based on the error of the fixed value, the output of the
neural network is gray data, so the gray network can be
used to analyze the neural network. Aiming at the problem
that the model cannot predict large fluctuation random
data, the principle of Markov chain is introduced. The
Markov chain limits the prediction residuals of the gray
residual BP neural network to a small range. According to
the Markov chain method, the deviation rule of the gray
residual BP network model is obtained, and the prediction
results are corrected according to this rule. The gray
residual BP network model improved by Markov chain can
make use of the three advantages and make the prediction
accuracy higher.
The Markov-Gray residual BP neural network combi-
nation model first uses non-equidistant gray GMð1; 1Þ to
predict the monitoring data, and then uses the output
 
residual sequence p ¼ eð0Þ ði  1Þ; eð0Þ ði  2Þ; eð0Þ ði  3Þ
as the input sequence of the BP neural network model and
 
T ¼ eð0Þ ðiÞ as the output. After the samples are deter-
mined, the gray BP neural network is trained to obtain the
 
new prediction value of eð0Þ ðiÞ , and the residual
sequence feðiÞg of the gray residual BP network combi-
nation model is obtained. The Markov chain is used to
modify feðiÞg to obtain the predicted value of the Markov-
Gray residual BP neural network combination model. In
the Markov-Gray residual BP neural network combination
model, the input layer of the network is set to 5 nodes, the
output layer is set to 1 node, and the number of nodes in the
hidden layer is obtained by continuous training to obtain
the optimal number of nodes. That is, the first three residual Fig. 7 Flowchart of Markov-Gray residual BP neural network
sequences predicted by the gray residual BP network combined model
model are used as the input of the BP network, and the
fourth residual value is used as the output. The modeling (4) The state transition probability and state transition
steps are shown in Fig. 7: matrix are determined. If Mij ðmÞ is the number of
original data samples transferred from state i to step j
(1) According to the original monitoring value, the gray
through m steps, and Mj is the number of original
GMð1; 1Þ is used to predict the monitoring data; 
  data samples in state i, then, Pij ðmÞ ¼ Mij ðmÞ Mj
(2) The residual sequence eð0Þ ðiÞ predicted by the
ði ¼ 1; 2; . . .; nÞ.Pij ðmÞ is the state transition proba-
GMð1; 1Þ
bility, and Markov’s one-step transition probability
n o matrix is represented by Pð1Þ:
p ¼ eð0Þ ði  1Þ; eð0Þ ði  2Þ; eð0Þ ði  3Þ
Pð1Þ ¼ ½P11 ð1Þ; . . .; P1n ð1Þ; . . .; Pn1 ð1Þ; . . .; Pnn ð1Þ
is used as the input sequence of the BP neural net-
 
work model and T ¼ eð0Þ ðiÞ is used as the output Then, the probability transition matrix of m steps
to train the BP network. is PðmÞ ¼ ½pð1Þm .
(3) The residual sequence obtained by the gray BP (5) For the one-step transition probability matrix pð1Þ,
combination model is divided into n states. we assume that the prediction object is in the k state.
If maxpð1Þkj ¼ pð1Þkl , it can be considered that the

