Influencing Factors and Prewarning of Unsafe Status of Construction Workers Based on BP Neural Network

Abstract

Scholars have paid considerable attention to the factors that affect the safety states of construction workers. However, only a few studies have focused on the safety assessment and security alerts of individual workers. In this study, the term ‘frequency statistics’ refers to the factors considered by domestic and foreign experts and scholars. The statistical results were combined with the interpretation of these factors to determine 22 factors that negatively influence the safety status of construction workers, which were used as the research object. The initial weight of the research results was integrated into the BackPropagation neural network, using the improved analytic hierarchy process to establish an early warning model for the unsafe status of construction workers. The mean squared error meets the requirements of the model and the prediction accuracy meets the requirements of the sample. The model can effectively provide an early warning and correct the initial weighting of the results. The early warning model was then applied to a project that involved the construction of a primary school in Suzhou. The follow-up results show that the safety status of the workers significantly improved. These results show that the early warning model was successfully used in the safety assessment to provide security alerts to individual workers. The data obtained by comprehensively considering both workers and experts are universal, unlike those obtained by considering only one of these two groups. Among the indicators, safety awareness, protection measures, and team cohesion most strongly negatively affected the safety statuses of the construction workers. The results of the early warning model combined with the sensitivity analysis are targeted and applicable in the practice of safety monitoring.

1. Introduction

In recent years, the safety management level in the field of engineering construction in China has been greatly improved by improving the relevant laws and regulations, implementing the measures of strengthening the management and control of partial projects with high risk, strengthening the accountability mechanism regarding safety accidents, and promoting the standardization of construction safety production [1]. However, casualties still occur in China’s engineering construction field; the number of these casualties remains high, and the worker security situation is still poor. According to the data released by the Ministry of Housing and Urban-Rural Development of the People’s Republic of China, the numbers of deaths that occurred as a result of safety accidents in housing and municipal engineering in China were 807, 840, 904, and 794 in 2017, 2018, 2019, and 2020, respectively [2,3,4]. It can be concluded that, from 2017 to 2019, there was no effective solution to avoid worker deaths at this time. In 2020, there was widespread construction stoppage, attributed to the COVID-19 pandemic, resulting in the decrease in deaths, instead of being due to the adoption of effective measures. According to the summary and analysis of the causes of construction accidents in China, in addition to the unsafe behavior of construction workers in China and abroad, the unsafe behavior of construction workers has a considerable impact on construction accidents. Therefore, an increasing number of scholars worldwide have begun to invest in research on the factors that negatively influence the safety status of construction workers.

In view of the safety accidents caused by the construction industry, management behavior, work experience, safety atmosphere, leadership attention, safety norms, safety attitude, safety ability, safety awareness, and other factors have been studied by scholars in China and abroad. Shi Yufang and Lu Jifa [5] took construction workers as the investigation objects from the perspective of model construction. They obtained data through a questionnaire survey and built a structural equation model to study the influence path and mechanism of construction workers’ fatigue on unsafe behavior. Wang Cuiying and Wang Qiang [6] used a new method based on an explanatory structural model to study the factors that negatively affect the safety behavior of construction workers. Their results show the hierarchical relationship of the influencing factors behind the unsafe behaviors of construction workers. Moreover, the basis for effectively reducing safety accidents was provided by analyzing each level in detail. On the basis of analyzing the characteristics of unsafe behaviors, Meng et al. [7] proposed an assessment method for the risk value of unsafe behaviors, built a risk assessment model for unsafe behaviors, and proposed different intervention strategies according to the risk values of unsafe behaviors. Shin et al. [8] established a psychological process model for construction workers based on system dynamics. They also analyzed the relationship feedback mechanism between workers’ safety attitudes and their safety behaviors. They pointed out that workers sharing accident information with one another or experiencing accidents indirectly through audiovisual materials can reduce the incidence of unsafe behaviors and avoid the formation of unsafe habits. Wu Fan et al. [9] established an evolutionary game model to explore the safety behavior strategies of the main contractors and construction personnel. Under this framework, the influence of the model parameter changes on the evolution of security behavior strategies was explored, and their evolution paths under different initial values were verified by numerical simulation. Byungjoo et al. [10] developed an empirical agent model that combines the theory of the social cognitive process of worker safety behavior with empirical research results. They conducted experiments to examine how social cognitive processes interact with safety management interventions and influence worker safety behavior under different site risk conditions.

Jiang Yueqi et al. [11] adopted a fuzzy comprehensive evaluation method (analytic hierarchy process (AHP)) to quantitatively analyze the propagation strength of some unsafe behaviors of construction workers and provide a theoretical basis and control strategy for the propagation of unsafe behaviors of construction workers from the perspective of research methods. Xu Rui et al. [12] used the AHP to establish a hierarchical structure model and calculate the weight and classification of the factors that affect the unsafe behaviors of coal mine employees. The results show that the most important factors that affect the unsafe behavior of miners are their own factors, organizational management, and the working environment. In comparison, the equipment factors and social factors have a minimal impact, thereby providing a theoretical basis for enterprises to prevent accidents. Zhao Hongchao et al. [13] constructed the AHP model for the severity of consequences of the unsafe behaviors of miners, calculated the weight ranking of various factors using MATLAB software, and determined the severity of the consequences of various unsafe behaviors. Zhao Jinna et al. [14] took the height falling accident as the research object according to the definition of the brittleness excitation degree. They also conducted the brittleness analysis of height falling accidents. Referring to the fuzzy evaluation model, Pei Zhuoqin [15] adopted the AHP to solve the problem of index weight distribution. They built a fuzzy comprehensive evaluation model of coal mine safety culture based on the AHP. Shapira et al. [16] interviewed 19 safety experts. They concluded through the AHP that the construction site supervision and the driver’s behavior have the greatest impact on tower crane safety.

