Feature Selection-Based Method for Scaffolding Assembly Quality Inspection Using Point Cloud Data

Abstract

The stability of scaffolding structures is crucial for quality management in construction. Currently, scaffolding assembly quality monitoring relies on visual inspections performed by designated on-site personnel, which are highly subjective, inaccurate, and inefficient, hindering the advancement of intelligent construction practices. This study proposes an automated method for scaffolding assembly quality inspection using point cloud data and feature selection algorithms. High-precision point cloud data of the scaffolding are captured by a Trimble X7 3D laser scanner. After registration with the forward design model, a 2D slicing comparison method is developed to measure geometric dimensions with an accuracy controlled within 0.1 mm. The collected data are used to build an SVM model for automated assembly quality inspection. To combat the curse of dimensionality associated with high-dimensional data, an optimized genetic algorithm is employed for the dimensionality reduction in the raw sample data, effectively eliminating data redundancy and significantly enhancing convergence speed and classification accuracy of the detection model. Case studies indicate that the proposed method can reduce feature dimensionality by 70% while simultaneously improving classification accuracy by 13.9%. The proposed method enables high-precision automated inspection of scaffolding assembly quality. By identifying the optimal feature subset, the method differentiates the priority of various structural parameters during inspection, providing insights for optimizing the quality inspection process.

1. Introduction

Scaffolding structures are widely recognized as essential temporary structures in the construction industry, which are crucial for supporting various building projects [1,2]. As one of the primary hazards in scaffolding operations, scaffolding collapse not only leads to structural damage and significant financial losses for ongoing projects but also poses a critical risk to the lives of construction workers. According to data from the Occupational Safety and Health Administration (OSHA), scaffolding operations involve approximately 65% of the total workforce in the construction industry. However, due to poor management of scaffolding work sites, these operations result in approximately 60 fatalities, 4500 injuries, and over USD 90 million in losses annually [3]. In China, based on data compiled by the Ministry of Housing and Urban–Rural Development (MOHURD) in 2020, scaffolding collapse accidents accounted for a significant proportion of 21.74% among the 23 major incidents related to housing and municipal construction projects. The primary cause of scaffolding collapse is the insufficient attention given to the safety monitoring of scaffolding assembly quality [2]. Currently, the safety monitoring of scaffolding still relies on traditional visual inspections performed by individuals with specific expertise at construction sites. However, this method is subjective and prone to errors, as it heavily depends on the visual judgment of individuals, which may result in the overlooking of critical safety risks. Furthermore, the large scale of scaffolding structures in construction projects exacerbates the labor-intensive nature and inefficiency of traditional visual inspections. The limited capacity of on-site management personnel to continuously monitor multiple scaffolding structures during the construction period further underscores the limitations of traditional scaffold safety monitoring practices [3,4].

Scholars have been exploring automated methods for scaffolding safety monitoring to address safety concerns associated with manual visual inspections. These methods fall into two categories: contact-based detection and non-contact-based detection. Contact-based detection utilizes sensors, radio frequency identification (RFID) tags, cameras, and other devices, employing Internet of Things (IoT) technology for real-time online scaffolding monitoring. This online monitoring approach offers advantages over traditional visual inspections, including lower risks, reduced costs, and improved efficiency [5]. However, implementing sensor networks for inspection may incur additional maintenance and management costs, including the installation of extra wires and cables at the construction site, which could introduce additional safety hazards [3].

With advancements in information technology, scholars have increasingly focused on non-contact detection methods based on computer vision, 3D laser scanning, and other technologies. Currently, computer vision-based methods are primarily used for assessing the quality of individual scaffolding components [6], with research on the inspection of overall assembly quality remaining limited. Meanwhile, 3D laser scanning technology has been widely applied to assess the assembly quality of structures, such as prefabricated buildings [7], bridges [8], and steel frameworks [9,10,11]. By aligning and comparing the as-built point cloud with the as-designed BIM model, this technology facilitates the evaluation of geometric dimension errors in actual building components. However, its application in scaffolding remains relatively limited.

In response to the issues of high subjectivity, labor intensity, and low efficiency associated with traditional manual visual inspection methods for scaffolding, an automated scaffolding assembly quality inspection method based on point cloud data and feature selection algorithms is proposed. This research aims to assess the impact of various structural parameters on scaffolding stability using feature selection algorithms, thereby identifying key factors in assembly quality inspections. Ultimately, a support vector machine (SVM) model is developed to enable the automated detection of scaffolding assembly compliance. This study employs laser scanning technology to acquire point cloud data of the scaffolding and compares it with the forward design model to achieve high-precision measurements of the scaffolding’s geometric structural parameters. To address the challenge of high-dimensional feature data resulting from the numerous scaffolding components, an optimized genetic algorithm (GA) is utilized for dimensionality reduction. It effectively circumvents the curse of dimensionality (COD) and improves the performance of the classification model. The specific contributions of this study are as follows:

• To measure scaffolding structural parameters, a 2D slicing comparison method based on point cloud data is proposed. This method replaces manual inspections with high-precision and automated 3D laser scanning technology, thereby significantly enhancing inspection accuracy and efficiency.
• By employing an improved GA to reduce the dimensionality of scaffolding structural parameter data, this study not only improves the classification accuracy of the quality inspection model but also identifies the feature subset with the most significant impact on overall assembly quality. This approach distinguishes the priority of different structural parameters during inspections, providing a data-driven basis for improving the inspection process and enhancing efficiency.