123
Neural Computing and Applications (2021) 33:239–255 251

system is most likely to change from state k to state l Table 1. In order to better predict the deformation data of
at the next moment. If there are two or more tunnel monitoring, ensure the construction safety and
probabilities in the Pð1Þ-th row of the matrix g, provide guidance for tunnel construction, this paper uses an
which are the same, we need to examine the two-step improved wavelet BP neural network model to analyze and
or n-step transition probability matrix. If there are predict the deformation data of tunnel deformation. The
two or more probabilities in the k-th row of the required items for tunnel deformation monitoring in this
matrix Pð1Þ, which are the same, we need to examine project are mainly surface subsidence, arch subsidence and
the two-step or n-step transition probability matrix. surrounding convergence.
(6) The prediction results of Markov-Gray residual BP According to the wavelet denoising principle, the best
combination model: Y ¼ y=ð1  qÞ. In the formula, y wavelet function used in this project is selected. From the
is the predicted value of the gray BP model, and q is settlement curve in Fig. 9, it can be seen that the overall
the cutoff value of the original state space. settlement of the tunnel surface begins to sink sharply and
gradually stabilizes. However, there are sudden or unsta-
According to the above modeling steps, the flowchart of
ble points in the change curve. These changes are usually
the entire Markov-Gray residual BP neural network com-
caused by errors that occur during observations. In order to
bination model is shown in Fig. 7.
avoid false early warning caused by sudden changes in the
curve, it is necessary to denoise the deformation monitor-
ing data, propose noise signals and obtain real deformation
7 Simulation analysis
data, which can more accurately reflect the changing trend
and ensure the safety of the tunnel.
The surface subsidence data of DKl505 ? 693 shallow
Write an experimental program through MATLAB,
buried section were selected for analysis. There are 5
using soft threshold, scale = min. The maximum signal-to-
points (points 1 to 5) in the ground settlement layout, and
noise ratio improvement achieved by different layered
the cumulative settlement curve is shown in Fig. 8.
wavelet functions is denoised at four thresholds: general
Point 1 is the measurement point directly above the
threshold, Stein unbiased risk threshold, maximum mini-
central axis of the tunnel. The measuring points from far to
mum criterion threshold and heuristic threshold. The
near to the left of Point 1 are points 2 and 4, respectively,
maximum SNR that can be achieved by the three wavelet
and the measuring points from far to near to the right of
functions is shown in Fig. 8. The results of the maximum
Point 1 are points 3 and 5, respectively. Among them, since
signal-to-noise ratio achieved by different wavelet func-
the No. 1 point is located directly above the central axis of
tions db (‘dbl’; ‘db2’; ‘db3’; ‘db4’; ‘db5’; ‘db6’; ‘db7’;
the tunnel excavation, the amount of deformation is the
‘db8’), sym( ‘syml’; ‘sym2’; ‘sym3’; ‘sym4’; ‘sym5’;
largest, and the more the remaining measuring points are
‘sym6’; ‘sym7’; ‘sym8’), and coif( ‘coifl’; ‘coit2’; ‘coif3’;
away from the central axis, the more the deformation
‘coif4’; ‘coil5’) are shown in Table 2. The statistical graph
gradually decreases. Firstly, the measurement data of point
is shown in Fig. 10.
1 are selected for wavelet analysis, and the cumulative
By comparing the maximum signal-to-noise ratios of
settlement observation data of point 1 are shown in
different wavelet functions in different layers, it can be
concluded that the signal-to-noise ratio of db wavelet and
16
sym wavelet is the same in the case of N = 1–4. For signal
Point 1
14 denoising with this requirement, both can be used. When
Cumulave selement value

Point 2
N = 4.8, the signal-to-noise ratio of db wavelet is greater
12
Point 3 than the signal-to-noise ratio of sym wavelet. Coifd
10 Point 4 wavelet is generally not used when N = l–2, although it
8 Point 5 shows obvious denoising effect. Generally, the engineering
examples encountered in the signal denoising generally use
6
3–6 layers of layering. By comparing the signal-to-noise
4 ratio, we can conclude that db wavelet is the best choice.
2
The denoising effect of different decomposition layers
of the wavelet is calculated to determine the optimal
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51
number of decomposition layers. The db5 wavelet and the
-2 soft threshold function are selected. The maximum signal-
Observaon period
to-noise ratio improvement achieved by different layered
Fig. 8 Cumulative deformation when the ground surface subsidence wavelet functions is denoised at four thresholds: general
at 1–5 points threshold, Stein unbiased risk threshold, maximum