In terms of statistical analysis, Lu et al. [17] and Mearns et al. [18] conducted a large-scale questionnaire survey of workers. They concluded that management behaviors, such as organizational safety policy and the concern of organizations and managers for workers’ safety, can positively influence workers’ behavioral choices. Siu et al. [19] conducted a questionnaire survey with construction workers at 27 sites in Hong Kong and found that safety attitude can predict occupational injury, workers’ psychological burden affects the on-site accident rate, and psychological stress plays a role in the transmission between safety attitude and accident rate. Moreover, Seo [20] conducted a survey with 722 workers in 102 regions of a multi-national company and demonstrated that among the five factors affecting human safety behavior, including perceived safety atmosphere, perceived danger level, perceived work pressure, perceived risks, and perceived obstacles, the perceived safety atmosphere had the greatest impact on the workers’ safety behavior. Finally, Neal et al. [21] discussed the impact of organizational climate on safety climate and safety performance. A survey of 525 employees from 32 teams in a large Australian hospital showed that organizational climate has a significant impact on safety climate, which in turn influences the workers’ self-reported compliance with safety regulations and their participation in safety-related activities.

In terms of theoretical framework and case analysis, Ye Gui et al. [22] believed that the influencing factors of unsafe behavior can be summarized into the following four aspects according to the four-factor theory of accident systems: human cognition, emotion, physical quality, and consciousness factors; placement and damage of objects factors; management methods and level factors; environmental quality and standard factors. Shengyu Guo et al. [23] explored two aspects of the statistical rule of time of workers’ unsafe behaviors through the case study method. The results showed that the time distribution between events presents a thick-tailed distribution, indicating that workers’ unsafe behaviors are characterized by suddenness and memorization. In addition, the strict rule of “construction phase → unsafe behavior” exists for different types of workers.

The literature reviewed in this paper highlights that scholars have extensively analyzed the influence mechanism of unsafe states from various perspectives. However, despite the significant amount of research on the topic, there are still shortcomings. First, in the study of unsafe behavior factors, the data source only reflects the ideas of a single group, leading to a lack of overall consideration, analysis, and comparison of the opinions of experts and workers. Second, AHP is a subjective method, using index weights can to some extent reduce the subjectivity of AHP. Finally, current studies mainly focus on the model algorithm, statistical analysis, and case analysis. However, studies on early warnings from the perspective of influencing factors are limited. Given that neural network technology is applicable to large-scale data and can explore the deep rules between the data with strong objectivity and a good prediction effect, the neural network was chosen for studying the influencing factors of unsafe states among construction workers and establishing an early warning model to enhance construction worker safety.

2. Research Methods and Technical Routes

2.1. The Index System Was Determined by the Statistical Method of Word Frequency in the Literature

We conducted a comprehensive literature search to identify the factors that influence the unsafe behavior of construction workers. China national knowledge infrastructure and Web of Science databases were used as the main sources for the literature search. As shown in Table 1, 22 kinds of high-frequency influencing factors were extracted by searching the abstracts, keywords, and word frequency statistics of 480 works.

Table 1. High-frequency influencing factors.

Serial Number	High-Frequency Words	Number of Articles
1	Safety climate [24,25]	66
2	Security awareness [26,27]	46
3	Security attitude [6,28]	40
4	Safety knowledge [29,30]	39
5	Safety education [31,32]	34
6	Workers’ emotions [33,34]	31
7	Security perception [35,36]	26
8	Security culture [37,38]	26
9	Security leadership [39,40]	21
10	Work experience [26,29]	19
11	Psychological factors [28,36]	17
12	Safety compliance [41]	16
13	Workers’ fatigue [5,42]	15
14	Skills level [27]	14
15	Safety input [30,35]	13
16	Safety supervision [28,31]	12
17	safety regulation [43]	10
18	Workmate impact [44,45]	9
19	Safety communication [24,27]	8
20	Construction conditions [46]	8
21	Management level [40,47]	6
22	Workers’ character [48,49]	4

According to the interpretation, the above high-frequency factors were preliminarily screened, and 22 factors were determined as research factors. The interpretation is shown in Table 2.

Table 2. Interpretation of high-frequency factors.

Serial Number	Factor	Interpretation
1	Safety climate	Safety level of working environment
2	Security awareness	Summarization of construction workers’ psychological experience of safety
3	Safety attitude	Some opinions and intentions of construction workers after evaluating the significance or importance of construction to safety in production
4	Safety education	Necessary means of transmitting safety experience
5	Workers’ emotions	The tendency and dynamic response of workers to objects based on their needs
6	Safety culture	Safe production mode in construction industry
7	Safety leadership	Managers’ ability regarding safety guidance and command for construction workers
8	Work experience	Work background of workers in construction industry
9	Safety compliance	Workers’ compliance with production safety regulations
10	Workers’ fatigue	The state of long-term and high-load physical labor of construction workers
11	Level of skill	Construction workers’ ability to work
12	Safety input	Relevant incentive policies developed to ensure all human, material and financial resources are invested in security
13	Safety supervision	Supervision and management of workers’ insecurity by managers
14	Safety regulation	Enterprise formulates rules and regulations on safe production
15	Workmate impact	Communicate, monitor and avoid unsafe behavior among employees
16	Safety communication	Safety leadership of management personnel and cooperation with workers; group communication efficiency
17	Fabrication condition	The construction environment of construction enterprises and construction teams under reasonable organization construction conditions
18	Level of management	Managers’ ability regarding safety guidance and command for construction workers
19	Workers’ character	Attitude to reality and more stable psychological characteristics in habitual behavior
20	Conservatory measure	Workers protect against threats to certain parts of the body during work
21	Worker age	Number of years workers have lived
22	Group cohesiveness	Collective activities organized by managers emphasize cognitive factors of cohesion and group solidarity based on the process of collective activities

2.2. Research Process

To fully consider the opinions of both experts and workers, we first collected expert opinions through a questionnaire survey and utilized an improved AHP method to determine the weight of each factor. Subsequently, we collected the opinions of workers through questionnaires and accordingly assigned them weights. The BackPropagation neural network was then employed to establish the prewarning model. The weight correction value could reflect the common opinions of experts and workers at this time. Therefore, the factor weight could be analyzed, and the model could be used for the safety assessment and early warning of individual workers. The specific research route is shown in Figure 1.

Figure 1. The specific research process.

2.3. Improved AHP

The theoretical basis of improvement was proposed by American operations research scientist, Professor T.L. Suaty. The original scale of 1–9 was replaced with a scale based on the natural index to determine the weight of the judgment matrix. The specific steps for improving the algorithm are outlined below.

