Abstract

The emergency management of chemical accidents plays an important role in preventing the expansion of chemical accidents. In recent years, the evaluation and research of emergency management of chemical accidents has attracted the attention of many scholars. However, as an important part of emergency management, the professional rescue team of chemicals has few evaluation models for their capabilities. In this study, an emergency rescue capability assessment model based on the PCA-BP neural network is proposed. Firstly, the construction status of 11 emergency rescue teams for chemical accidents in Shanghai is analyzed, and an index system for evaluating the capabilities of emergency rescue teams for chemicals is established. Secondly, the principal component analysis (PCA) is used to perform dimension reduction and indicators’ weight acquisition on the original index system to achieve an effective evaluation of the capabilities of 11 rescue teams. Finally, the indicators after dimensionality reduction are used as the input neurons of the backpropagation (BP) neural network, the characteristic data of eight rescue teams are used as the training set, and the comprehensive scores of three rescue teams are used for verifying the generalization ability of the evaluation model. The result shows that the proposed evaluation model based on the PCA-BP neural network can effectively evaluate the rescue capability of the emergency rescue teams for chemical accidents and provide a new idea for emergency rescue capability assessment.

1. Introduction

Due to the properties of hazardous chemicals, such as toxicity, corrosiveness, explosiveness, flammability, and combustion support, there are huge risks in their production, transportation, storage, sales, use, and disposal. Once a hazardous chemical accident occurs, it is easy to cause many casualties, huge property losses, and serious environmental pollution and bring catastrophic consequences to both enterprises and the society. For example, the explosion of a dangerous good warehouse in Tianjin Port on August 15, 2015, resulted in 165 deaths, 798 casualties, and 8 missing. The direct economic loss reached 6.866 billion yuan [1]. Therefore, the emergency treatment of chemical accidents must be timely and efficient to prevent accidents from expanding and causing even greater losses.

In recent years, the assessment of emergency management capabilities for hazardous chemical accidents has attracted the attention of many scholars. Wang et al. [2] proposed a disaster management control capability assessment model based on the Capability Maturity Model (CMM). This model evaluates the capability of the organization from eight aspects and divides the capability assessment results into four levels, which provides general assessment guidelines for different types of emergency management organizations. Lin [3] analyzed the nature of emergency rescue capabilities from the perspective of the entire city and established an urban emergency rescue capability evaluation system based on AHP and Fuzzy Comprehensive Evaluation (FCE). Yang et al. [4] analyzed many factors that affect the emergency capacity of enterprises, established an emergency capacity evaluation index system, and determined the weight for each indicator through the Analytic Hierarchy Process (AHP). Yu and Guan [5] analyzed the current situation and difficulties of emergency treatment of hazardous chemical accidents, discussed the emergency training system, and provided a reference for improving the emergency capabilities of professional teams for fire and hazardous chemical rescue. Zhu et al. [6] used Bayesian networks to propose a framework for dynamically evaluating explosion accidents in chemical plants to support prevention, management, and real-time warning. He et al. [7] established a Petri net model of emergency process of chemical accidents in order to evaluate the emergency capability, which can dynamically evaluate the emergency capability of chemical accidents.

In addition to the abovementioned traditional evaluation methods, the application of artificial neural networks to the evaluation of chemical accidents has made some progress. Yuan et al. [8] used back propagation neural networks, generalized regression neural networks and radial basis function neural networks to evaluate the safety production management of chemical companies and found that the prediction ability of the radial basis function neural network is more accurate. Aiming at the shortcomings of the current chemical production safety evaluation system and combining the knowledge of artificial neural networks, Yang established a new evaluation index system and proposed the advantages of applying neural networks to the chemical production safety evaluation system [9].

These studies mainly focus on the establishment of an emergency management assessment system for hazardous chemical accidents and the application of emergency assessment methods. Or, consider the emergency management activity itself as a project management process and study the capacity assessment model for emergency management control. However, as an important part of emergency management, the professional rescue team of hazardous chemicals has few evaluation models for their capabilities. In addition, traditional evaluation methods, such as Analytic Hierarchy Process (AHP), are greatly affected by human factors in the implementation process, and it is difficult to obtain objective evaluation results. When there are many evaluation indicators, it will complicate the structure of the artificial neural network model and increase the computational complexity. Shanghai is an important petrochemical and fine chemical industry base in China, with a solid chemical industry foundation. Through the assessment of the emergency response capabilities of the 11 professional rescue teams for hazardous chemicals in Shanghai, the capabilities of the rescue teams can be strengthened using targeted countermeasures.

In order to reasonably evaluate the capabilities of professional emergency rescue teams for hazardous chemical accidents, this study surveyed 11 professional rescue teams in Shanghai, analyzed the status of these teams, constructed a rescue capacity assessment index system, and built a rescue capability evaluation model combined with BP neural network. At the same time, in order to determine the indicators’ weight and reduce the number of neurons in the input layer of the backpropagation (BP) neural network, the principal component analysis (PCA) was used to reduce the dimension of the evaluation index system and obtain weight. The dimensionality-reduced feature factors were used as the input units of the BP neural network. This method can not only reduce the influence of human factors in the evaluation process but also simplify the structure of the artificial neural network and reduce the computational complexity of the evaluation model. The trained BP neural network evaluation model can well evaluate the capabilities of professional emergency rescue teams for hazardous chemical accidents, providing a new idea for emergency rescue capability assessment.