123
252 Neural Computing and Applications (2021) 33:239–255

Table 1 Observation data of

Observation period Cumulative settlement value Observation period Cumulative settlement value
cumulative settlement at point
1 0 26 11.0292
2 0.5353 27 11.9988
3 1.8079 28 12.12
4 1.1312 29 12.8674
5 2.1513 30 12.8674
6 3.3835 31 12.9482
7 4.343 32 13.5643
8 4.8379 33 13.5643
9 5.454 34 13.029
10 5.0601 35 13.0492
11 5.7267 36 12.6755
12 6.6155 37 12.5846
13 6.2721 38 12.3321
14 7.1407 39 12.221
15 7.5548 40 12.8371
16 8.484 41 13.5542
17 7.4841 42 13.332
18 8.4032 43 13.5542
19 8.4032 44 13.332
20 8.5749 45 13.13
21 9.4435 46 13.3421
22 9.797 47 13.029
23 10.4333 48 12.9482
24 10.6252 49 12.9482
25 11.413 50 13.0391

12 Table 2 Statistical table of the

db sym coif
signal-to-noise ratio of different
Maximum signal-to-noise rao (dB)

11.5 wavelet functions 1 8.23 8.23 10.24

11 2 9.97 9.97 10.85
3 11.23 11.23 10.85
10.5
4 11.23 11.23 11.08
db
10 5 11.68 11.23 11.08
sym 6 11.68 11.23 0
9.5
coif 7 11.68 11.23 0
9 8 11.68 11.23 0

8.5

8 increases with the increase in the number of decomposition

1 2 3 4 5 6 7 8 layers. When the fourth layer decomposition is reached, the
Wavelet base denoising signal-to-noise ratio of the four threshold func-
tions of the improvement amount reaches the maximum,
Fig. 9 Effect of different wavelet functions on the denoising effect
and the subsequent course decreases. Stein’s unbiased risk
threshold is better than the other three when the number of
minimum criterion threshold and heuristic threshold. As
decomposition layers is greater than 4. By comparing the
shown in Fig. 11, the denoising processing conditions of
denoising signal-to-noise ratio, it can be intuitively deter-
the four threshold functions at different levels of decom-
mined that the number of decomposition layers is four.
position are used to obtain the denoising signal-to-noise
According to the monitoring data of surface points, it
improvement of each threshold function at different levels
can be known that the cumulative change value increases
of decomposition. The overall range of the denoising effect
with time. In BP neural network prediction, the test sample

123
Neural Computing and Applications (2021) 33:239–255 253

14 50

12
40 Cumulave selement value (mm)
Signal to noise rao

10

Cumulave value (mm)

Regression analysis value (mm)
30
8 db

sym
6 20
coif
4
10

0
0 0 1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930
0 2 4 6 8 10
Serial number -10
Observaon me (d)
Fig. 10 Statistical diagram of the signal-to-noise ratio of different
wavelet functions Fig. 12 Comparison of original observations and regression analysis
at point 1
14
16
12 nigrsure
Denoising SNR improvement (dB)

sqtwolog 14
10
heursure
8 12
Cumulave value (mm)

minimaxi
6 10

4
8
Cumulave value (mm)
2
6 Predicted value (mm)
0
1 2 3 4 5 6 7 8 9 10 4
-2
2
-4
Decomposion level
0
Fig. 11 Relationship between the number of different decomposition 0 10 20 30 40 50 60 70
levels and the signal-to-noise ratio under the four thresholds Observaon me (d)