2.3.1. Establishment of Hierarchical Structure Model

When applying the AHP to analyze a decision problem, the problem should be adjusted and stratified to construct a hierarchical structure model. Under this model, complex problems are decomposed into components of elements. Then, these element components form several elements at the upper level as criteria according to their attributes and relationships and play a dominant role in the related elements at the next level.

2.3.2. Judgment Matrix Construction

A pairwise comparison of factors is performed to establish a pairwise comparison matrix. Two factors $X_i$ and $X_j$ are taken each time, and $a_{ij}$ represents the ratio of the influence of $X_i$ and $X_j$ to $Z$. All the comparison results are represented by matrix $A={(a_{ij})}_{n\times n}$. Matrix $A$ is called the pairwise comparison judgment matrix. If the ratio of the effects of $X_i$ and $X_j$ on $Z$ is $a_{ij}$, then the ratio of the effects of $X_j$ and $X_i$ on $Z$ should be $a_{ji}=\frac{1}{a_{ij}}$.

Luo Zhengqing [50] conducted a study on different scales used in the AHP method, including order preservation, consistency, uniformity, memory, and perception. After the research, they concluded that sorting under a single criterion should use the 1–9 scale to determine the weight. Moreover, the scale based on natural exponents should be used for the multicriteria sorting problem that requires high precision to determine the weights. In the present study, the scale based on the natural index is used to determine the subjective weight, as shown in Table 3.

Table 3. The meaning of different scales.

Scale	Implication
$e^{0/4}$	Factor i is as important as factor j
$e^{2/4}$	The i factor is slightly more important than the j factor
$e^{4/4}$	The i factor is more important than the j factor
$e^{6/4}$	The i factor is more important than the j factor
$e^{8/4}$	The i factor is more important than the j factor

Furthermore, $e^{1/4}$, $e^{3/4}$, $e^{5/4}$, and $e^{7/4}$ indicate that the importance of the i-th factor relative to the j-th factor is between two adjacent ranks. A judgment matrix $A={(a_{ij})}_{n\times n}$ is formed. $a_{ij}$ is used to express the importance of the i-th factor relative to the j-th factor. Then, $a_{ij}$ satisfies Formula (1).

$$\{\begin{array}{c}{a}_{ij}>0\\ a_{ji}=\frac{1}{a_{ij}}\\ a_{ii}=1\end{array}$$

(1)

2.3.3. Hierarchical Single Ordering and Consistency Check

(1)

Hierarchical single sorting

The judgment matrix A corresponds to the eigenvector W of the largest eigenvalue ${\lambda }_{\max }$. After normalization, it is the ranking weight of the relative importance of the corresponding factors at the same level as the factor at the previous level.

The above method of constructing a paired comparison judgment matrix can reduce the interference of other factors. However, it is inevitable that some degree of inconsistency may arise when all the results are compared. Therefore, a consistency test for the judgment matrix provided by decision makers must be conducted to determine whether it can be accepted.

(2)

Calculation of consistency index CI

$$CI=\frac{{\lambda }_{\max }-n}{n-1}$$

(2)

(3)

Calculation of the corresponding average random consistency index RI

The judgment index used is different from the 1 to 9 scale. Thus, the value of the RI must be solved. A total of 10,000 sample matrices are constructed with the random method. The numbers are randomly extracted from $e^{0/4},e^{2/4},e^{4/4},e^{6/4},e^{8/4},e^{1/4},e^{3/4},e^{5/4},e^{7/4}$ and their reciprocals to construct the reciprocal matrix. The mean value ${\lambda }_{\max }^{\prime }$ of the largest eigenroot is searched, and RI is defined as Formula (3).

The final RI value is shown in Table 4.

$$RI=\frac{{\lambda }_{\max }^{\prime }-n}{n-1}$$

(3)

Table 4. The value of RI.

n	3	4	5	6	7	8	9	10
RI	0.2843	0.4682	0.5850	0.6607	0.7010	0.7341	0.7618	0.7833

(4)

Calculation of consistency ratio CR

$$CR=\frac{CI}{RI}$$

(4)

2.3.4. Hierarchical Total Sorting and Consistency Test

(1)

Hierarchical total ranking

What we obtained above is a set of weight vectors of the elements with respect to a certain element in its upper layer. Total hierarchical sorting must be performed to obtain the sorting weight of each element, particularly the sorting weight of each scheme in the bottom layer. For the total sorting weight, the weights under the single criterion should be synthesized from top to bottom, as shown in Table 5.

Table 5. Hierarchical total sort.

	Layer A	$A_1$	$A_2$	…	$A_m$	Total ranking weight of layer B
Layer B		$a_1$	$a_2$	…	$a_m$
$\begin{array}{l}{B}_1\\ \\ B_2\\ \\ \dots \\ \\ B_n\end{array}$		$\begin{array}{l}{b}_{11}\\ \\ b_{21}\\ \\ \dots \\ \\ b_{n1}\end{array}$	$\begin{array}{l}{b}_{12}\\ \\ b_{22}\\ \\ \dots \\ \\ b_{n2}\end{array}$	$\begin{array}{l}\dots \\ \\ \dots \\ \\ \dots \\ \\ \dots \end{array}$	$\begin{array}{l}{b}_{1m}\\ \\ b_{2m}\\ \\ \dots \\ \\ b_{nm}\end{array}$	$\begin{array}{l}{\displaystyle \sum _{j=1}^m{b}_{1j}{a}_j}\\ {\displaystyle \sum _{j=1}^m{b}_{2j}{a}_j}\\ \dots \\ {\displaystyle \sum _{j=1}^m{b}_{nj}{a}_j}\end{array}$

One must assume that the upper layer (layer A) contains $A_1$, …, $A_m$ factors, and their total ranking weights are $a_1$, …, $a_m$. Then, one must also assume that the next layer (layer B) contains n factors $B_1$, …, $B_m$, and their weights for the hierarchical single ordering of $A_j$ are $b_{1j}$, …, $b_{nj}$ (when $B_i$ is not associated with $A_j$, $b_{ij}=0$). Now, one can calculate the weight of each factor in the B layer for the overall goal, that is, to find the total ranking weight $b_1$, …, $b_n$ of each factor in the B layer, and calculate it according to $b_i={\displaystyle \sum }_{i=1}^m{b}_{ij}{a}_j$, i = 1, …, n.