2. Methods

2.1. Construction of Rescue Capability Evaluation Index System

Shanghai has a total of 11 emergency rescue teams for production safety, as shown in Table 1. At present, the 11 emergency rescue teams for production safety are managed by the company where they work. The Shanghai Emergency Management Bureau is responsible for providing business guidance. As a result of a thorough investigation, following construction problems were found with these teams:(1)Inefficient cooperation mechanism: there is a lack of coordination between the emergency rescue team and other departments. The team’s responsibilities are unclear, and there is no unified command.(2)Slow emergency response: the lack of classification and corresponding response plans based on the type and scale of hazardous chemical accidents makes the emergency rescue scene more chaotic.(3)Inappropriate team building: the positioning of these rescue teams is unclear, and the rescue areas are not divided. There are no long-term full-time members in these rescue teams, and these team members have not received any specialized training in dealing with hazardous chemical accidents.(4)Insufficient emergency equipment, materials, and maintenance funds. The maintenance of professional rescue equipment for hazardous chemical accidents lacks government financial support. The necessary equipment and materials cannot be timely supplemented.(5)Noncompliant emergency handling. There is a dearth of norms and standards on the emergency handling of hazardous chemical accidents. There is no targeted emergency response plan for different hazardous chemical accidents. Due to the problem of team building, rescuers with nonprofessional characteristics, and temporary combination, it is difficult to conduct emergency response scientifically and quickly.

According to the above construction status, an emergency rescue team assessment index system was established. The emergency response capabilities of professional rescue teams for hazardous chemical accidents include the following elements:(1)Emergency cooperation: the main consideration is whether the division of responsibilities within the rescue team is reasonable and clear and whether smooth information can be communicated between various departments; whether the management of human resources has considered a perfect reward and punishment system, employee benefits and incentives; and whether the communication is efficient enough to ensure the normal operation of the emergency mechanism.(2)Emergency command: the emergency rescue of hazardous chemical accidents mainly includes two aspects: an emergency disposal plan and the emergency professional and technical personnel. Different schemes are needed to respond to different chemical accidents as the quantity and types of hazardous chemicals always vary between different plants and regions. Therefore, certain requirements are put forward for the pertinence and completeness of the emergency response plan and the allocation of emergency professional and technical personnel.(3)Emergency foundation: personnel, materials, equipment, and funds are the basis for emergency rescue of hazardous chemical accidents. In this study, the factors that affect basic emergency support are divided into four parts: (1) the emergency team, considering whether the stability, quantity, and quality of emergency personnel; (2) emergency equipment (including personnel protection equipment), the functionality, safety, quality, and quantity of equipment should satisfy the emergency disposal requirements; (3) whether emergency materials could meet different types of hazardous chemical accidents; (4), emergency funding, whether governments and enterprises had been given economic support to ensure better operation of the emergency rescue teams.(4)Training and education: consider the training of professional emergency rescue knowledge and skills. Assess the improvement of the emergency rescue ability of the corresponding emergency personnel.(5)Emergency drills: consider the workload, such as whether the number and time-frequency of drills is reasonable to meet the demands. In addition, the factors that need to be considered are the effects of the emergency drills, whether the personnel is familiar with the emergency procedures and more effective in responding to special chemical accidents through emergency drills.

Accordingly, this research proposed a rescue capability evaluation index system that includes 5 first-level indicators, 14 second-level indicators, and 28 third-level indicators, as shown in Table 2:

2.2. Dimension Reduction of Evaluation Indicators and Weight Acquisition

The above index system is too complicated; using the original indicator as the input unit of the BP neural network will face problems of such as high data dimensions, poor fitting effects, and inaccurate prediction results. Therefore, the principal component analysis (PCA) was required to reduce the dimensions of the indicators to eliminate the correlation [10].

Principal component analysis (PCA) is an important statistical method that uses the idea of dimensionality reduction to transform multiple indicators into a few comprehensive indicators. These comprehensive indicators are not explanatory but retain most of the original information [11]. The new comprehensive indicator is a linear combination of all the original indicators which remain independent of each other. The principal component analysis (PCA) can reduce the number of evaluation indicators, thereby reducing the number of neurons in the input layer to simplify the structure of the BP neural network.

In a geometric sense, the principal component analysis (PCA) method is to project the original data onto a new coordinate axis, which is the principal component. In order to enable the principal component to contain more information about the original data, the variance of the principal component must be maximized. The process of principal component analysis (PCA) is to find linear combination coefficients. The coefficients must maximize the variance of the principal components, and the sum of the squares of the coefficients must be equal to one. In addition, starting from the second principal component, each principal component must be independent of the existing principal components.