2. Literature Review

2.1. Scaffolding Safety Management

In scaffolding safety management, the involvement of designated on-site personnel is essential. These individuals, duly trained in safety protocols [3], are responsible for performing routine and comprehensive visual inspections to evaluate the structural integrity and safety of scaffolding systems [12]. Their primary responsibility is to identify potential risks and instances of non-compliance with safety regulations, thereby enabling prompt corrective actions. However, assessing the safety status of scaffolding through manual visual inspections, which heavily rely on subjective judgments, has limitations in efficacy and reliability [1,13]. Furthermore, manual inspections are labor-intensive, particularly given the extensive scale of scaffolding on construction sites, leading to time-consuming and physically demanding processes [14]. Moreover, relying solely on individual management personnel is insufficient for continuously monitoring multiple scaffolding structures during the construction period on construction sites [3].

To address the limitations of traditional scaffolding safety management, scholars have extensively explored automated methods for scaffolding safety monitoring. A common approach involves establishing monitoring systems based on IoT technology, utilizing various devices, such as RFID tags, sensors, and cameras. RFID tags facilitate the effective tracking and management of scaffolding components, allowing for the identification of aging parts that need replacement [15]. However, this method alone cannot automate the analysis of load-bearing conditions and structural deformations in assembled scaffolding structures. Consequently, scholars have turned to sensor-based solutions for real-time monitoring of scaffolding stress and structural integrity. Sensors allow for continuous monitoring of load-bearing conditions in scaffolding structures by measuring critical parameters, such as load distribution, strain, and displacement. As such, they provides valuable insights into the structural integrity and performance of the scaffolding [16]. Integrating trained machine learning or deep learning algorithms into the monitoring system enables the automatic assessment of scaffolding safety status [3,17]. This allows for the early detection of potential safety risks, enabling timely interventions and preventive measures. However, the use of sensors may involve additional costs for maintenance and management due to the installation of wiring and cables on-site [3].

Therefore, scholars have endeavored to develop non-contact scaffolding monitoring methods. For instance, Feng et al. [18] focused on visual sensing technology and used industrial cameras to capture continuous images or videos of temporary structures. Relevant algorithms are employed to extract real-world spatial information and calculate deflection and strain. However, the visibility of visual sensors can be obstructed by the complex and dynamically changing construction environment, as well as by scaffolding boards and safety nets, thereby limiting the method’s applicability. Kim et al. [6] proposed a scaffolding component quality assessment system based on computer vision technology. It utilizes the Mask R-CNN algorithm for instance segmentation of scaffolding components to identify rusted areas, cracks, and surface deformations indicative of component quality. However, this method incurs high computational costs and is currently unable to assess the overall quality level of the entire scaffolding structure. Dzeng et al. [19] utilized the YOLOv5 algorithm to automatically detect missing scaffolding components and developed an AR-enabled system to assist on-site inspectors in identifying these defects. However, this method cannot measure the geometric dimensions necessary to evaluate the compliance of scaffolding structural parameters.

2.2. Quality Inspection Based on Point Cloud Data

Three-dimensional laser scanning has attracted considerable attention in the architecture, engineering, and construction (AEC) field as an emerging measurement technology, due to its notable advantages, such as high automation, non-contact operation, and efficient large-scale sampling capabilities [20]. Three-dimensional laser scanners use laser rangefinders to measure the distance between the scanner and the target object. They capture the geometric position, color, and intensity of the object’s surface and save these data as point clouds [21,22]. Currently, point clouds are widely used in 3D vision research and are extensively applied in 3D model reconstruction, deformation measurement, quality assessment, and progress control.

In quality assessments within the AEC industry, integrating point clouds with building information modeling (BIM) has become a common practice, known as the scan-vs.-BIM approach [23,24,25]. This involves loading both the as-built point cloud and the as-designed BIM model into the same digital environment and aligning them using software or algorithms to match common features. This alignment is carried out in a shared coordinate system, enabling the evaluation of the constructed project’s quality by comparing the differences between the models [26]. Currently, this method has been applied to varying extents in assessing quality and progress across diverse engineering objects, including piping systems [25,27], prefabricated concrete elements [22,28,29], steel structures [11,30,31], and bridges [32].

Despite the promising applications of laser scanning in the AEC sector, its utilization on scaffolding construction sites is challenged by complex environments and significant occlusions. Existing research primarily focuses on achieving high-quality 3D reconstruction of scaffolding structures using point clouds, by integrating unmanned aerial vehicles (UAVs) [14], robotic dogs [33], and visual sensors [34] with deep learning algorithms. Additionally, some studies focus on the semantic segmentation of scaffolding point clouds. Lin et al. [35] used the Hough transform and k-means clustering algorithms to generate masks for scaffolding components, and produced semantically segmented scaffolding component point clouds to detect missing components. Kim et al. [36] proposed an automated method to verify the proper installation of safety nets around scaffolding. The method employs the YOLOv5 algorithm to identify safety net boundaries and matches them with 3D points in the point cloud, thereby detecting scaffolding areas lacking safety nets. However, there is limited research on the application of point cloud data for assembly quality inspection in scaffolding structures. Wang [13] proposed an automated method using 3D point cloud data to inspect the compliance of scaffolding work platforms, but this method is limited to relatively simple configurations and cannot be extended to overall scaffolding structure compliance checks. Luo et al. [2] proposed a method utilizing multi-thread LiDAR point clouds to monitor temporary structural deformations, which aligns point data from multiple stations, identifies scaffolding rods using the RANSAC algorithm, and measures 3D deformations by comparing scaffolding axes over time. However, this method is limited to low-density point clouds, requires substantial planar features in the scene, and cannot assess the initial compliance of scaffolding. Kim et al. [37] employed the RandLA-Net model to segment various components in scaffolding point clouds and identify regulation violations by evaluating the point cloud density of different component types. However, this method can only evaluate the attachment conditions of working platforms, stairs, and guardrails, and cannot ascertain whether the geometric dimensions of the scaffolding comply with standards. Overall, there is currently no established theoretical framework for using laser scanning to evaluate the overall structural compliance of scaffolding.

