machines

Article
A Reverse Modeling Method Based on CAD Model Prior and
Surface Modeling
Linlin Fan 1,2,3 , Junyi Wang 1,2, *, Zhigang Xu 1,2 and Xiao Yang 1,2

1 State Key Laboratory of Robotics, Shenyang Institute of Automation, Chinese Academy of Sciences,
Shenyang 110016, China
2 Institutes for Robotics and Intelligent Manufacturing, Chinese Academy of Sciences, Shenyang 110169, China
3 University of Chinese Academy of Sciences, Beijing 100049, China
* Correspondence: [email protected]

Abstract: Reverse engineering is an important process of real model digitization. However, the
existing methods are limited by the characteristics of the modeling object; the modeling is low
efficient, has poor versatility, and unstable accuracy. To complete the reverse engineering of a real
workpiece with irregular deformation based on the ideal computer-aided design (CAD) model, a
high-precision reverse engineering method of the workpiece based on the CAD model prior was
presented. Through the registration of the ideal CAD model and the point cloud model of the real
workpiece, the geometric feature position information and feature constraint information contained
in the CAD model are transmitted to the modeling process, which helps to improve the accuracy and
efficiency of reverse engineering. This method is applied to the reverse engineering process of the
engine compartment with slight irregular deformation, the modeling accuracy reaches 0.04 mm, and
the high-precision reverse engineering of the engine compartment is successfully realized.

Keywords: reverse engineering; 3D modeling; NURBS curve; surface modeling

Citation: Fan, L.; Wang, J.; Xu, Z.;

Yang, X. A Reverse Modeling Method
Based on CAD Model Prior and 1. Introduction
Surface Modeling. Machines 2022, 10,
Reverse engineering is the process of obtaining the CAD (computer-aided design)
905. https://doi.org/10.3390/
model of an object according to its physical entity. At first, it is often applied to modeling
machines10100905
without drawings, by measuring the surface data points of the real object and fitting regular
Academic Editor: Panagiotis features in reverse to obtain the CAD model of the object. With the development of geologi-
Kyratsis cal and topographic modeling and the smart city, reverse engineering has been applied to
Received: 6 September 2022
a wider range of application scenarios such as human body structure reconstruction [1],
Accepted: 4 October 2022
ancient cultural relics reconstruction [2], agriculture [3], manufacturing [4,5] and so on [6,7].
Published: 8 October 2022
In the field of industrial manufacturing, the application of reverse engineering without
drawings has been gradually replaced by modeling with drawings. For example, in the
Publisher’s Note: MDPI stays neutral
field of industrial digital twins, reverse engineering is applied to the model construction of
with regard to jurisdictional claims in
real workpieces with a slight deformation in the digital space, based on design drawings.
published maps and institutional affil-
The deformation of these workpieces may come from their manufacturing, transportation,
iations.
or other processes, which is unavoidable. This deformation does not affect the normal work
of the workpiece but may affect some precision assembly links. A digital twin constructed
by using the high precision constructed model based on a real deformation workpiece can
Copyright: © 2022 by the authors.
more accurately reflect various processes in the physical space. The traditional reverse
Licensee MDPI, Basel, Switzerland. engineering method uses 3D (3-dimension) measuring equipment to collect the surface
This article is an open access article point data of the object, and the operators select the surface points belonging to the same
distributed under the terms and characteristics and use an appropriate parameterized surface to fit the model surface or fit it
conditions of the Creative Commons to a regular figure. The accuracy and time for reverse modeling depend on the complexity
Attribution (CC BY) license (https:// of the modeled object and the proficiency of the operator, but even for skilled operators, it
creativecommons.org/licenses/by/ takes about a few hours to complete high-precision reverse modeling of a common part.
4.0/). This modeling method is time-consuming and relies entirely on manual operation, and the

Machines 2022, 10, 905. https://doi.org/10.3390/machines10100905 https://www.mdpi.com/journal/machines

Machines 2022, 10, 905 2 of 18

accuracy of the model is not stable because of the use of manual judgment. For example,
Lukáš Jančar et al. used Geomagic to reconduct a differential gear case, and the maximum
deviation of the created model was ±0.1 mm [8]. Feng Li et al. rebuild a component in the
SolidWorks 2013 environment, and the RMS (root mean square) of the reverse model with
a single source of data is 0.5 mm [9]. With the improvement of the accuracy of 3D measure-
ment sensors and measurement methods, the time cost and accuracy loss brought by the
reverse engineering process affects the value of high-precision 3D data. In the literature [9],
authors used a tactile probe to compensate for data from a laser line scanner to improve
the speed and accuracy of obtaining 3D data. The data collection speed increased by 90%,
and the root means square error of the regular feature model is reduced from 0.53 mm to
0.07 mm. Zhaohui Geng proposed an algorithm to generate the mean configuration based
on the data clouds collected from several scans and multiple parts. A Bayesian model with
prior knowledge of production processes and scanners is specified to model the statisti-
cal properties of the mean configuration, which could provide a more precise result [10].
However, due to the limitation of modeling methods, this high precision of 3D scanning
equipment and point cloud data is difficult to be reflected in reverse engineering [11],
especially when fitting irregular features.