size can be constructed. In this study, MARS machine Fig. 13 Comparison of the predicted value of the BP neural network
at point 1 with the original observation
learning regression algorithm is used for data processing.
The comparison between the original observations at point
periods can be used to predict the data of the next 5 peri-
1 and the regression analysis is shown in Fig. 12.
ods, and the accuracy comparison can be more intuitive.
2. The raw data are taken by BP neural network The structure of the sample data is to predict the settlement
prediction: amount of the sixth period using the settlement data of the
first five periods, and the number of output layer nodes is
The constructed surface subsidence data are used to
one. There are many rules of thumb for determining the
predict, and the prediction result of BP neural network is
correct number of neurons to use in the hidden layer, but
compared with the original cumulative value as shown in
they usually ignore the number of training cases, the
Fig. 13. In order to reflect the prediction effect more
amount of noise in the target and the complexity of the
intuitively, the data of the last five periods are enlarged and
function. Therefore, the final choice of the neural network
displayed.
architecture is trial and error, and the number of nodes in
Since it cannot be compared with unknown data, it is not
the hidden layer is determined to be 10. According to the
possible to use existing data to predict future settlement
BP neural network activation function algorithm learning,
values. According to the data given, there are 50 periods of
it is determined that the activation function of the hidden
cumulative monitoring values. The data of the first 45
layer is logsig, and the activation function of the output

123
254 Neural Computing and Applications (2021) 33:239–255

layer is purelin. The training parameters are usually spec- this paper can meet the ideal accuracy requirements and is
ified according to requirements, the maximum number of suitable for the prediction analysis of deformation moni-
iterations is 2000, and the expected error is 0.0001. toring data of this tunnel project, which has guiding sig-
The statistical chart of the residual values of the three nificance for tunnel construction and provides guarantee
prediction results of the sinking point of the arch is shown for tunnel construction safety.
in Fig. 14.
By comparing the three types of predicted values with Acknowledgements The research is funded by the National Natural
Science Foundation of China (NSFC) under Grant No. 41790434 and
the observed raw data, it is concluded that the prediction the Key Research and Development Program of China Railway
processing of the new threshold function wavelet denoising (Grant No. K2019G033).
data while carrying out this research method is more
accurate than using BP neural network to denoise the Compliance with ethical standards
original data and the new threshold function. In the pre-
diction and analysis of the results of the last five periods, Conflict of interest The authors declare that they have no conflict of
by comparing the prediction and residual value analysis, it interest.
can be concluded that the improved model proposed in this
paper can be well applied to the analysis and prediction of
deformation monitoring data of the tunnel arch subsidence.
References
1. Junyu LL (2010) Application remote video monitor system in
cross-river tunnel engineering. Shanghai Construct Sci Technol
8 Conclusion 3:16
2. Xie X, Feng L (2011) Development of wireless sensor network
technology and its challenges in subway tunnel engineering.
During the tunnel construction process, risks will occur Yanshilixue Yu Gongcheng Xuebao/Chin J Rock Mech Eng
between the connection of various procedures. Before the 30(66):4047–4055
next process starts, it is necessary to timely analyze the 3. Wang F, Zhao Y, Li C et al (2018) An experimental study on the
possible risk points of the next process according to the corrosion characteristics of the karst tunnel engineering area in
Southwest China. Bull Eng Geol Environ 78:4047–4061
construction site conditions, and to do the protection and 4. Bui HG, Meschke G, Schillinger D (2018) Application of the
monitoring preparations before construction to minimize finite cell method to tunnel engineering simulation. PAMM
the risks. In this paper, the application of perimeter con- 18(1):e201800011
vergence and arch subsidence deformation monitoring data 5. Zhong Y (2018) Technical innovation for Sanyang road cross-
river tunnel project. Front Eng Manag 5(3):411–415
on the tunnel surface section to verify that the combined 6. Hu XD, Deng SJ, Wang Y (2018) Mechanical tests on bearing
model based on BPNN and MASR proposed in this paper capacity of steel pipe-frozen soil composite structure applied in
can be used to predict and analyze tunnel deformation Gongbei Tunnel. Yantu Gongcheng Xuebao/Chin J Geotech Eng
monitoring data. Based on the above conclusions, the 40(8):1481–1490
7. Dingli Z, Zhenyu S (2018) Structural stability of complex tunnel
combined model based on BPNN and MASR proposed in surrounding rock and its control. J Hydroelectr Eng 37(2):1–11
(in Chinese)
8. Mori T, Iizuka S, Nakayama T (2017) Material engineering for
0.6 silicon tunnel field-effect transistors: isoelectronic trap technol-
ogy. MRS Commun 7(3):1–10
0.5 9. Umili G, Bonetto S, Ferrero AM (2018) An integrated multiscale
approach for characterization of rock masses subjected to tunnel
Residual value 1 excavation. J Mech Geotech Eng 10(3):513–522
0.4
10. Zhang H (2018) The longest railway tunnel in China. Engineering
Residual value (mm)