2.3. Data Dimensionality Reduction

Scaffolding structures typically consist of numerous rods, couplers, and other components. Ensuring that each component meets stringent quality standards is crucial for assembly compliance. However, the multitude of structural components inevitably results in high-dimensional quality features, which poses significant challenges for data analysis. High-dimensional data occur when the number of features exceeds the sample size [38]. This not only significantly increases the computational and storage complexities of classification models, but also tends to lead to overfitting, ultimately resulting in degraded model performance [39]. This phenomenon is commonly known as the curse of dimensionality [40]. Dimensionality reduction is a common approach to address this issue, which includes feature extraction and feature selection. Feature extraction involves projecting original high-dimensional features into a new low-dimensional feature space through linear or nonlinear combinations, whereas feature selection directly chooses specific subsets from the high-dimensional feature set [40].

Currently, feature extraction methods, notably principal component analysis (PCA), are widely used in structural health monitoring (SHM) [38,41]. However, in the context of scaffolding structures, composite features derived from feature extraction lack practical relevance and do not streamline the collection of scaffolding assembly quality data. This study aims to identify key quality features from the numerous existing ones that significantly influence overall scaffolding assembly quality. Prioritizing these features during quality inspection facilitates a quick assessment of assembly quality, thereby simplifying the inspection process and improving efficiency. Therefore, feature selection methods are deemed more appropriate for this study compared to feature extraction.

Feature selection methods are widely employed preprocessing techniques in machine learning, aimed at improving model accuracy by removing irrelevant and redundant data [42]. Feature selection methods are categorized into three types: filter methods, wrapper methods, and embedded methods [43]. Filter methods use statistical techniques to assign weights to features and remove those deemed irrelevant based on a specified threshold. Wrapper methods utilize sequential or heuristic search algorithms to identify variable subsets that optimize the objective function, such as classification performance. Embedded methods incorporate feature selection directly into the training process to enhance computational efficiency. Comparatively, filter methods generally provide faster computational efficiency, whereas wrapper methods often yield superior performance [42]. Currently, feature selection methods have not been effectively applied to structural quality inspection, particularly for scaffolding structures.

3. Methodology

Traditional manual visual inspection methods for scaffolding heavily rely on the subjective judgment of designated on-site personnel, which is labor-intensive, time-consuming, and often unreliable. Although some studies have proposed sensor-based contact automated inspection methods, they may incur additional maintenance and management costs. There is a lack of effective non-contact methods for verifying the compliance of assembled scaffolding. This paper presents an automated inspection method for scaffolding assembly quality based on point cloud data, utilizing an SVM algorithm as the classification model. An enhanced GA is employed to identify the most relevant features impacting overall assembly quality. Training the SVM model with these features enhances its detection capabilities, resulting in higher accuracy in structural compliance inspections. The specific workflow of this study is illustrated in Figure 1. Initially, geometric parameters that may influence scaffolding assembly quality are identified from current standards and regulations to establish quality assessment criteria. Subsequently, high-precision point cloud data of the scaffolding are collected using a terrestrial 3D laser scanner and registered with the forward design model. In the registered model, different sections of the scaffolding are extracted as 2D slices to directly measure the deviation between selected structural parameters and the design values. After processing, these deviation data serve as the raw sample dataset for training the SVM model. To improve the classification accuracy of the SVM model, the GA is utilized for dimensionality reduction in the raw sample data, and a PSO algorithm is employed for SVM parameter optimization. Given the poor convergence of the GA, the Relief algorithm is employed to optimize population initialization and mutation iterations, enabling the rapid elimination of less important features and an efficient search for the optimal feature subset.

3.1. Data Preparation

3.1.1. Quality Features Determination

Scaffolding can be classified into several types based on their connection forms, including tube and coupler scaffolding, frame scaffolding, and disk lock scaffolding. Among these, disk lock scaffolding has gained widespread engineering applications due to its high load-bearing capacity, strong stability, low labor intensity, and lower overall cost. It has now become the mainstream form of scaffolding. A disk lock scaffolding system is assembled from various components, including standards, ledgers, base jacks, etc., as shown in Figure 2.

In China, a series of standards and regulations have been issued to specify the quality requirements for scaffolding construction. Relevant information on these standards and regulations is provided in

Table 1. Current standards and regulations for scaffolding in China.

Name	Level	Publication Year	Source
Unified standard for the safety of scaffolding in construction	National	2016	[44]
General code for scaffolding in construction	National	2022	[45]
Technical code for the safety of steel tubular scaffolding with couplers in construction	Industry	2011	[46]
Technical standard for the safety of disk lock steel tubular scaffolding in construction	Industry	2021	[47]

. This paper aims to study the geometric compliance of the assembled scaffolding overall. Therefore, the structural parameters of the scaffolding are selected as quality features, and thresholds for these parameters are set according to existing standards and regulations, as shown in

Table 2. Summary of scaffolding structural parameters and their thresholds.