(2)

Consistency test

A consistency check is also required for the total ordering of the hierarchy. Similar to the total ordering of layers, the inspection is performed layer by layer, from a high layer to a low layer. The reason for this is that inconsistencies at various levels may accumulate when a comprehensive investigation is performed, leading to serious inconsistencies in the final analysis results.

One can suppose that the upper layer (layer A) contains m factors, A₁, …, A_m, and the pairwise comparison judgment matrix of the factors related to A_j in layer B is tested for consistency in a single ranking test. In this case, the single-ranking consistency index is obtained as $CI(j)$, $(j=1,\dots ,m)$. The corresponding average random consistency index is $RI(j)$. Then, the random consistency ratio of the total ranking of layer B can be obtained with Formula (5).

$$CR=\frac{{\displaystyle \sum _{j=1}^{m}CI(j)a_j}}{{\displaystyle \sum _{j=1}^{m}RI(j)a_j}}$$

(5)

When CR < 0.10 [51], the consistency of the judgment matrix is acceptable; otherwise, the judgment matrix should be appropriately modified.

2.4. BP Neural Network

A neural network is proposed and developed based on modern neuroscience, aiming to reflect an abstract mathematical model of the structure and function of the human brain [49]. The BP neural network is a forward multi-layer neural network based on the BP algorithm. Its topology is a hierarchical forward network, which consists of an input layer, a hidden layer, and an output layer.

2.4.1. Artificial Neuron Model

A basic neuron model is shown in Figure 2, which generally contains the following basic elements: a set of connections where the connection strength is represented by the weight of each connection. A positive weight indicates activation, whereas a negative weight indicates inhibition. A summation unit is used to obtain a weighted sum (linear combination) of the input signals. A non-linear activation function acts as non-linear mapping and limits the neuron output amplitude to a certain range, also known as a threshold. The above basic elements can be expressed by Formula (6).

$$u_k={\displaystyle \sum _{j=1}^p{\omega }_{kj}{x}_j,v_k=u_k-{\theta }_k,y_k=\phi (v_k)}$$

(6)

Figure 2. Basic neuron model.

2.4.2. Establishment of the Number of Hidden Neurons

Kolmogorov’s theorem states that the number of hidden neurons is as follows:

$$n_2=\sqrt{n_1+m+1}+a,a=1,\dots ,10$$

(7)

$n_2$ is the number of neurons in the hidden layer, $n_1$ is the number of neurons in the input layer, and m is the number of neurons in the output layer.

2.4.3. Training Function

The training function TRAINLM uses the Levenberg–Marquardt (LM) algorithm to find the optimum value. The LM algorithm [52] is a combination of the gradient descent method and the Gauss–Newton method. This neural network algorithm combines the advantages of these two methods and overcomes the problems of slow convergence and easily falling into local minimum points to a certain extent.

2.4.4. Data Preprocessing

Equation (8) is used to normalize the data.

$${\stackrel{⌢}{a}}_{ij}=\frac{a_{ij}-{\min }_j}{{\max }_j-{\min }_j},i=1,\dots ,n,j=1,\dots m$$

(8)

The expert opinion must be considered in the early warning model. Thus, the data are trimmed using Equation (9).

$${\tilde{a}}_{ij}={\widehat{a}}_{ij}\cdot {\omega }_j,i=1,\dots n,j=1,\dots m$$

(9)

2.4.5. Solving the Weight of Each Factor

The sum of the absolute values of the connection weights between each input layer node and all the hidden layer nodes is calculated. The results are normalized to obtain weights for 22 factors. The formula is shown in Equation (10).

$${\omega }_i=\frac{{\displaystyle \sum _{j=1}^m}\left|v_{ij}\right|}{{\displaystyle \sum _{i=1}^n}{\displaystyle \sum _{j=1}^m}\left|v_{ij}\right|},i=1,2,\dots ,n$$

(10)

3. Data Source and Analysis

3.1. The Improved AHP Is Used to Solve the Expert Weight

3.1.1. Establishing a Hierarchical Structure Model

The 22 factors identified in Table 1 are taken as the index layer, while the three aspects, namely, human factors, management factors, and environmental factors, are taken as the criterion layer to build the hierarchical structure of the unsafe states of construction workers, as shown in Table 6.

Table 6. A hierarchical structure table.

Criterion Layer	Factor
Human factor	Workmate impact, skills level, age of workers, security attitude, work experience, workers’ fatigue, safety compliance, workers’ character; workers’ emotions
Management factor	Group cohesion, safety communication, management level, safety education, security culture, safety regulation, safety supervision; safety input
Environmental factor	Safety climate, construction conditions; protection measures

3.1.2. Constructing Judgment Matrix

Questionnaires are sent to 80 experts, including the project safety director, project manager, the person responsible for safety management, company manager and safety supervisor. All have more than 10 years of engineering practice experience and certain achievements in the field of safety management. Out of the 50 questionnaire results recovered, 37 qualified questionnaires were selected, thereby reaching the number of experts required by the AHP. The judgment matrix was constructed according to the 37 expert questionnaires, where each expert has the same weight.

The judgment matrix of the target layer and the criterion layer is constructed as follows:

$$A=\left[\begin{array}{ccc}1.000& 1.169& 1.313\\ 0.855& 1.000& 1.169\\ 0.762& 0.855& 1.000\end{array}\right]$$

The judgment matrix of human factors at the criterion layer and its index layer is constructed.