2. Literature Review
Scholars use high-precision 3D data to solve problems by studying various 3D model
construction methods. Different 3D models have been produced successively. For example,
Kun Zhang et al. constructed a lightweight point cloud model using the data collected
by the 3D measurement sensor through the feature-preserved point cloud simplification
method [12]. Use methods such as the Delaunay triangulation [13], mapping method [14],
or the adaptive methods [15,16] construct the triangular mesh model [17]. In addition, the
skin surface model obtained by adding tolerance information to the point cloud gradually
formed based on the discrete shape method presented by Min Zhang et al. [18,19].
These 3D information models can reflect the real 3D information of the object as
accurately as possible and are suitable for solving different problems, respectively; however,
the expression form of the model is relatively large, and the calculation efficiency is seriously
affected when the model is used for a related calculation.
In intelligent manufacturing and other industrial fields, the role of the reverse model
is not only to demonstrate the shape of real objects in digital space but also to participate
in many simulation calculations as a basic model. Considering the lightweight model
expression form and the follow-up model-based simulation calculation, the model obtained
by surface modeling has more obvious advantages, but it is still an essential problem in
the process of reverse engineering on how to improve the efficiency of modeling while
reconstructing the surface accurately.
At present, there are several surface modeling methods, such as a free form to deform
the surface, parametric surface defined by different parametric equations, and implicit
surface expressed by implicit function, which has different characteristics and is suitable for
reverse engineering in different scenes. A free-form deform surface has the characteristics of
a smooth and intuitive construction process and can complete the construction of the surface
according to physical deformation. Researchers have optimized the deformation process
of the surface through different methods; for example, the force density method (FDM)
presented by Jean-Philippe Pernot et al. forms the basis of a new set of minimizations based
on the bar network geometry [20]. However, the free form deform surface is not suitable to
reconstruct a model that is not closed or contains branch structure, and the deformation
is an iterative process; if the point cloud data with low quality, the iterative speed will be
slow, and the iterative process is not easy to converge. An implicit surface can reduce the
influence of noise points in the data set, and the reconstruction result is smooth. However,
it is difficult to control the reconstruction efficiency and the complexity of the algorithm.
The process of the parametric surface reconstruction is simple, and it is a commonly used
design curve in the aerospace industry. There are two parametric surfaces that were widely
Machines 2022, 10, 905 3 of 18

used, the B-spline surface and the NURBS (non-uniform rational B-splines) surface. NURBS
is used in computer-aided design/computer-aided manufacturing (CAD/CAM) design in
the industry, and it has been identified as the standard of industrial product data exchange.
It is also a commonly used parametric surface form when constructing a surface from point
cloud data. For example, Lee J. Wells and others use a 3D point cloud data to construct the
NURBS surface to perform a SPC evaluation by detecting the control parameters of the
NURBS surface [21].
Methods for constructing and applying NURBS surfaces are well established [22].
H. Park presents an approximate lofting method for B-spline surface fitting to a functional
surface within a specified accuracy. This method adopts adaptive sampling and multiple
B-spline curve fitting; for surface approximation with tolerance δ = 10−4 mm, the fitting
accuracy can reach for 10−6 mm [23]. J. Yang et al. applicated NURBS in turbomachinery
design and obtained satisfactory results, and the redesigned efficiency increased by 3% [24].
Subsequently, Bopit Chouychai et al. proposed a tangent function method to find the
turning point of curve function by the derivative of the function at a point, to improve
the fitting speed of NURBS [25]. J. Apolinar Mu Muñoz Rodr Rodríguez et al. and
Xiaoqiang Tian et al. used a genetic algorithm and BP neural network-based on particle
swarm optimization to optimize the control point and knot sequence of the NURBS surface,
respectively. There is improved accuracy and speed of fitting NURBS surfaces [26,27]. Jinho
Song et al. divide unorganized points into boundary points and internal points using a
deep neural network, in order to facilitate more explicit parameterization of boundary
points during NURBS modeling [28]. Deming Kong et al. improved the current NURBS
free-form surface construction method based on discrete stationary wavelet transform.
The elevation image of the free-form surface point cloud and the elevation sequence of
the elevation image are obtained by using the grid method, and the feature points of the
free-form surface are extracted by using the discrete stationary wavelet transform; the
NURBS freeform surface can be constructed with high precision [29]. Tuo-Ran Wang et al.
used the weights iterative optimization method for the NURBS curve fitting, where the
geometric property of weight has been adopted to iteratively obtain the adjusting values of
the weights with the least square method. the proposed method reduces the maximum error
of the fitted surface from 0.01544 to 0.0066; the RMS error of the fitted surface is reduced
from 0.0066 to 0.0045. The fitting accuracy of NURBS can meet most of the demands of
reverse engineering [30].
However, there are many restrictions on NURBS’ surface in the application of reverse
engineering. It is not easy to deal with the joining of different NURBS surfaces. The surface
knot parameters and control point parameters also need to be adjusted according to the
characteristics of the modeling object. To solve these problems, for the reverse engineering
problem that the ideal CAD model is known, the information of the CAD model can be
combined to help build the reverse model of the real workpiece, in order to improve the
accuracy and efficiency of the reverse engineering. Thus, this paper presented a high
precision reverse engineering method based on the CAD model prior, taking a large engine
cabin as the example, combined with its design CAD model as the prior information of
reverse modeling; the reverse engineering of fusing CAD model information is carried out.
The rest of this paper is structured as follows: in the part of the methodology, the reverse
engineering method of fusing the CAD model information will be introduced in detail
from the following steps: point cloud model construction, processing of point cloud data,
and the construction and combination of the subregional surface. The experiment part
will discuss the quality of the reverse model with different parameter settings. The final
comments and results will be given in the results section.

3. Materials and Methodology

The process of reverse engineering by fusing CAD model information can be divided
into three steps: the point cloud model construction, processing of point cloud data, and
construction and combination of the subregional surface.
Machines 2022, 10, 905 4 of 18

The point cloud data collected by the 3D scanning equipment is preprocessed, includ-
ing point cloud denoising and point cloud feature point extraction. The purpose of point
cloud denoising is to remove noise and obtain more accurate point cloud data. The purpose
of point cloud feature point extraction is to save computing time for the subsequent point
cloud coarse registration. After data pretreatment and carrying out the registration on this
point cloud data and the CAD model of modeling target, we take the information such
as characteristic positions, and attributes of the CAD model as constraint conditions of
modeling, and segment the point cloud to different regions according to the characteris-
tics. A piece of the NURBS parametric surface is used to fit one region of the point cloud
data, and several NURBS surfaces obtained by the fitting are recombined according to the
constraints provided by the CAD model; finally, the digital space reconstruction model
of the real workpiece is obtained. In this framework, the above steps are realized by the
algorithm compiled by Matlab.
The flow chart of the reverse engineering method based on the CAD model prior is
shown in Figure 1.