Parameter	Description	Geometric Requirements	Source
$R$	The height-to-length ratio of independent scaffolding structures.	<3.0	[45]
$H_{scaff}$	The maximum erection height of independent scaffolding structures.	<30 m	[46]
$h$	The lift height.	<2 m	[44]
$H_{br}$	The height of base ledgers/transoms above ground level.	<550 mm	[47]
$L_{bl}$	The bay length.	<1.5 m	[44]
$L_{bw}$	The bay width.	<1.5 m	[44]
$L_{se}$	The height of standards extending above top ledgers/transoms.	<650 mm	[47]
$L_{bj}$	The exposed thread length of base jacks.	<300 mm	[44]
$L_{tj}$	The exposed thread length of top jacks.	<300 mm	[47]
$V$	Vertical alignment of standards.	<90 mm at 30 m height, proportional to height.	[46]

.

3.1.2. Dataset Establishment Based on Scan-vs.-BIM

To obtain high-precision geometric information on assembled scaffolding, this study utilizes a Trimble X7 3D laser scanner to capture point cloud data. During the scanning process, factors, such as scanner accuracy, surface reflectivity, and environmental conditions, inevitably introduced noise into the collected point cloud data, resulting in data redundancy and affecting point cloud accuracy. Therefore, CloudCompare 2.12.4 software is employed to trim the raw point cloud data in this study, extracting the target scaffolding point cloud. A bilateral filtering algorithm is applied to reduce noise and enhance data accuracy. The processed point cloud data are shown in Figure 3a.

The point cloud data contain geometric information about various structural parameters of target scaffolding. To measure the deviation of these parameters from the design values, a scan-vs.-BIM method is employed to align the point cloud with the corresponding forward design model, and the deviation is visualized using RGB colors. Initially, the point cloud and the model are imported into the same digital environment, and alignment is achieved by matching corresponding feature points. This process is carried out using Geomagic Control X 2020.1 software. The alignment results, shown in Figure 3b, achieve an accuracy of 0.1 mm. Structural parameters that reflect lengths, heights, and other geometric information of scaffolding components cannot be directly obtained from the alignment results. Therefore, a 2D slicing comparison method is adopted, involving the creation of multiple 2D planes in the alignment environment to sequentially slice different sections of the model and calculate the geometric contour deviations of the scaffolding in each section, as shown in Figure 3c. To collect geometric data, such as lift heights and bay lengths, the slicing planes should be vertical or horizontal and pass through the central axes of scaffolding components. The actual values of structural parameters can be calculated using the following Formula (1):

$$P_{point}=P_{model}+\Delta l$$

(1)

where $P_{point}$ represents the actual value of structural parameters in the point cloud, $P_{model}$ is the design value of these parameters in the forward design model, and $\Delta l$ is the deviation obtained through a 2D slicing comparison. Various geometric structural parameters of different samples are calculated and summarized to obtain the raw data of scaffolding assembly quality features. To avoid the impact of magnitude differences between different parameters on the accuracy of the classification model, the raw data were normalized using the following Formula (2):

$$x\prime ={\displaystyle \frac{x-x_{min}}{x_{max}-x_{min}}}$$

(2)

where $x$ represents the original feature data, and $x_{max}$ and $x_{min}$ are the maximum and minimum values of the same type of data, respectively.

3.2. Improved Relief-GA-Wrapper Algorithm

Assembled scaffolding typically involves numerous components, resulting in high-dimensional structural parameter data that can lead to the curse of dimensionality. In previous practice, the GA-Wrapper method, which uses the GA to search for the optimal feature subset that enhances classifier performance, has been commonly used for dimensionality reduction. The GA is inspired by the concept of evolution through natural selection and random mutations in biology. It represents feature subsets as binary strings of length $N$ (where $N$ denotes the total number of features in the original data), referred to as chromosomes. The $i$-th bit (0 or 1) indicates whether feature $i$ is discarded or retained. A population of chromosomes undergoes selection, crossover, and mutation to iteratively evolve and find the feature subset with the best classification performance. The GA-Wrapper algorithm demonstrates good robustness and excellent global optimization capability but has a slow convergence rate and heavily depends on the quality of the initial population. To improve the initial population and optimize the population’s iterative search process, this study introduces the Relief algorithm to enhance the performance of the GA.

The workflow of the improved Relief-GA-Wrapper algorithm is illustrated in Figure 4. First, the Relief algorithm calculates the weight of each feature, which serves as prior information for initializing the GA population and thereby improves the quality of the initial population. Additionally, different mutation rates are assigned to various features based on these weights, enabling the population to retain features with higher weights during the iteration process. Consequently, the retained features will exhibit a stronger correlation with the scaffolding assembly quality. Then, the population undergoes selection, crossover, and mutation, with classifier accuracy and corresponding fitness values calculated in each generation to identify the optimal feature subset.

3.2.1. Initialization of the GA Population

In this study, the population is initialized based on feature weights calculated by the Relief algorithm. The Relief algorithm, a well-established filter method, is primarily used for feature selection in binary classification problems. It assigns weights by evaluating the relevance of each feature to the target concept. The p-dimensional Euclid distance between samples serves as an evaluation criterion, suggesting that important features should cluster similar samples and distinguish different ones [48]. If the distance between a selected sample $R$ and a near-hit $H$ on a feature $A$ is smaller than the distance between $R$ and a near-miss $M$ on the same feature, then feature $A$ is considered effective for classification. Consequently, its weight should be increased. The weights will be updated according to the following Formula (3):

$$W\left[A\right]=W\left[A\right]-diff{\left(A,R,H\right)}^2/i_R+diff{\left(A,R,M\right)}^2/i_R$$

(3)

where $i_R$ denotes the iteration count, and the function $diff\left(A,X_1,X_2\right)$ calculates the difference in feature values between samples $X_1$ and $X_2$ on feature $A$. When the feature values are numerical, Equation (4) is as follows:

$$diff\left(A,X_1,X_2\right)={\displaystyle \frac{\left|value\left(A,X_1\right)-value\left(A,X_2\right)\right|}{\max \left(A\right)-\min \left(A\right)}}$$

(4)

where $value\left(A,X\right)$ represents the feature value of sample $X$ on feature $A$.