$$B_1=\left[\begin{array}{cccccccccc}1.000& 1.636& 1.666& 1.268& 1.356& 1.720& 1.557& 1.499& 1.927& 2.360\\ 0.611& 1.000& 1.636& 1.666& 1.268& 1.356& 1.720& 1.557& 1.449& 1.927\\ 0.600& 0.611& 1.000& 1.636& 1.666& 1.268& 1.356& 1.720& 1.557& 1.449\\ 0.788& 0.600& 0.611& 1.000& 1.636& 1.666& 1.268& 1.356& 1.720& 1.557\\ 0.737& 0.788& 0.600& 0.611& 1.000& 1.636& 1.666& 1.268& 1.356& 1.720\\ 0.581& 0.737& 0.788& 0.600& 0.611& 1.000& 1.636& 1.666& 1.268& 1.356\\ 0.642& 0.581& 0.737& 0.788& 0.600& 0.611& 1.000& 1.636& 1.666& 1.268\\ 0.667& 0.642& 0.581& 0.737& 0.788& 0.600& 0.611& 1.000& 1.636& 1.666\\ 0.519& 0.667& 0.642& 0.581& 0.737& 0.788& 0.600& 0.611& 1.000& 1.636\\ 0.424& 0.519& 0.667& 0.642& 0.581& 0.737& 0.788& 0.600& 0.611& 1.000\end{array}\right]$$

A judgment matrix of management factors at the criterion level and its index level is constructed.

$$B_2=\left[\begin{array}{ccccccccc}1.000& 1.655& 0.909& 0.922& 0.919& 0.986& 0.939& 0.859& 0.873\\ 0.604& 1.000& 1.655& 0.909& 0.922& 0.919& 0.986& 0.939& 0.859\\ 1.100& 0.604& 1.000& 1.655& 0.909& 0.922& 0.919& 0.986& 0.939\\ 1.085& 1.100& 0.604& 1.000& 1.655& 0.909& 0.922& 0.919& 0.986\\ 1.088& 1.085& 1.100& 0.604& 1.000& 1.655& 0.909& 0.922& 0.919\\ 1.014& 1.088& 1.085& 1.100& 0.604& 1.000& 1.655& 0.909& 0.922\\ 1.065& 1.014& 1.088& 1.085& 1.100& 0.604& 1.000& 1.655& 0.909\\ 1.164& 1.065& 1.014& 1.088& 1.085& 1.100& 0.604& 1.000& 1.655\\ 1.145& 1.164& 1.065& 1.014& 1.088& 1.085& 1.100& 0.604& 1.000\end{array}\right]$$

The judgment matrix of environmental factors at the criterion level and its index level is constructed.

$$B_3=\left[\begin{array}{ccc}1.000& 1.147& 1.260\\ 0.872& 1.000& 1.147\\ 0.793& 0.872& 1.000\end{array}\right]$$

3.1.3. Hierarchical Single Ordering and Consistency Check

The results of hierarchical single ordering and consistency checks can be obtained by combining Formulas (2) and (4). The results are shown in Table 7.

Table 7. Hierarchical single sorting and consistency check.

Matrix	λmax	Eigenvector	CI	CR	n	Is CR Less than 0.1?
A	3.000	(0.381, 0.331, 0.287)	0.000	0.000	3	Yes
B1	10.256	(0.149, 0.131, 0.118, 0.110, 0.102, 0.092, 0.085, 0.080, 0.071, 0.061)	0.028	0.036	10	Yes
B2	9.247	(0.108, 0.106, 0.109, 0.110, 0.111, 0.113, 0.115, 0.117, 0.111)	0.031	0.041	9	Yes
B3	3.000	(0.375, 0.332, 0.293)	0.000	0.000	3	Yes

3.1.4. Hierarchical Total Ordering and Consistency Check

The above is the weight vector of a set of elements in comparison to an element in the previous layer. Subsequently, the ranking weight of each scheme in the lowest layer for the target is obtained. The results are shown in Table 8.

Table 8. The weight of influencing factors.

Factor	Weight	Billing	Factor	Weight	Billing
Protection measures	0.1076	1	Workmate impact	0.0367	12
Age of workers	0.0953	2	Workers’ character	0.0367	12
Group cohesion	0.0841	3	Safety regulation	0.0367	12
Safety climate	0.0568	4	Safety supervision	0.0361	15
Security awareness	0.0499	5	Skills level	0.0357	16
Security attitude	0.0450	6	Safety input	0.0351	17
Safety education	0.0419	7	Security culture	0.0351	17
Workers’ emotions	0.0390	8	Security leadership	0.0324	19
Management level	0.0387	9	Work experience	0.0305	20
Construction conditions	0.0381	10	Workers’ fatigue	0.0271	21
Safety communication	0.0374	11	Safety compliance	0.0232	22

$CR=0.033$ is calculated with Formula (5). Given that $CR=0.033<0.1$, the overall ranking result of the hierarchy has a satisfactory consistency. The analysis result is also accepted.

3.1.5. Analysis of Results

According to the results, protection measures, workers’ age, group cohesion, safety climate, and safety awareness have the greatest influence on the safety states of construction workers. Therefore, these five factors should be the primary focus of daily production and construction. Although safety input, safety culture, safety leadership, work experience, worker fatigue, and safety compliance have minimal influence, they cannot be disregarded.

When building an early warning model, expert opinions must be taken into consideration. AHP remains subjective, even though it has been improved. The neural network model is also used for calculating the revised weights to modify the primary weights.

3.2. Using a Neural Network to Build an Early Warning Model

3.2.1. Questionnaire Design

The “Questionnaire Survey on Influencing Factors of Unsafe State of Construction Workers” was developed through data inquiry and expert consultation. The questionnaire comprises fifty-seven questions in three parts. The first part includes basic personal information, while the second part is the measurement of factors, primarily consisting of 22 indicators. Each indicator is reflected through two questions. Each question has five options, corresponding to different points. The third part measures worker safety state levels through six questions, each with five options. The questionnaires were distributed to multiple construction workers. A total of 500 eligible questionnaires were eventually returned. The survey objects are shown in Table 9.

Table 9. Statistics of survey objects’ information.

Worker Information	Attributes	Proportion
Age	18–26	8%
	27–35	22%
	36–44	24%
	45–53	40%
	53 and above	6%
Has the worker experienced an accident	Experienced an accident	24%
	No accident	76%
Area	Sichuan	17%
	Shandong	19%
	Jiangsu	20%
	Jilin	13%
	Others	31%
Education level	Bachelor’s degree and above	1%
	Specialist	15%
	High school	31%
	Junior high school and below	53%
Type of work	Skilled worker	27%
	Ordinary worker	52%
	Manager	14%
	Others	7%
Gender	Female	22%
	Male	78%

3.2.2. Input of the Early Warning Model

Each of the twenty-two indicators is reflected through two questions, with each question having five options, corresponding to 1, 2, 3, 4, and 5 points. Two scores that correspond to each factor are added to obtain the score for that factor. The scores for each factor are normalized using Formula (8). Then, expert opinion is then taken into consideration using Formula (9).