Construct and
Construct point Process point combine
cloud model cloud data subregional
surfaces

Registrate point Determining Control

Point cloud data
denoising
cloud coordinate Points and Node
system Sequences
Workpiece
3D scan
reconstruction
point cloud
Segment point cloud model
Calculate normal vector Construct NURBS
data based on regional
and characteristic degree Surfaces
growing of CAD model

Divide ideal CAD

Extract point cloud Reorganize Surfaces
model surface mesh and
feature point set based on CAD model
grid offset

Figure 1. The flow chart of reverse engineering method based on prior CAD model.

3.1. Materials
We use the cabin section sample of an aerospace engine as the experimental body. As
shown in Figure 2, it is the nozzle of an aerospace engine cabin, with a length of 1.5 m and
a maximum diameter of 0.7 m. Since the manufacturing error and the deformation under
the influence of subsequent processes will affect the assembly accuracy of the workpiece,
to ensure the accuracy of the assembly simulation calculation, it is necessary to use the
real 3D model of the workpiece obtained by inverse modeling for subsequent simulation
calculations. As mentioned above, current reverse modeling methods are highly dependent
on the manual operations and take a long time to model. If the cabin sections with different
deformation characteristics in the mass production process are reversely modeled one
by one, it will seriously limit the efficiency of the production. The products produced
in the same batch have the same CAD model. Through the reverse modeling method in
this paper, the prior information in the CAD model only needs to be manually extracted
once in the production process to quickly complete the modeling of the same batch of
products, which can effectively reduce the requirements of the operator’s technical level,
and increase productivity.
Machines 2022, 10, 905 5 of 18

Figure 2. The experimental body: the cabin section sample of an aerospace engine.

The 3D data acquisition equipment used in the experiment is the Metra SCAN 3D laser
scanner and C-Track optical tracker from the CREAFORM, with a measurement resolution
of 0.025 mm, a measurement accuracy of 0.025 mm, and a measurement rate of 1,800,000
times/second. The measuring working range is 0.2–6 m, and the physical map of the
product is shown in the following Figure 3.

(a) (b)
Figure 3. The physical map of the product. (a) The Metra SCAN 3D laser scanner. (b) The C-Track
optical tracker.

3.2. Point Cloud Acquisition and Processing

Due to the influence of measurement conditions in the process of data acquisition, the
point cloud data obtained by 3D data acquisition equipment may contain noise points. To
remove the noise point data and obtain more accurate point cloud data, it is necessary to
denoise the original point cloud. The data points collected by high-resolution 3D scanning
equipment are very dense, and massive data points can accurately express the appearance
information of the object surface, but it also greatly limits the processing speed of the point
cloud data. Therefore, we can extract the feature points of the point cloud data to remove
some data points that have little change in the shape of the workpiece surface and use the
extracted feature point data to perform coarse registration of the point cloud to improve
the modeling efficiency.

3.2.1. Filter the Point Cloud

During the 3D scanning process, the parameter control of the equipment is completed
through the software named VXelements, and the resolution of the 3D scanning equipment
is adjusted to its full resolution, which is 0.025 mm. The Metra SCAN 3D laser scanner
Machines 2022, 10, 905 6 of 18

is clamped by the robotic arm and moves around the test piece uniformly to complete
the collection of the three-dimensional data. Then, the collected 3D data is exported in
the form of points. The initial point cloud collected by the 3D scanning equipment is a
collection of available points and noise points. The existence of noise points will lead to a
low accuracy of subsequent surface reconstruction, holes, and edges, and even lead to the
failure of the surface reconstruction algorithm. For this reason, point cloud noise reduction
must be performed. The sources of point cloud noise are divided into two categories: the
system error of the 3D scanning system and the environmental factors during acquisition.
The generated noise is mainly divided into the following types: scattered noise, outlier
gathering noise, internal repeated noise, and redundant noise. The core of the point
cloud denoising method is to construct a certain mathematical function according to the
distribution characteristics of the noise points and to screen and record the points for
unified removal according to the characteristics of the noise points. The noise removal
method used is as follows.
Statistical filtering: This method uses the 3σ theory in probability theory to calculate
the average value of the distance between a point and a point in its field. The calculated
value of the whole point cloud distance is counted, and the statistical result can form
a Gaussian distribution. Open the two function interfaces of mean and variance, and
remove the relevant point cloud data according to different input values. The denoising
method is suitable for uniformly sampled point sets, and the main type of noise removed is
scattered noise.
Radius filtering: First, take data as the radius, and draw a ball with this radius at each
point. When the number of points in the ball meets the requirements, the point will be
retained, otherwise, it will be deleted. This method will filter according to the density of
the point cloud data. When the density of a part of the point cloud data does not meet the
requirements, it will be deleted. This method mainly removes isolated outlier noise points.

3.2.2. Feature Point Extraction of Point Cloud Data

The resolution of high-precision 3D data acquisition equipment can reach more than
0.025 mm, and the data volume of point cloud obtained by 3D scanning of large workpieces
is very large, but in the process of processing some point cloud data, it is not necessary to
use all point cloud data to participate in the operation; the point cloud can be simplified by
extracting feature points from the point cloud. Using lightweight point cloud data in some
calculation processes, the calculation efficiency of the subsequent calculation process will
be improved.
The method based on PCA (principal components analysis) is used to calculate the
normal vector of the point cloud, and then extract the point cloud feature points, which can
effectively reduce the number of points in the point set and retain the shape of the feature
points of the point cloud. The main implementation methods are as follows.
The covariance matrix is created according to the k-nearest neighbors of the point
cloud, and the normal information of the point cloud is calculated by PCA. The normal
direction of the point cloud is redirected according to the direction of the point cloud mean
neighbor points.
For some point pi and k neighboring points pi1 , pi2 , pi3 , . . . , pik , compute its mean
neighborhood point pm :

1
pm = ( p + pi2 + pi3 + . . . + pik ). (1)
k i1
The normal vector ni direction of pi is defined as

ni , ( pi − p m ) · ni > 0
(2)
− ni , ( pi − p m ) · ni < 0
Machines 2022, 10, 905 7 of 18