The GA population will be initialized as follows:

• Set the selection probability of the feature with the maximum weight to $p_{max}$ and that of the feature with the minimum weight to $p_{min}$. In this study, $p_{max}$ and $p_{min}$ are set to 0.8 and 0.2, respectively.
• For all other features, calculate their selection probability using the following Formula (5):

  $$p_i={\displaystyle \frac{\left(w_i-\min \left(w\right)\right)\left(p_{max}-p_{min}\right)}{\left(\max \left(w\right)-\min \left(w\right)\right)+p_{min}}}$$

  (5)
• Initialize the population individuals based on the selection probabilities of each feature.

3.2.2. Genetic Operations

The genetic operations in the GA include selection, crossover, and mutation. Before selection, the fitness value of each individual in the population is calculated using the evaluation function. In the proportional selection scheme, the probability of an individual being selected is directly proportional to its fitness value. The selection process is then carried out using the roulette wheel method.

In this study, the crossover probability parameter $P_c$ determines whether crossover will occur for different selected parent individuals. A two-point crossover operation, as illustrated in Figure 5, is used. For the two parent chromosomes selected for crossover, two crossover points $t_1$ and $t_2$ are randomly chosen, satisfying $1\le t_1\le t_2\le n$. The gene segments between these points are exchanged to generate two offspring chromosomes.

To achieve better dimensionality reduction, the number of features should be minimized while maintaining or improving classification accuracy. During population initialization, features included in each individual are selected based on their selection probabilities, effectively excluding most irrelevant features. As the population evolves, features with lower relevance should continue to be removed. This involves two steps: first, reducing the less relevant features included in individuals, generation by generation according to their weights, and second, preventing these features from reappearing in new individuals. Therefore, in this study, mutation operations are restricted to allow only changes from 1 to 0, ensuring that feature dimensions in the population only decrease. Additionally, the mutation rate for each feature is calculated using the following Formula (6):

$$P_{mi}=P_{m0}\times \left(1-p_i\right)$$

(6)

where $P_{m0}$ represents the base mutation rate, and $p_i$ denotes the selection probability of features as calculated by Formula (5). Under the constraint of Formula (6), features with lower weights have higher mutation rates and are more likely to be deselected. Conversely, features with higher weights are more likely to be retained in the offspring.

3.2.3. Evaluation Function

In the proposed Relief-GA-Wrapper algorithm, both classifier accuracy and the size of feature subsets jointly determine the effectiveness of dimensionality reduction. Therefore, the evaluation function should consider both factors, ensuring that individuals with higher classifier accuracy and smaller feature subsets are assigned higher fitness values. Inspired by the literature [49], this study proposes the modified evaluation function shown in Formula (7), as follows:

$$f\left(X\right)=\alpha \exp \left(1-{\displaystyle \frac{\left|X\right|}{n}}\right)+\left(1-\alpha \right)\exp \left({\displaystyle \frac{F_1\left(X\right)-\left(1-\beta \right)F_0}{\beta F_0}}\right)$$

(7)

where $\left|X\right|$ denotes the number of features in individual $X$, $n$ is the total number of original features, $F_1\left(X\right)$ represents the mean F1 score obtained by the classifier using k-fold cross-validation on the sample data with the selected features of individual $X$, and $F_0$ represents the F1 score of the classifier trained on the entire feature set. The first term on the right side of Formula (7) represents the contribution of feature subset size to the fitness value, which decreases as the number of features increases. The second term reflects the impact of classifier accuracy, which increases as the classifier accuracy improves. The constant $\alpha$ balances the evaluation function’s preference for feature count versus classifier accuracy; a larger $\alpha$ increases the influence of feature count while decreasing the influence of classifier accuracy. The constant $\beta$ is used for controlling the allowed rate of decrease in classifier accuracy, constraining $F_1\left(X\right)\ge \beta F_0$. When $F_1\left(X\right)<\beta F_0$, the second term significantly decreases, thereby reducing the fitness function value and eliminating the individual.

3.3. PSO-SVM Algorithm

3.3.1. SVM Algorithm for Non-Linearly Separable Problems

As previously mentioned, an individual’s fitness value is determined by both classifier accuracy and the size of the feature subset. Given the high-dimensional, small-sample, and nonlinear characteristics of the scaffolding quality feature data, this study employs the SVM algorithm as the classifier within the GA. The SVM algorithm is a binary classification model known for its excellent generalization performance and resistance to overfitting. For linearly separable problems, it searches for a classification hyperplane in the sample space, which serves as the decision boundary, and maximizes the minimum distance from this boundary to the edge points of the two sets, ensuring correct classification. For non-linearly separable problems, a kernel function maps the samples from the original feature space to a higher-dimensional space, transforming the problem into a linearly separable one in the high-dimensional space. The decision function of a nonlinear SVM, derived using the Lagrangian formula, is expressed as shown in the following Formula (8):

$$f\left(X\right)=sgin\left({\displaystyle \sum _{i=1}^m{\alpha }_i{Q}_i\kappa \left(X_i,X\right)+b}\right)$$

(8)

where ${\alpha }_i$ is the Lagrangian multiplier, $m$ is the number of samples; $Q_i$ is the label of sample $X_i$, indicating whether the scaffolding quality is acceptable; $b$ is a numerical parameter calculated from all samples satisfying $0<{\alpha }_i<C$; $\kappa \left(X_i,X\right)$ is the kernel function used, and this study employs the radial basis function (RBF) kernel, as in Equation (9):

$$\kappa \left(X_i,X\right)=\exp \left(-{\displaystyle \frac{{‖X_i-X‖}^2}{2{\sigma }^2}}\right)$$

(9)

where $‖X_i-X‖$ denotes the Euclidean distance, and $\sigma$ is the bandwidth parameter controlling the width of the kernel.