3.2.3. Output of the Early Warning Model

Similar to the input, the safety state is reflected by six questions. The scores that correspond to the questions are added to obtain the safety states score for the worker, which is then normalized using Formula (8).

According to relevant literature [53,54,55] and after consulting experts in the field of safety management, the unsafe states were divided into the following four levels: 1, 2, 3, and 4. The four levels correspond to security, mild risk, moderate risk, and severe risk. The corresponding safety states score intervals are shown in Table 10.

Table 10. Division of security states.

Grade	State	Fractional Interval
1	Security	(0, 0.25)
2	Mild risk	(0.25, 0.5)
3	Moderate risk	(0.5, 0.75)
4	Severe risk	(0.75, 1)

3.2.4. Training of Early Warning Model

(1)

Determining the number of input neurons and the input matrix

Given that 22 influencing factors and 500 questionnaires exist for workers, the number of input neurons is 22, and the size of the input matrix is 500 × 22. The input matrix must be transposed because of the particular requirements of Matlab. The final input matrix size is 22 × 500.

(2)

Determining the number of output neurons and the output matrix

The safety states of workers are reflected by a score, and the number of worker questionnaires is 500. Thus, the number of output neurons is 1, and the size of the output matrix is 500 × 1. The output matrix must be transposed because of the particular requirements of Matlab. The final output matrix size is 1 × 500.

(3)

Determining the number of hidden neurons

According to the Kolmogorov theorem, the number of hidden layer nodes is chosen to be eight. The final neural network model is shown in Figure 3. (W is the weight of the input to the neuron, and b is the bias of the neuron.)

Figure 3. Neural network model.

(4)

Levenberg–Marquardt algorithm

The BP algorithm has some shortcomings, such as slow convergence, low training efficiency and poor numerical stability. The non-linear neural network learning algorithm named the Levenberg–Marquardt (LM) algorithm can greatly improve the above defects [50]. The LM algorithm is a combination of the gradient descent method and quasi-Newton method. The algorithm expects to achieve a high-order training speed without calculating the Hessian matrix and the formula is shown in Equation (11).

$$x_{k+1}=x_k-{[J^{T}J+\lambda I]}^{-1}{J}^T$$

(11)

In the formula, $\mathrm{J}$ is the Jacobian matrix. When using the Levenberg–Marquardt algorithm, if the decline is too fast, one can use a smaller $\mathsf{\lambda }$ to increase its similarity to the Gaussian Newton method. Alternatively, one can use a larger $\mathsf{\lambda }$ to increase its similarity to the gradient descent method. The Matlab toolbox provides the TRAINLM function for the calculation of the Levenberg–Marquardt algorithm.

(5)

Importing data and determining parameters

In this study, the neural network toolbox in Matlab was utilized to train the neural network model. The input and output matrices were imported into Matlab separately. During the model training process, the training function was TRAINLM, the adaptive adjustment learning function was LEARNGDM, the performance function was the mean square error, the number of network layers was 2, the number of training times was 1000, the target error was 0, and the other parameters were set to the default values of the toolbox.

3.2.5. Training the Model

During the training process, the neural network toolbox divides the samples into the following three parts: training samples, validation samples, and test samples. The variation in the mean square error with the number of iterations is shown in Figure 4. The training result demonstrates good performance for the fourth iteration, with a mean squared error of 0.019765.

Figure 4. The variation in mean square error.

3.2.6. Model Accuracy Test

Following the completion of the model training, the questionnaire survey was conducted again, and resulted in the acquisition of 214 samples of data. The 214 samples of data were predicted, and the predicted output was compared with the expected output of the sample to test the validity of the model. The comparison results of the selected samples are shown in Table 11.

Table 11. Results of model testing.

Sample Number	True Score	Actual Grade	Actual States	Predictive Score	Prediction Grade	Prediction State	Is the State Predicted Correctly?
1	0.304	2	Mild risk	0.265	2	Mild risk	Y
2	0.391	2	Mild risk	0.380	2	Mild risk	Y
3	0.782	4	Severe risk	0.723	3	Moderate risk	N
4	0.782	4	Severe risk	0.720	3	Moderate risk	N
5	0.304	2	Mild risk	0.273	2	Mild risk	Y
6	0.695	3	Moderate risk	0.729	3	Moderate risk	Y
7	0.043	1	Security	0.257	2	Mild risk	N
8	0.739	3	Moderate risk	0.680	3	Moderate risk	Y
9	0.608	3	Moderate risk	0.685	3	Moderate risk	Y
10	0.173	1	Security	0.252	2	Mild risk	N
11	0.217	1	Security	0.257	2	Mild risk	N
12	0.565	3	Moderate risk	0.690	3	Moderate risk	Y
13	0.652	3	Moderate risk	0.728	3	Moderate risk	Y
14	0.260	2	Mild risk	0.286	2	Mild risk	Y
15	0.391	2	Mild risk	0.401	2	Mild risk	Y

The experimental results indicate that with 180 test samples, accurate predictions were made with an accuracy rate of 84.11%. The expected level and the actual level were found to be in close alignment, indicating the reliability of the experimental outcomes. This indicates that the early warning model can accurately predict and further research can be conducted.

3.2.7. Weight Correction of Each Factor Indicator

The connection weight matrix is exported from the input layer of the model to the hidden layer. The connection weight matrix is shown in Table 12.

Table 12. The connection weight matrix.