Calculate the point feature degree according to the normal vector information and
extract the point cloud feature points; reducing the redundancy degree of the point cloud
model, the characteristic degree of pi is calculated as follows:

1
σi = (|n · n | + |ni2 · ni | + |ni3 · ni | + . . . |nik · ni |) (3)
k i1 i
The average feature degree of the point cloud is calculated as follows:
n
1
σ=
n ∑ σi (4)
i =1

Marking the points of which the point characteristic degree is less than the point cloud
mean characteristic degree, as the point cloud characteristic points, to obtain a characteristic
point set p0 of the point cloud 
P0 = pi | σi < σ (5)

3.3. Processing of Point Cloud Data

3.3.1. Registration of CAD Model and Point Cloud Data
In the process of real product modeling, the purpose of point cloud model registration
is to transform the coordinate system of the CAD model and point cloud data. Thus, the
known feature attribute data such as feature location and feature type carried by the CAD
model can be transferred to the point cloud data to be processed through the consistent
coordinate location relationship, which can be used as the constraint condition of the
modeling process. The first step of the registration process is to determine the target point
cloud and the point cloud to be registered. In this framework, as shown in Figure 4, the
point cloud to be registered is the point cloud data obtained by the 3D scanning equipment
after preliminary data processing, and the target point cloud is the data point sampled
from the surface of the ideal CAD model.

(a) (b) (c)

Figure 4. This is target point cloud and the point cloud to be registered. (a) The target point cloud.
(b) The point cloud to be registered. (c) The initial coordinate system of two point clouds.

The registration process is divided into two stages: coarse registration and fine regis-
tration. The coarse registration transformation based on local feature description is carried
out by using the feature point set of the point cloud to improve the registration efficiency,
and the ICP (iterative closest Point) fine registration transformation is carried out by using
all the point cloud data after the coarse registration to ensure registration accuracy.
Firstly, the feature point sets Q0 and P0 of the point cloud to be registered Q and the
target point cloud P are extracted, respectively, according to the normal vector information
of the point cloud. Then, the local feature descriptors of the point cloud are constructed.
The purpose of constructing the feature descriptor is to express the local features of the
point cloud on the mathematical level. This mathematical expression is not affected by
the global coordinate system where the point cloud is located and has the invariance of
rigid transformation, which can describe the local features of the point cloud. The HoPPF
Machines 2022, 10, 905 8 of 18

(histograms of point pair features) descriptor is selected as the local feature descriptor of
the point cloud, and the RANSAC ( random sample consensus) method is used to match
the descriptors to determine the coarse registration relationship between the point cloud to
be registered and the target point cloud. Finally, the ICP algorithm is used to complete the
fine registration of the complete set of point clouds, wherein the transformation relation
is calculated as follows: The transformation matrix of point cloud coarse registration is
calculated from the corresponding matching key points KPS = {qs1 , qs2 , qs3 , . . .} T and
KPT = { pt1 , pt2 , pt3 , . . .} T in the point set to be registered and the target point set.
Initial location matrix:
 
xqs1 yqs1 zqs1
 xq
 s2 yqs2 zqs2 

KPS = {qs1 , qs2 , qs3 , . . .} T =  xq (6)
 s3 yqs3 zqs3 

.. .. ..
. . .

Target location matrix:

 
x pt1 y pt1 z pt1
x p y pt2 z pt2 
 t2
KPT = { pt1 , pt2 , pt3 , . . .} T =  x p (7)

 t3 y pt3 z pt3 

.. .. ..
. . .

Accord to the initial position matrix and the target position matrix, the transformation
matrixes R a and Ta in the coarse registration can be converted into a least square problem

3
( R a , Ta ) = arg min ∑ k( R a · kpsi + Ta ) − kpti k2 (8)
i =1

After decentering the transformation and SVD (singular value decomposition), the
transformation matrices R a and Ta in the coarse registration can be obtained. Performing
the coordinate transformation on the point cloud Q to be registered is conductedby using
the transformation relation obtained by the coarse registration:

Q0 = R a Q + Ta (9)

The transformed target point cloud Q0 is used as the initial value of fine registration.
After the coarse registration transformation, all the point cloud data are used for ICP
algorithm-based fine registration transformation, the closest point of the point cloud to
be registered is searched in the target point cloud as a matching point, and the minimum
sum of the distances of all the matching points is used as an iterative registration target; a
rotation transformation matrix R and a translation transformation matrix T are calculated
after multiple iterative transformations, performing precise registration transformation on
the 3D point cloud data to be registered. The CAD model and the point cloud data of the
workpiece will convert into the same coordinate system.

3.3.2. Point Cloud Segmentation Based on Region Growing and Clustering

The initial point cloud features obtained by 3D scanning equipment are highly coupled,
and to ensure the modeling accuracy, it is necessary to segment the point cloud model
according to the features. Divide the overall point cloud of the model into point cloud
regions belonging to different regions and reconstruct the model in different regions. The
region growing algorithm is used to segment the point cloud according to the characteristics.
Starting from the seed point set, the region growing algorithm is used to add the points
with similar characteristics nearby, and the region growing is continuously carried out to
obtain the largest growing region on the premise of meeting the error requirement. The
Machines 2022, 10, 905 9 of 18

process of point cloud segmentation based on the region growing algorithm is shown in
Figure 5.

Start

No
Select Yes
Take points in Choose an Mark the Remove the
Yes No unlabeled Whether the points
Are all points Select new the appropriate point and seed point
neighborhood in the neighborhood
marked as split? seed point neighborhood feature add to the from the
points of the are all processed?
in turn measure seed stack seed stack
seed point

End Take the

No Is the
next seed
seed point stack
point in the
empty?
stack
Yes

This round of regional

growth is complete

Figure 5. The flow chart of segmentation of point cloud based on region growing method.