3.3.2. Parameter Optimization Using PSO Algorithm

The penalty factor $C$ and the bandwidth parameter $\sigma$ in the Gaussian kernel jointly determine the classification accuracy and the generalization capability of SVM. To enhance the performance of the classification model, an optimization framework based on the PSO algorithm is employed, as illustrated in Figure 6.

First, initialize the swarm size $N_{PSO}$ and randomly set the position and velocity of each particle within the given constraints. Since there are two parameters to optimize, the dimensionality of the particles is set to 2. As a result, each particle’s position vector is denoted as $x_i=\left(x_{i1},x_{i2}\right)$ and its velocity vector is set as $v_i=\left(v_{i1},v_{i2}\right)$. The variables $x_{i1}$ and $x_{i2}$ of the position vector correspond to the values of $C$ and $\sigma$, respectively, while the variables $v_{i1}$ and $v_{i2}$ of the velocity vector represent changes in $C$ and $\sigma$, respectively. For each particle, an SVM model is constructed based on its position vector, and the particle’s fitness is quantified by the classification performance of the SVM model. The F1 score is used to evaluate this performance, with the mean of the k-fold cross-validation F1 scores serving as the fitness value for each particle.

During the iterations, update the positions and velocities of each particle, compare the fitness values across different iterations, and determine the personal best value ($pbest$). Once the personal best values converge, compare them across particles to identify the global best value ($gbest$). The SVM classification performance at the global best position represents the optimal performance within the swarm. Consequently, the optimal values for the penalty factor $C$ and the bandwidth parameter $\sigma$ are found at the position vector parameters $x_{gbest1}$ and $x_{gbest2}$ of the global best particle, respectively.

4. Experimental Results and Analysis

4.1. Data Collection and Preprocessing

To validate the feasibility of the proposed method, an experiment and analysis were conducted through a case study of a support scaffolding project in an underground space in Nanjing, Jiangsu Province, China. A Trimble X7 3D laser scanner was used for on-site point cloud data collection. The average error of the point cloud data obtained from on-site 3D laser scanning after registration is 1.4 mm, with an average confidence level of 99.11%. Following preprocessing, the overall scaffolding point cloud was segmented into groups, each consisting of 15 units of bay arranged in 3 rows and 5 columns. The components were then numbered, as illustrated in Figure 7. For each group of point cloud data, registration and comparison were conducted against the forward design model. The geometric structural parameters of each scaffolding group were measured using 2D comparison, and their compliance was evaluated based on standards, regulations, and practitioner opinions, ultimately forming a dataset for scaffolding assembly quality.

A total of 200 sample datasets were collected for this experiment, with a 1:1 ratio of qualified to unqualified samples. Each sample consists of 84-dimensional feature data, including 1 item of $R$, 2 items of $h$, 1 item of $H_{br}$, 5 items of $L_{bl}$, 3 items of $L_{bw}$, 24 items of $L_{se}$, 24 items of $L_{tj}$, and 24 items of $V$. The definitions of the different parameters are provided in Table 2. The numbering of $h$, $L_{bl}$, and $L_{bw}$ is given in Figure 7, while the numbering of $L_{se}$, $L_{tj}$, and $V$ corresponds to the numbering of standards ($S$). After normalization, the statistical information for experimental data under different features is shown in

Table 3. Normalized sample feature data statistics.

	${\mathit{h}}_1$	${\mathit{h}}_2$	${\mathit{L}}_{\mathit{t}\mathit{j}-1}$	${\mathit{L}}_{\mathit{s}\mathit{e}-1}$	${\mathit{V}}_1$	…	${\mathit{H}}_{\mathit{b}\mathit{r}}$	$\mathit{R}$
Sample Size	200	200	200	200	200	…	200	200
Mean	0.558	0.303	0.450	0.518	0.365	…	0.498	0.313
Variance	0.218	0.175	0.228	0.214	0.260	…	0.287	0.161
Minimum	0.000	0.000	0.000	0.000	0.000	…	0.000	0.000
Lower Quartile	0.500	0.167	0.282	0.354	0.182	…	0.261	0.190
Median	0.583	0.333	0.459	0.530	0.319	…	0.556	0.313
Upper Quartile	0.667	0.417	0.616	0.683	0.545	…	0.778	0.406
Maximum	1.000	1.000	1.000	1.000	1.000	…	1.000	1.000

.