	Hidden Layer 1	Hidden Layer 2	Hidden Layer 3	Hidden Layer 4	Hidden Layer 5	Hidden Layer 6	Hidden Layer 7	Hidden Layer 8
Input layer 1	0.367	0.344	−0.523	0.445	0.117	−0.261	−0.312	0.162
Input layer 2	0.541	0.404	−0.502	0.513	0.492	0.156	0.086	−0.221
Input layer 3	−0.111	0.293	0.134	0.571	0.134	−0.325	0.182	0.686
Input layer 4	0.179	0.067	0.036	0.109	0.117	−0.147	−0.099	−0.257
Input layer 5	−0.254	−0.482	−0.315	0.425	0.199	−0.045	0.331	−0.544
Input layer 6	−0.075	−0.172	0.158	0.422	−0.397	0.106	−0.193	0.241
Input layer 7	0.262	0.224	−0.396	0.259	0.156	−0.234	−0.49	0.021
Input layer 8	0.526	−0.291	−0.07	−0.174	0.25	−0.101	−0.144	0.415
Input layer 9	0.444	0.368	0.033	−0.27	−0.458	−0.102	0.045	−0.326
Input layer 10	0.344	−0.388	0.429	−0.069	−0.391	−0.091	−0.033	0.113
Input layer 11	−0.182	0.232	−0.053	0.171	0.433	−0.034	−0.005	0.234
Input layer 12	−0.135	0.003	−0.675	−0.372	−0.004	0.476	0.41	−0.392
Input layer 13	0.095	−0.016	−0.334	−0.283	−0.413	0.424	−0.236	−0.059
Input layer 14	−0.384	0.084	−0.21	0.266	0.187	0.392	−0.247	−0.384
Input layer 15	−0.298	0.359	−0.275	0.335	0.029	0.669	−0.515	−0.063
Input layer 16	−0.425	0.495	−0.381	0.262	0.359	0.429	−0.506	−0.186
Input layer 17	−0.292	0.219	0.019	0.537	−0.414	−0.094	−0.581	−0.325
Input layer 18	0.421	−0.018	−0.171	−0.164	0.37	0.312	−0.444	0.051
Input layer 19	−0.103	−0.459	−0.168	−0.288	−0.408	−0.136	−0.358	−0.33
Input layer 20	0.462	0.3	−0.318	0.219	−0.242	0.646	−0.424	−0.218
Input layer 21	0.337	−0.138	−0.096	−0.073	−0.504	−0.249	0.277	0.465
Input layer 22	0.079	−0.507	−0.341	−0.121	0.341	−0.405	−0.625	−0.243

According to Formula (10), the revised weight of the factor is calculated, as shown in Table 13.

Table 13. The revised weight.

Factor	Weight	Criterion Layer	Factor	Weight	Criterion Layer
Safety communication	0.062	Management factor	Safety regulation	0.044	Management factor
Security awareness	0.060	Human factor	Age of workers	0.044	Human factor
Protection measures	0.058	Environmental factor	Workers’ fatigue	0.042	Human factor
Group cohesion	0.054	Management factor	Security leadership	0.042	Management factor
Workers’ emotions	0.053	Human factor	Work experience	0.040	Human factor
Workmate impact	0.052	Human factor	Management level	0.040	Management factor
Safety climate	0.052	Environmental factor	Safety supervision	0.038	Management factor
Construction conditions	0.051	Environmental factor	Safety compliance	0.038	Human factor
Safety input	0.050	Management factor	Security culture	0.036	Management factor
Security attitude	0.050	Human factor	Skills level	0.027	Human factor
Workers’ character	0.046	Human factor	Safety education	0.021	Management factor

3.3. Analysis of Results

3.3.1. Weight Correction Analysis

Based on the weight ranking results obtained by analyzing the weight correction value, we determined that security communication and security awareness accounted for the largest proportion, which is above 6%, as shown in Table 13. Thus, workers should be regularly invited to seminars. They should also be encouraged to share safety experiences, discuss safety methods, and enhance communication among workers. Moreover, some experts should be invited to share safety knowledge to improve the safety awareness of construction workers. Security communication and security awareness are followed by protection measures, group cohesion, worker sentiment, worker influence, safety atmosphere, construction conditions, safety input, and safety attitude. The weight of these eight indicators is between 5% and 6%, and the impact on workers’ unsafe behavior cannot be underestimated. On the one hand, these factors are mostly human factors, which require great attention. The relevant indicators should be monitored as key indicators. On the other hand, improving a safe environment involves vigorously publicizing the safety protection slogan and conducting regular environmental checks to protect workers and ensure their safety.

3.3.2. Comparative Analysis of Primary Weight and Revised Weight

The serial numbers corresponding to each factor are defined, as shown in Table 14. The comparison of the primary weight and the revised weight is shown in Figure 5.

Table 14. The serial number of factors.

Serial Number	Factor	Serial Number	Factor
1	Safety climate	12	Safety input
2	Security awareness	13	Safety supervision
3	Security attitude	14	Safety regulation
4	Safety education	15	Workmate impact
5	Workers’ emotions	16	Safety communication
6	Security culture	17	Construction conditions
7	Security leadership	18	Management level
8	Work experience	19	Workers’ character
9	Workers’ fatigue	20	Protection measures
10	Safety compliance	21	Age of workers
11	Skills level	22	Group cohesion

Figure 5. The comparison of the primary weight and the revised weight.

Figure 5 shows that workers and experts differ considerably in their perceptions of certain factors, including safety education, protective measures, worker age, group cohesion, worker sentiment, safety engagement, coworker influence, safety communication, and construction conditions. Experts place greater emphasis on safety education, protective measures, worker age, and group cohesion, whereas workers prioritize worker sentiment, safety input, coworker influence, safety communication, and construction conditions more than experts do. These differences can be attributed to the factors to which workers and experts are exposed in their daily routines.

4. Engineering Application

The above shows that the mean square error of the neural network reaches 0.019765, and the prediction accuracy of the selected 214 questionnaires was 84.11%. This predictive model can be applied to actual engineering projects. To carry out further research, a construction project of a primary school in Suzhou was selected as the research object.

4.1. Worker Forecast

Five types of jobs were selected from the workers of the primary school construction project, including concrete workers, steel workers, masonry workers, frame workers, and plasterers. Eight people were randomly selected from each job type to give warning. The warning results are shown in Table 15.

Table 15. Prediction results of workers in the primary school construction project.