Segmentation of the point cloud based on the region growing is a segmentation

algorithm based on the surface. The algorithm extends outward from the seed point, and
according to the angle between the normal vectors of the point cloud, the adjacent points
that meet the characteristic change range are classified into one class to form different
regional divisions. To some extent, the selection of seed points determines the effect of the
final point cloud segmentation. Therefore, in order to improve the efficiency and accuracy
of the point cloud segmentation, the CAD model is used as the prior condition of point
cloud segmentation, and the CAD model is divided into different feature regions, as shown
in Figure 6. Some data points in each region are selected as the initial seed points to
segment the whole point cloud data. Divide the CAD model into different feature regions
and choose suitable points in each region as the growth seed is completed by manual
operation. It is a process of extracting prior information from CAD models.

Figure 6. The CAD model is divided into feature regions, and some data points are selected as
seed points.

The main steps of the point cloud segmentation method based on region growing
are as follows: (1) Select the characteristic measure. The curvature information of the
point cloud is used as a feature measure to determine the region similarity and the growth
process of the constraint region. (2) Calculate the characteristic measure. The method
traverses all the point cloud data and calculates the feature vector estimation of each point,
namely the curvature information of the point cloud. (3) Select the seed point. In this
framework, different features of the CAD model are used to determine the initial seed
point. (4) Region segmentation. Starting from a first seed point in the initial seed point
Machines 2022, 10, 905 10 of 18

set, judging whether points nearby the first seed point in the 3D space have the same or
more similar characteristics, and if so, adding the point into a seed point stack until all
neighboring points are processed. (5) Select a new seed point. Select the next point in the
initial seed point set as a new seed point, and jump to step (3) to repeat the region growing
processing until all points in the point cloud data set are processed. Figure 7 shows the
results of point cloud segmentation.

(a) (b)
Figure 7. Point cloud segmentation. (a) The initial state of the point cloud. (b) The result of point
cloud segmentation.

3.3.3. Point Cloud Data Extraction Based on Grid Point Offset

In the process of surface construction from point cloud data, the quality of point cloud
data directly affects the quality of the subsequent surface construction. To obtain a smooth
surface, it is necessary to smooth the modeling point cloud data. For the disordered point
cloud data, we need to determine the order and position of the point cloud data before data
smoothing. The current point cloud sorting methods have the problems of large amounts
of computation and low efficiency.
To solve the problem of the low efficiency of point cloud sorting and select the appro-
priate point cloud data points from the massive point cloud data for subsequent surface
modeling, the grid offset method is used to sort and filter the point cloud. We divide
regular grid points on the surface of the model according to the ideal CAD model, take the
grid points as reference points, and carry out the grid point offset according to the point
cloud data to obtain the grid point data of the deformation surface type value of the real
surface of the product.
The mesh offset algorithm is as follows: for each feature of the CAD model, a mesh
model is established separately, and the mesh size is defined according to the accuracy of
the target model, as shown in Figure 8. Each grid point of the CAD grid is connected with
the point cloud data divided by feature through a unified coordinate system relationship.
By finding the nearest neighbor of the CAD grid point in the point cloud model of the
deformed workpiece, determine the offset distance of the reference grid points, offset all
the reference grid points to obtain the shape value grid points of the deformed workpiece,
and use these grid points as the shape value points to construct the NURBS surface of the
real workpiece.
Smoothing the model value points of the NURBS curved surface of the workpiece
according to the u direction and the v direction of a grid, respectively, eliminates the data
fluctuation caused by a 3D scanning data acquisition device and obtains a group of model
data points for constructing the NURBS curved surface.
Machines 2022, 10, 905 11 of 18

(a) (b) (c)

(d) (e) (f)

Figure 8. The process of grid point offset. (a) The grid point of CAD model. (b) The points of point
cloud data. (c) Each grid point of the CAD grid is connected with the point cloud data through a
unified coordinate system relationship. (d) Nearest neighbor of the CAD grid point in the point cloud
model. (e) Grid point of the deformed workpiece. (f) A NURBS of real workpiece constructed by
shape value grid point.

3.4. Reconstruction Surface from Point Cloud by Subarea

3.4.1. Determination of NURBS Surfaces
NURBS is determined by u direction order, v direction order, u direction control point,
v direction control point, u direction knot points and v direction knot points. The key to
parametric surface reconstruction in the process of reverse engineering is how to get the
control points of the surface according to the measured point cloud data. Once the control
points are determined, the chord length accumulation method is used to determine the
knot points, and the reconstructed surface can be easily obtained according to the formula.
A NURBS surface of degree p in the u direction and degree q in the v direction s(u, v) can
be described by the following formula:
n m

x (u, v)
 ∑ ∑ Ni,p (u) Nj,q (v)Wi,j Pi,j n m
i =0 j =0
s(u, v) =  y(u, v)  = n m = ∑ ∑ Ri,j (u, v) Pi,j (10)
z(u, v) ∑ ∑ Ni,p (u) Nj,q (v)Wi,j i =0 j =0
i =0 j =0

Ni,p (u) Nj,q (v)Wi,j

Ri,j (u, v) = n m (11)
∑ ∑ Nk,p (u) Nl,q (v)Wk,l
k =0 l =0

where Pi,j represents the control vertices of the NURBS surface; Wi,j : weight factor associ-
ated with the control vertex; n + 1: number of control points in u direction; m + 1: number
of control points in v direction; p: the degree of the NURBS surface in the u direction; q: the
degree of the NURBS surface in the v direction; Ni,p (u): the p-th order basis function on
the knot vector of u, which is calculated by the knot vector of u according to the recursive
formula; Nj,q (v): the q-th order basis function on the knot vector of v, which is calculated by
the knot vector of v according to the recursive formula; Ri,j (u, v): bivariate rational basis
functions; Ni,p (u),Nj,q (v) are, respectively, calculated by the recurrence formula of B-spline
Machines 2022, 10, 905 12 of 18

curve. The basis functions on the knot vector are calculated according to the B-spline de
Boor-Cox recursive definition as follows:

1 i f ui ≤ u ≤ ui + 1

 Ni,0 (u) = 0

 




 others
u − ui u i + k +1 − u

(12)