4.2. Feature Weight Calculation

The Relief algorithm was applied to calculate the weights of the 84-dimensional features based on normalized sample feature data. The results are presented in Figure 8a. A positive feature weight indicates that the feature positively contributes to distinguishing between qualified and unqualified scaffolding, whereas a negative weight indicates a weak correlation with assembly quality. The results reveal that only 10 out of the 84 features possess negative weights. This implies that the majority of features are strongly correlated with scaffolding assembly quality. Among the eight feature categories, the weight differences in the vertical alignment of standards ($V$) are particularly pronounced, displaying a polarized distribution. From a feature definition perspective, this is because current standards and regulations impose the strictest quality requirements on this feature, with the smallest allowable deviation range, as shown in Table 2. Even minor angular deviations at the lower end can lead to unacceptable displacements at the top, which, if exceeding safety thresholds, would render the scaffolding non-compliant. This explains the high sensitivity of overall assembly quality to variations in this feature. In contrast, other feature categories measure specific dimensions of the scaffolding, such as length, height, and distances between components. Within the same scaffolding project, these parameters typically show minimal variation, and the quality standards set by regulations are more lenient, allowing for greater flexibility. Consequently, these features exhibit similar weight values and a relatively smooth decline within their categories. Overall, considering both feature weights and the sensitivity of assembly quality to changes in feature values, the importance of the eight feature categories can be ranked as follows: (1) vertical alignment of standards ($V$); (2) lift height ($h$); (3) bay length ($L_{bl}$); (4) height of standards extending above top ledgers/transoms ($L_{se}$); (5) exposed thread length of base jacks ($L_{tj}$); (6) bay width ($L_{bw}$); (7) height of base ledgers/transoms above ground level ($H_{br}$); (8) height-to-length ratio of independent scaffolding structures ($R$).

The weight values of all features are sorted from highest to lowest, and the trend is illustrated in Figure 8b. It is apparent that the overall trend initially drops sharply, then gradually, and finally sharply again, with distinct inflection points around the 20th and 70th positions. The weights of vertical alignment of standards ($V$) and lift height ($h$) are significantly higher than those of other quality characteristics, resulting in a steep initial decline in the weight curve. The bay length ($L_{bl}$), the height of standards extending above top ledgers/transoms ($L_{se}$), and the exposed thread length of base jacks ($L_{tj}$) exhibit very similar weights, and their combined proportion exceeds half of the total. This results in a long, gentle decline in the middle section of the curve. As the rank approaches the lower end, irrelevant features with negative weights become predominant, causing the curve to drop sharply once more. Overall, it can be preliminarily concluded that the top 20 features by weight have the strongest impact on assembly quality.

4.3. Feature Selection Results

Once the weight information is obtained, the selection probability of different features during the genetic operation can be calculated, which will also serve as prior information for initializing the genetic algorithm’s population. The parameters of the GA were set as follows: population size $N_{GA}$ of 50, maximum iterations $i_{GA}$ of 200, crossover probability $P_c$ of 0.85, base mutation probability $P_{m0}$ of 0.05, preference constant $\alpha$ of 0.50, and classifier accuracy allowance reduction ratio $\beta$ of 0.05. The fitness value in the genetic algorithm was calculated using 5-fold cross-validation, with classifier accuracy evaluated using the F1 score, and the classification model employed was PSO-SVM. The parameters for the PSO algorithm were set as follows: a particle swarm size $N_{PSO}$ of 20, maximum iterations $i_{PSO}$ of 100, an inertia factor $\omega$ of 0.90, and learning factors $c_1$ and $c_2$ both set to 2.00.

The fitness value variation curve during the genetic operation is shown in Figure 9a. The results show that the proposed Relief-GA-Wrapper algorithm outperforms the traditional GA-Wrapper algorithm in feature selection for scaffolding assembly quality. The traditional GA-Wrapper algorithm converges slowly, reaching a local optimum at the 42nd generation with a fitness value of 6.569 and a corresponding F1 score of 0.849, selecting only nine features. In contrast, the proposed Relief-GA-Wrapper algorithm rapidly improves the fitness value early on in the training, achieving an optimal value by the 22nd generation, with a fitness value of 15.354, a corresponding F1 score of 0.889, and selecting 25 features. This indicates that the proposed algorithm not only doubles the optimization speed of the traditional GA-Wrapper algorithm but also significantly improves classifier accuracy.

Considering both the F1 score and feature subset size variation curves, the traditional GA-Wrapper algorithm quickly reduces the feature subset size but consistently achieves lower classification accuracy compared to the proposed algorithm. This discrepancy arises because the traditional GA-Wrapper algorithm randomly selects feature subsets when generating the initial population, leading to the omission of some high-weight features and the inclusion of less relevant ones. As iterations progress, these irrelevant features are quickly discarded, resulting in a rapid reduction in the feature subset size. In this study, mutation operations are restricted to changing features from 1 to 0, making it unlikely for initially discarded important features to reappear in the offspring population, which results in decreased classification accuracy. In contrast, the proposed method prioritizes highly relevant features based on their weights from the start and uses mutation operations to eliminate relatively less important features during iterations. This approach allows the feature subset to gradually converge to the Relief algorithm’s selection results, thus improving search efficiency. In conclusion, the proposed algorithm outperforms the traditional algorithm in searching for the optimal feature subset.

The optimal feature subset obtained by the proposed algorithm is shown in

Table 4. Composition of the optimal feature subset and its feature weights.

Parameter	Number	Weight	Number	Weight	Average Weight
$V$	1	0.038540	14	−0.007118	0.024585
	3	0.027063	15	0.084471
	7	−0.013026	16	0.010180
	9	0.033467	17	0.028631
	11	0.030232	18	−0.009158
	12	0.022318	19	0.027120
	13	0.011102	20	0.060374
$L_{tj}$	4	0.008545	18	0.011076	0.010464
	9	0.008117	21	0.012867
	12	0.010207	23	0.011974
$L_{se}$	9	0.011309	13	0.010802	0.011056
$h$	1	0.036448	/	/	0.036448
$L_{bl}$	1	0.018430	/	/	0.018430
$H_{br}$	/	0.007077	/	/	0.007077

. This feature subset consists of 25 features across 6 categories, including 14 of $V$, 6 of $L_{tj}$, 2 of $L_{se}$, and 1 each of $h$, $L_{bl}$, $H_{br}$. The appearance of negative weight features (including $V_7$, $V_{14}$, and $V_{18}$) in the results is due to the proposed algorithm reaching its maximum fitness value as early as the 22nd generation, at which point these negative weight features had not yet been screened out. The impact of these negative weight features on classification accuracy will be discussed in Section 4.4. Considering both the number of features and their weights, it can be determined that $V$ has the strongest influence on scaffolding assembly quality, followed by $h$, and subsequently $L_{tj}$, $L_{se}$, $L_{bl}$, and $H_{br}$. This is consistent with the results calculated by the Relief algorithm.