Worker Number	Type of Work	Predictive Score	Prediction Grade	Early Warning States	Worker Number	Type of Work	Predictive Score	Prediction Grade	Early Warning States
1	Plasterer	0.32	2	Mild risk	17	Frame workers	0.267	2	Mild risk
2		0.339	2	Mild risk	18		0.268	2	Mild risk
3		0.361	2	Mild risk	19		0.27	2	Mild risk
4		0.367	2	Mild risk	20		0.272	2	Mild risk
5		0.378	2	Mild risk	21		0.721	3	Moderate risk
6		0.385	2	Mild risk	22		0.724	3	Moderate risk
7		0.386	2	Mild risk	23		0.731	3	Moderate risk
8		0.393	2	Mild risk	24		0.735	3	Moderate risk
9	Steel workers	0.282	2	Mild risk	25	Masonry workers	0.32	2	Mild risk
10		0.284	2	Mild risk	26		0.339	2	Mild risk
11		0.287	2	Mild risk	27		0.361	2	Mild risk
12		0.289	2	Mild risk	28		0.367	2	Mild risk
13		0.29	2	Mild risk	29		0.378	2	Mild risk
14		0.713	3	Moderate risk	30		0.385	2	Mild risk
15		0.715	3	Moderate risk	31		0.386	2	Mild risk
16		0.718	3	Moderate risk	32		0.71	3	Moderate risk

Based on the statistics, 36 people were found to be in a mild risk state, and 12 workers were in the moderate risk state. Given that majority of the workers were observed in an unsafe state, it is imperative to promptly adjust workers’ behavior.

Based on the work type statistics, the safety states were sorted into plasterers, masonry workers, concrete workers, steelworkers, and frame workers. Steelworkers and frame workers were found to have a higher probability of being in an unsafe state. Thus, construction units should give special attention to their safety states.

4.2. Sensitivity Analysis

Several workers’ safety states were observed to be approaching severe danger. To further investigate this, we selected the four workers closest to severe danger and controlled the variables. We focused on one factor at a time for each worker, adjusted the input, and made predictions. The ordinal numbers that corresponded to the factors in Table 14 were utilized.

In the data, 0 values were set to 0.001 to ensure that the data could be adjusted. The changes in the safety states of four workers were predicted when a single factor increased by 500%, 700%, 900%, and 1100%. The data with large changes in the safety state were selected. The results are shown in Table 16 (I: security; II: mild risk; III: moderate risk; IV: severe risk).

Table 16. Sensitivity analysis results.

Worker Number	Factor	500% Change		700% Change		900% Change		1100% Change
		Score	State	Score	State	Score	State	Score	State
1	Safety climate	0.674	III	0.6	III	0.497	II	0.406	II
1	Security leadership	0.496	II	0.364	II	0.307	II	0.289	II
1	Safety communication	0.673	III	0.604	III	0.516	III	0.43	II
2	Safety climate	0.43	II	0.313	II	0.287	II	0.282	II
2	Security leadership	0.581	III	0.477	II	0.38	II	0.32	II
2	Workmate impact	0.632	III	0.554	III	0.488	II	0.438	II
2	Safety communication	0.65	III	0.579	III	0.503	III	0.432	II
2	Protection measures	0.652	III	0.588	III	0.524	III	0.471	II
3	Security leadership	0.606	III	0.527	III	0.441	II	0.369	II
3	Workmate impact	0.632	III	0.569	III	0.518	III	0.478	II
4	Safety climate	0.679	III	0.618	III	0.521	III	0.42	II
4	Workmate impact	0.632	III	0.569	III	0.508	III	0.456	II
4	Safety communication	0.48	II	0.365	II	0.312	II	0.31	II
4	Protection measures	0.643	III	0.595	III	0.544	III	0.495	II

Table 16 shows that safety climate, safety leadership, safety communication, worker influence, and protective measures were the key aspects to be adjusted.

4.3. Follow-Up Research

We reanalyzed the safety states of these four workers after 1 month. The results are shown in Table 17. As indicated in Table 16, the predicted scores of the four workers were observed to decrease.

Table 17. Comparison of safety states before and after improvement.

Worker Number	Before Training			After Training
	Score	Grade	State	Score	Grade	State
1	0.735	3	Moderate risk	0.67	3	Moderate risk
2	0.731	3	Moderate risk	0.671	3	Moderate risk
3	0.724	3	Moderate risk	0.694	3	Moderate risk
4	0.721	3	Moderate risk	0.672	3	Moderate risk

5. Conclusions

The unsafe behavior of construction workers is the primary cause of engineering safety accidents. This study focuses on the construction workers group and evaluates 22 safety factors as research objects. An improved AHP and BP neural network model is developed to build a safety warning system. Furthermore, the study incorporates sensitivity analysis into safety monitoring and warning by comprehensively analyzing data from both construction workers and experts. The prewarning model is applied to a primary school project in Suzhou, where a sensitivity analysis of the four workers closest to the severe danger state is conducted, and suggestions for safety management measures are given. The conclusions are as follows: (1) experts and construction workers have different concerns regarding safety factors. Experts believe that protection measures, workers’ age, group cohesion, safety atmosphere, and safety awareness significantly influence the unsafe states of construction workers. Conversely, workers believe that safety communication, safety awareness, protection measures, and group cohesion have a great impact. Nevertheless, both groups concur on the weight of construction workers’ influencing factors, where protection measures, group cohesion, and safety awareness are highly significant, while safety compliance and security culture are insignificant. (2) The safety warning model established in this study effectively monitors the unsafe states of construction workers. (3) The safety prewarning model and sensitivity analysis enable targeted proposals for safety management measures.

In theory, this study considers the opinions of workers and experts on the influencing factors. On the one hand, the results can guide real-world measures. On the other hand, they can also deepen the understanding and communication between workers and experts. Moreover, the prewarning model constructed can be applied to specific projects. It can predict the safety state of workers to warn individuals and can also suggest corrective measures through sensitivity analysis. Once an accident happens, the safety states of the workers can also be re-evaluated. The changed factors and relevant measures were observed to provide a reference for workers and their peers and raised awareness of these important factors. However, there are still some shortcomings in this study, which need to be further discussed. Although the AHP has been improved, eliminating the influence of experts and workers’ subjectivity on the research results remains difficult. Data can be further collected to reduce the influence of subjectivity on the research and ensure the accuracy of the model training. The prediction accuracy of the model can be improved by adjusting the neural network parameters and structure or using advanced technology. Moreover, this practice should be strengthened, its application must be given attention, and the prewarning model should be applied for continuous improvement.