 Ni,k (u) =
 Ni,k−1 (u) + Ni+1,k−1 (u)


 u i + k − u i u i + k +1 − u i +1
set 0 = 0



0

1 i f vi ≤ v ≤ vi + 1


  N
 i,0
 ( v ) =



 0 others
v − v v i + k +1 − v


 Ni,k (v) = i (13)
 Ni,k−1 (v) + N (v)


 vi +k − vi vi+k+1 − vi+1 i+1,k−1
set 0 = 0



0
u = [u0 , u1 , . . . . . . , un+k+1 ] and v = [v0 , v1 , . . . . . . , vn+k+1 ] represent the knot vectors in the
u and v directions respectively.
Therefore, to construct a NURBS surface, it is necessary to first determine the order
of the surface p in the u directions and q in the v direction, the knot sequence in the u and
v directions, and then inversely calculate the control points of the NURBS surface and
determine the corresponding weights according to the point cloud data points in different
regions, to determine a NURBS surface.

3.4.2. NURBS Surface Construction and Reconstruction

Because the construction of the NURBS surface requires that the number of control
points of each parameter curve in the u direction and v direction of the surface should be
consistent, the number of control points of each parameter curve in the construction of the
NURBS surface should also be consistent, but the scattered and disordered characteristics of
point cloud data brings difficulties to the determination of control points. The point cloud
data sequence extracted by the grid point offset method can realize the batch equal-length
processing and arrangement of the point cloud data according to u and v parameter curve
directions of the NURBS surface; we obtain the grid points of the single characteristic
surface type value of the real surface of the product, and further inversely calculate the
control points of the NURBS surface. Parametric surfaces in 3D space are represented in
the following format:
s(u, v) = ∑ bi f i (u, v), (u, v) ∈ D (14)
i∈ I

Surface features si (u, v) at different positions of the workpiece are constructed by

using the offset point cloud of the mesh segmented by regions in the Section 3.3.3. Then,
we split and recombined the surface features according to the topological relationship in
the CAD model. The feature splitting and recombination constraints determined by the
topological relationship of the CAD model are shown in the Figures 9 and 10:
Machines 2022, 10, 905 13 of 18

Figure 9. The feature splitting constraints determined by the topological relationship of the
CAD model.

Figure 10. The feature recombination constraints determined by the topological relationship of the
CAD model.

Then, we obtain the surface model of reverse engineering, as shown in Figure 11.

Figure 11. The result of surface modeling.

4. Experiment and Analysis

In the process of the assembly and manufacture of the aerospace engine, the deforma-
tion of the engine cabin caused by the process of manufacturing and loading brings great
difficulties to the assembly of the engine cabin, and seriously affects the quality stability of
engine products. The assembly simulation method can be used to simulate different as-
sembly process parameters, calculate the interference, collision, stress, and other assembly
conditions in the assembly process, determine the best assembly control parameters, and
then ensure the assembly quality of the cabin.
To simulate the assembly process of real workpieces, 3D scanning equipment was
used to collect the surface information of real deformed workpieces, and the real shape
feature model of the deformed cabin section of the engine was constructed to obtain the
high-precision shape model of the workpiece, which was used to simulate the assembly
process. To improve the efficiency and accuracy of the assembly simulation process, and
Machines 2022, 10, 905 14 of 18

then ensure the production efficiency and assembly quality of the aerospace engine, a
modeling method is presented in this paper, in which the CAD design model is used as the
prior condition, and the reverse engineering of the aerospace engine cabin is carried out to
improve the efficiency and accuracy of modeling. In this process, the size of the regular grid
points of the CAD model will affect the accuracy and speed of modeling; thus, this chapter
will discuss the impact of parameter values on the quality and efficiency of modeling. The
experiments will be carried out in the Matlab environment, and the modeling results will
be displayed with the help of NX1926.
First, discrete the CAD model to a point cloud. This process is completed in Meshlab.
Then, the different characteristics of the CAD model are distinguished and some points are
defined as the seed points of the area segmentation. The point cloud data of the CAD model
and the coordinates of the seed points are used as the input Matlab for the next step of
processing; these prior processes need to be conducted manually at present. After that, use
the program algorithm written in Matlab to complete the processing of the scanned point
cloud data, the registration of CAD model point cloud and scanned data, and complete the
point cloud according to the prior input seed point data region growing segmentation and
the fitting of the NURBS parametric surfaces. The surface obtained by fitting is in an IGES
format, which is exported to NX for display. When fitting a NURBS surface, the mesh size
selected when dividing the CAD will directly affect the number of point clouds involved
in the surface reconstruction, thus affecting the modeling quality of the surface. Taking
the resolution of the point cloud data as the basic unit, changing the size of the CAD grid,
and recording the modeling time and modeling error, the final modeling result is shown in
Figures 12 and 13, and the specific analysis data results are shown in Figure 14 and Table 1.
Respectively, take the size of the CAD grid as several times the resolution of the
point cloud, perform the feature division and point cloud segmentation of the CAD model
according to the method described in the part of methodology, find the nearest point cloud
data of the CAD grid, and carry out the CAD model network. Offset the grid points to
obtain the point cloud data grid points, use the grid points of the point cloud to construct
the NURBS surface, and reorganize all the NURBS surfaces according to the topological
relationship of the CAD model to obtain the modeling result.

(a) (b) (c)

(d) (e) (f)

Figure 12. The process of grid point offset. (a) The CAD grid points with a resolution of 6 mr.
(b) Nearest neighbors of the CAD grid point in the point cloud model. (c) The NURBS of real
workpiece constructed by grid point. (d) All of the CAD model grid point. (e) The offset data points
according grid point. (f) Modeling results.
Machines 2022, 10, 905 15 of 18

Figure 12 is a schematic diagram of the CAD grid map generated in the modeling
process when the CAD grid size is 6 mr. The point cloud data are selected according to
the principle of the nearest neighbors of grid points and the point cloud construction of
the NURBS surface, as the grid of the overall CAD model. Divide and select the point
cloud data based on the closest point of the grid, and the final modeling result is shown in
this figure.
When the CAD grid size is refined to 0.4 mr, the CAD grid map is generated during the
modeling process, the point cloud data is selected according to the principle of the nearest
grid point, and the point cloud constructs a schematic diagram of the NURBS surface, the
grid of the overall CAD model, the point cloud data, is divided and selected according to
the nearest point of the grid point, and the final modeling results are shown in Figure 13.