The results demonstrate that the proposed method reduced the dimensionality of the scaffolding assembly quality data from 84 to 25 dimensions, achieving a 70% reduction. Furthermore, the algorithm’s efficiency and accuracy were significantly improved.

4.4. Ablation Study

To evaluate the contributions of different modules in the proposed algorithm, an ablation study was conducted on the established scaffolding assembly quality dataset. The SVM algorithm, with hyperparameters optimized via GridSearchCV, served as the baseline method. It achieved a classification accuracy of 0.750 when trained on the full feature set. Various algorithm modules were then incrementally added to the baseline model to create different classification models, and their performances were assessed. As the Relief algorithm only calculates feature weights without extracting subsets, the top 25 features were manually selected for Models 4 and 6 to match the feature count of Model 8. Additionally, Models 3 and 7, respectively, used the same initial population as Models 5 and 8 to control for experimental variables. The performance results of the different models are listed in

Table 5. Test results of the classification model with different algorithm modules on the established dataset.

Test Model	Settings				Fitness Value	Classifier Accuracy	Number of Feature
	Relief	GA	PSO	SVM
1				√	0.940	0.750	84
2			√	√	1.209	0.769	84
3		√		√	5.961	0.845	11
4	√			√	1.794	0.773	25
5	√	√		√	12.475	0.879	16
6	√		√	√	2.340	0.794	25
7		√	√	√	6.569	0.849	9
8	√	√	√	√	15.354	0.889	25

.

The experimental results clearly demonstrate the effectiveness of the optimized methods proposed in this study. The SVM algorithm showed significant improvement in classification performance for scaffolding assembly quality when applied to the reduced-dimension dataset. The Relief algorithm enhanced the correlation between feature subsets selected by the GA and assembly quality, resulting in more reasonable feature subset sizes and a 2.3% to 4.0% increase in SVM classification accuracy. The PSO algorithm further improved SVM classification accuracy by 0.4% to 1.9% through parameter optimization under the same conditions. Furthermore, the comparison between Model 6 and Model 8 indicates that, although the feature subset selected by the proposed method includes features with negative weights (as shown in Table 4), the classification performance of the algorithm is significantly improved compared to Model 6, which selected only the 25 features with the highest weights. Overall, the proposed method is highly suitable for the problem of scaffolding assembly quality inspection. It reduced feature dimensionality from 84 to 25 while significantly enhancing classifier accuracy by 13.9% compared to the SVM algorithm without optimization.

5. Conclusions

Ensuring the structural stability of scaffolding is crucial for construction safety. This study addresses the traditional scaffolding inspection challenges of high manual dependency, low accuracy, and inefficiency by proposing an automated inspection method based on point cloud data and feature selection algorithms. High-precision scaffolding point clouds are captured by 3D laser scanning technology, and a 2D slicing comparison method is developed to measure geometric parameters, achieving an accuracy of up to 0.1 mm. An optimized SVM algorithm is employed as the classification model to inspect scaffolding assembly quality. This model enables automated classification and inspection of assembly quality by using preprocessed geometric feature data. To circumvent the curse of dimensionality associated with high-dimensional data, an improved GA is used to reduce the dimensionality of raw sample data. It effectively eliminates data redundancy and significantly improves the convergence speed and classification accuracy of the classification model. In case studies, the proposed method demonstrates a strong capability to identify the optimal feature subset, reducing the dimensionality of the scaffolding’s structural parameter data by 70% and increasing the classification model’s accuracy by 13.9%. This demonstrates that prioritizing key structural parameters and eliminating less relevant features enhances the accuracy of scaffolding assembly quality inspections. The reduction in feature numbers also simplifies the inspection process and improves efficiency. This approach provides construction companies with a new method for improving scaffolding assembly quality inspections. On-site inspections should prioritize features closely related to assembly quality, such as vertical alignments of standards and lift heights. Overall, the proposed method significantly improves the accuracy of measuring scaffolding’s structural parameters, reduces data redundancy, and enables automated detection of assembly quality using the SVM model, thus substantially enhancing inspection accuracy and efficiency. This represents a significant advance in intelligent construction and has substantial innovative value.

Despite its contributions, this study has certain limitations. The scaffolding structure used in the experiment excludes components, such as toeboards and diagonal braces, which are also subject to quality standards and regulations, resulting in an incomplete dataset. Additionally, the automatic and rapid segmentation of the target scaffolding from the scanned construction scene point cloud remains a significant challenge. From an algorithmic perspective, although the PSO algorithm effectively optimizes the SVM classifier’s hyperparameters, it suffers from long iteration times and high training costs. Future research should aim to develop automatic segmentation methods for scaffolding point clouds, apply these methods across various construction scenarios, and establish a more comprehensive dataset for scaffolding assembly quality. Furthermore, further optimization of algorithms is necessary to reduce training costs and facilitate real-time monitoring.