Residual value 2
0.3 4(2):165–166
Residual value 3
11. Boultbee N (2018) Upper Lillooet river hydroelectric project: the
challenges of constructing a power tunnel for run-of-river hydro
0.2
projects in mountainous British Columbia. Engineering
4(2):260–266
0.1
12. Vignais N, Weresch J, Keir PJ (2017) Posture and loading in the
pathomechanics of carpal tunnel syndrome: a review. Crit Rev
0
25 30 35 40 45 Biomed Eng 44(5):397–410
13. Li Y, Wang Z (2017) Wind tunnel experiment on polypeptide
-0.1 derivatives in controlling wind erosion. Nongye Gongcheng
Xuebao/Trans Chin Soc Agric Eng 33(5):149–155
-0.2 14. Zhang N, Fang Q, Li Y et al (2017) Mechanical analysis of
Observation time (d)
secondary lining of high-speed railway tunnel. KSCE J Civ Eng
Fig. 14 Comparison of three predicted residual values of arch 22(5–6):2384–2389
subsidence

123
Neural Computing and Applications (2021) 33:239–255 255

15. Joshi A, Kiran R, Sah AN (2017) An experimental analysis to 21. Vassallo R, Mishra M, Santarsiero G et al (2019) Modeling of
monitor and manage stress among engineering students using landslide–tunnel interaction: the Varco d’Izzo case study. Geo-
Galvanic Skin Response meter. Work 56(3):1–12 tech Geol Eng 37:5507–5531
16. Yan S, Yan X (2019) Using labeled autoencoder to supervise 22. Wang X, Cai M (2020) A DFN–DEM multi-scale modeling
neural network combined with k-nearest neighbor for visual approach for simulating tunnel excavation response in jointed
industrial process monitoring. Ind Eng Chem Res rock masses. Rock Mech Rock Eng 53:1053–1077
58(23):9952–9958 23. Liu T, Gong GF, Yang HY et al (2019) Trajectory control of
17. Gong JF, Zhu Y, Zhang GZ (2018) The grade discrimination of tunnel boring machine based on adaptive rectification trajectory
large deformation of layered surrounding rock in tunnel and its planning and multi-cylinders coordinated control. Int J Precis Eng
treatment countermeasures. J Railw Eng Soc 35(12):51–55 Manuf 20(6):1721–1733
18. Shi YM, Gao FP (2017) Lateral instability and tunnel erosion of a 24. Gong YF, Wang B, Wei HB et al (2018) Surface subsidence law
submarine pipeline: competition mechanism. Bull Eng Geol of double-line shield tunnel based on Peck formula. Jilin Daxue
Environ 1:1–12 Xuebao (Gongxueban)/J Jilin Univ Eng Technol Ed
19. Huang M, Zhang Z, Tan Z (2017) Construction model of deep 48(5):1411–1417
tunnel for urban flood control and drainage in China. Tumu
Gongcheng Xuebao/China Civ Eng J 50:76–81 Publisher’s Note Springer Nature remains neutral with regard to
20. Wang HT, Zhuang XX, Sun HY (2019) Prediction of pipeline jurisdictional claims in published maps and institutional affiliations.
displacement induced by water leakage during tunnel construc-
tion. Geotech Geol Eng 37:5383–5389

123