(a) (b) (c)

(d) (e) (f)

Figure 13. The process of grid point offset. (a) The CAD grid points with a resolution of 0.4 mr.
(b) Nearest neighbors of the CAD grid point in the point cloud model. (c) The NURBS of real
workpiece constructed by grid point. (d) All of the CAD model grid point. (e) The offset data points
according grid point. (f) Modeling results.

From the figure, it can be observed that the number of CAD grid points and the points
used to construct the surface increased. The points cloud data are also denser, and the
surface quality of the final inverse model is smoother.
Set different CAD mesh sizes and record the modeling error and required time; the
final results are shown in Figure 14 and Table 1 below. The modeling error in the table is
calculated as: sampling the reconstructed surface, finding the closest point in the denoised
point cloud data for each sampling point as a matching point pair, and calculating the
maximum matching point pair distance as the final modeling error. The time required for
modeling is the time after point cloud registration and point cloud segmentation based
on features, including dividing the ideal grid points, finding the closest point of the point
cloud, and surface modeling.
Machines 2022, 10, 905 16 of 18

Figure 14. Model error and modeling time at different mesh sizes.

Table 1. Model error and modeling time at different mesh sizes.

Mesh Size (mr) Maximum Error (mr) Time (s)

0.4 1.36 91.03
2 1.56 42.72
4 2.18 30.88
6 2.97 27.16
8 4.02 24.71
10 4.70 23.72
12 5.31 23.22
14 5.62 22.46
16 5.66 22.14
18 5.73 23.65
20 - -

It can be observed from Figure 14, obtained from the experiment, that with the increase
in the mesh size, the time required for modeling is gradually reduced, and the model error
is also gradually increased. After the mesh size reaches 14 mr, the maximum modeling
error remains at about 5.6 mr. When the mesh continues to increase to more than 20 mr, due
to the limitation of point cloud data resolution, the selected modeling data points are too
few to construct a NURBS surface of more than three orders. It can also be observed from
Figures 12 and 13 that when the mesh size is 6 mr, the surface joints of the reconstructed
model have obvious wrinkles, and when the mesh size is 0.4 mr, the constructed NURBS
surface is smooth and flat. Because the influence of the mesh size on the modeling computa-
tion is exponential, the modeling time decreases sharply when the mesh size increases from
approximately 0mr to 4 mr and tends to be stable when the mesh size is larger than 4 mr.
The point cloud resolution obtained by the 3D scanning equipment used in the experiment
is 0.025 mm. When the CAD grid resolution is set to 0.4 mr, the error of the reconstructed
model is 1.36 mr, which is less than 0.04 mm. The model with this accuracy can be used as
a cabin assembly. Simulated computational models help improve the efficiency and quality
of cabin assembly.
Machines 2022, 10, 905 17 of 18

5. Conclusions
In this paper, a CAD model as a prior information of reverse engineering method is
presented. Through the registration of the CAD model and point cloud data to complete
the consistency conversion of the two model coordinate systems, the CAD model provides
the position information of each feature and the connection and constraint information
between features as a prior condition of the modeling process. Workpiece point cloud data
is segmented according to different feature regions, and a NURBS surface is used to recon-
struct a single feature surface and reconstruct multiple feature surfaces. Compared with the
conventional manual modeling depending on manual operation and the conventional reg-
ular feature modeling depending on feature recognition, the method is suitable for reverse
engineering with drawings of workpieces with slight irregular deformation compared with
a CAD design model, which is particularly suitable for constructing a batch product of real
models of the same kind of workpieces and can greatly improve the modeling efficiency
and the modeling precision stability. The method presented in this paper still has certain
limitations, and the generality of products with complex combination relationship and
complex modeling surfaces features needs to be further determined through experiments.
The influence of the mesh size defined by the CAD model on the model construction still
needs to be experimented with 3D data of various scenes. In the next step, we will change
the parameters of the 3D data acquisition equipment and the experimental environment
to verify the robustness of this method in different environments, models and equipment
parameters, and further clarify the usage scenarios of the method proposed in this paper
and the selection method of specific parameters.

Author Contributions: Conceptualization, L.F. and J.W.; methodology, L.F.; software, L.F.; validation,
J.W. and X.Y.; formal analysis, X.Y.; investigation, L.F.; writing—original draft preparation, L.F.;
writing—review and editing, L.F.; supervision, Z.X.; project administration, Z.X. All authors have
read and agreed to the published version of the manuscript.
Funding: This research was funded by the National Natural Science Foundation of China (92067205)
and Central leading local science and Technology Development Fund Project (2022JH6/100100014).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The data presented in this study are available upon request from the
first author. The data are not publicly available due to intellectual property protection.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Turek, P.; Filip, D.; Przeszłowski, Ł.; Łazorko, A.; Budzik, G.; Snela, S.; Oleksy, M.; Jabłoński, J.; S˛ep, J.; Bulanda, K.; et al.
Manufacturing Polymer Model of Anatomical Structures with Increased Accuracy Using CAx and AM Systems for Planning
Orthopedic Procedures. Polymers 2022, 14, 2236. [CrossRef] [PubMed]
2. Ullah, A.S.; Watanabe, M.; Kubo, A. Analytical point-cloud based geometric modeling for additive manufacturing and its
application to cultural heritage preservation. Appl. Sci. 2018, 8, 656.
3. Kwon, K.; Mun, D. Iterative offset-based method for reconstructing a mesh model from the point cloud of a pig. Comput. Electron.
Agric. 2022, 198, 106996. [CrossRef]
4. Popis, ter, F.; Popescu, D.; Păcurar, A. Mathematical Approach in Complex Surfaces Toolpaths. Mathematics 2021, 9, 1360.
[CrossRef]
5. Wang, Z.; Shi, Y.; Wang, X. NURBS Function Closed-Loop Mapping Trajectory Planning of Serial Robotic Plasma Cladding for
Complex Surface Coatings. Int. J. Adv. Manuf. Technol. 2022. [CrossRef]
6. Geng, Z.; Bidanda, B. Review of reverse engineering systems–current state of the art. Virtual Phys. Prototyp. 2017, 12, 161–172.
[CrossRef]
7. Vitali, A.; Rizzi, C. Acquisition of customer’s tailor measurements for 3D clothing design using virtual reality devices. Virtual
Phys. Prototyp. 2018, 13, 131–145. [CrossRef]
8. Jančar, L.; Začal, J.; Folta, Z. Reverse Engineering of the Gearbox Case. In Current Methods of Construction Design; Springer:
Berlin/Heidelberg, Germany, 2020; pp. 79–87.
Machines 2022, 10, 905 18 of 18

9. Li, F.; Longstaff, A.P.; Fletcher, S.; Myers, A. Rapid and accurate reverse engineering of geometry based on a multi-sensor system.
Int. J. Adv. Manuf. Technol. 2014, 74, 369–382. [CrossRef]
10. Geng, Z.; Sabbaghi, A.; Bidanda, B. Reconstructing original design: process planning for reverse engineering. IISE Trans. 2022,
1–14. [CrossRef]
11. Helle, R.H.; Lemu, H.G. A case study on use of 3D scanning for reverse engineering and quality control. Mater. Today: Proc. 2021,
45, 5255–5262. [CrossRef]
12. Zhang, K.; Qiao, S.; Wang, X.; Yang, Y.; Zhang, Y. Feature-preserved point cloud simplification based on natural quadric shape
models. Appl. Sci. 2019, 9, 2130. [CrossRef]
13. Sarawit, A.; Kim, Y.; Bakker, M.; Peköz, T. The finite element method for thin-walled members-applications. Thin-Walled Struct.
2003, 41, 191–206. [CrossRef]
14. Hannaby, S. A mapping method for mesh generation. Comput. Math. Appl. 1988, 16, 727–735. [CrossRef]
15. Wilson, J.; Topping, B.H.V. Parallel adaptive tetrahedral mesh generation by the advancing front technique. Comput. Struct. 1998,
68, 57–78. [CrossRef]
16. Topping, B.; Cheng, B. Parallel and distributed adaptive quadrilateral mesh generation. Comput. Struct. 1999, 73, 519–536.
[CrossRef]
17. Romero-Jarén, R.; Arranz, J. Automatic segmentation and classification of BIM elements from point clouds. Autom. Constr. 2021,
124, 103576. [CrossRef]
18. Zhang, M.; Anwer, N.; Mathieu, L.; Zhao, H. A discrete geometry framework for geometrical product specifications. In
Proceedings of the 21st CIRP Design Conference, Kaist, Korea, 1 January 2011.
19. Zhang, M.; Anwer, N.; Stockinger, A.; Mathieu, L.; Wartzack, S. Discrete shape modeling for skin model representation. Proc. Inst.
Mech. Eng. Part B J. Eng. Manuf. 2013, 227, 672–680. [CrossRef]
20. Pernot, J.P.; Guillet, S.p.; Le´on, J.C.; Facidieno, B.; Giannini, F. A new approach to minimisations for shape control during
free-form surface deformation. In Proceedings of the International Design Engineering Technical Conferences and Computers
and Information in Engineering Conference, Chicago, IL, USA, 2–6 September 2003; Volume 37009, pp. 711–720.
21. Wells, L.J.; Dastoorian, R.; Camelio, J.A. A novel NURBS surface approach to statistically monitor manufacturing processes with
point cloud data. J. Intell. Manuf. 2021, 32, 329–345. [CrossRef]
22. Takizawa, K.; Kostov, N.; Puntel, A.; Henicke, B.; Tezduyar, T.E. Space–time computational analysis of bio-inspired flapping-wing
aerodynamics of a micro aerial vehicle. Comput. Mech. 2012, 50, 761–778. [CrossRef]
23. Park, H. An approximate lofting approach for B-spline surface fitting to functional surfaces. Int. J. Adv. Manuf. Technol. 2001,
18, 474–482. [CrossRef]
24. Yang, J.; Liu, Y.; Wang, X.; Wu, H. An improved steady inverse method for turbomachinery aerodynamic design. Inverse Probl.
Sci. Eng. 2017, 25, 633–651. [CrossRef]
25. Chouychai, B. Point cloud data reduction with tangent function method for nurbs curve and surface fitting. In Proceedings
of the 2015 7th International Conference on Knowledge and Smart Technology (KST), Chonburi, Thailand, 28–31 January 2015,
pp. 137–142.
26. Rodríguez, J.A.M. Efficient NURBS surface fitting via GA with SBX for free-form representation. Int. J. Comput. Integr. Manuf.
2017, 30, 981–994. [CrossRef]
27. Tian, X.; Kong, L.; Kong, D.; Yuan, L.; Kong, D. An Improved Method for NURBS Surface Based on Particle Swarm Optimization
BP Neural Network. IEEE Access 2020, 8, 184656–184663. [CrossRef]
28. Song, J.; Lee, J.; Ko, K.; Kim, W.D.; Kang, T.W.; Kim, J.Y.; Nam, J.H. Unorganized point classification for robust NURBS surface
reconstruction using a point-based neural network. J. Comput. Des. Eng. 2021, 8, 392–408. [CrossRef]
29. Kong, D.; Tian, X.; Kong, D.; Zhang, X.; Yuan, L. An improved method for nurbs free-form surface based on discrete stationary
wavelet transform. IEEE Access 2020, 8, 67015–67023. [CrossRef]
30. Wang, T.R.; Liu, N.; Yuan, L.; Wang, K.X.; Sheng, X.J. Iterative Least Square Optimization for the Weights of NURBS Curve. Math.
Probl. Eng. 2022, 2022, 5690564. [CrossRef]