The specific steps of principal component analysis (PCA) are as follows: Step 1: standardize the raw data. Assuming the original data is an n × m matrix: The rows in the data matrix represent different samples, and the columns represent different evaluation indicators. It can be seen that there are n samples and m evaluation indicators. The matrix can be normalized by the following formula: among them, Variance: Step 2: calculate the correlation coefficient matrix of the sample indicators: among them, Step 3: calculate the eigenvalue and eigenvector of the correlation coefficient matrix. Step 4: select the principal component . All the eigenvalues are arranged in descending order. The larger the eigenvalue, the more the system information contained in the principal component. Calculate the contribution rate of each principal component by the following formula: It is generally considered that the cumulative contribution rate of the first principal components exceeds 85% is reasonable, indicating that the total amount of system information they contain exceeds 85%. At this time, the principal component is the characteristic index after dimensionality reduction. Step 5: find the unit orthogonal feature vector of the first feature vectors. Each principal component is a linear combination of all the original indicators, and the coefficient is the element of the unit orthogonal eigenvector: Then, the expression of the principal component of the -th sample is Step 6: the comprehensive evaluation function of the -th sample is shown in the following formula:

 The weight of the principal component is the contribution rate .

2.3. Construction of Evaluation Model Based on BP Neural Network

BP neural network is a multilayer feedforward network model. Its network is mainly composed of three parts: input layer, hidden layer, and output layer. As shown in Figure 1, it can map -dimensional data to -dimensional data. The neurons in each layer of the BP neural network are not connected, and the output of the neurons in each layer only affects the output of the next layer. At the same time, the network will backpropagate errors during operation to continuously adjust the weights and thresholds of the network to achieve self-adjustment [12]. BP neural network is a nonlinear adaptive system, so it is more suitable for dealing with fuzzy or nonlinear problems. This method can effectively reduce subjective factors in the evaluation process and reduce the evaluation time [13].

Figure 1 

Structure of BP neural network.

The main parameter settings of the BP neural network are the number of network layers, the number of nodes in the hidden layer, the transfer function, and the training function, which will be introduced one by one as follows.

2.3.1. Number of Network Layers

As the number of network layers increases, the structure of the BP neural network will become more and more complex. Correspondingly, the complex BP neural network will prolong the learning time and cause the phenomenon of “overfitting.” Through previous testing and research, the neural network is generally set up as a three-layer network, that is, the input layer, the hidden layer, and the output layer [14].

2.3.2. Number of Hidden Layer Nodes

Too few hidden layer neurons may not train the desired network, or the trained network is not strong enough and has poor generalization ability. To contrary, too many hidden layer neurons will increase the learning time and the error may not be smaller. The number of neurons in the hidden layer can be obtained by the following empirical formula [15]:where is the number of neurons in the hidden layer, is the number of neurons in the input layer, is the number of neurons in the output layer, and is a constant between [1, 10].

2.3.3. Transfer Function

There are three main transfer functions: logsig function, tansig function, and purelin function. The choice of the transfer function of the hidden layer and the output layer has a greater impact on the prediction accuracy of the BP neural network. The logsig transfer function is an S-type logarithmic function, the tansig function is an S-type hyperbolic tangent function, and both are nonlinear functions. After that, the purelin function is a linear function. Generally, the transfer function of the hidden layer selects the logsig function or the tansig function, and the transfer function of the output layer selects the purelin function. As far as the nonlinear transfer function is concerned, if the output of the samples is greater than zero, the logsig function is mostly used; otherwise, the tansig function is used.

2.3.4. Training Function

The common training functions are as follows: trainlm: LevenbergMarquardt method traingd: gradient descent method traingdm: gradient descent method with momentum factor traingda: gradient descent method with adaptive learning rate traingdx: gradient descent method with adaptive learning rate and momentum factor

3. Data Processing

3.1. Data Sources

The abovementioned evaluation index system was determined by analyzing the construction status of 11 emergency rescue teams for hazardous chemical accidents in Shanghai. Then, this paper produced a score sheet for these professional rescue teams. To reduce the subjective arbitrariness when scoring, the results were divided into 5 levels, and each level was guaranteed to take a positive value, from the worst to the best, respectively, 1, 2, 3, 4, and 5 points. See Table 3 for scoring criterion. The scorer could select the corresponding score according to the actual situation. Table 4 shows the scores of the 11 rescue teams:

3.2. Data Dimensionality Reduction and Weight Determination

In this research, SPSS 25.0 software was used to perform the principal component analysis (PCA) of the original data to achieve data reduction and weight determination. The eigenvalues of the correlation coefficient matrix between 28 indicators, the contribution rate and the cumulative contribution rate of the principal components (Table 5), and the factor load matrix (Table 6) could be automatically calculated through the factor analysis tool in SPSS software. As can be seen from Table 5, there are 7 eigenvalues greater than 1, and the cumulative contribution rate of the corresponding 7 principal components reaches 94.743%, which meets the requirement that the cumulative contribution rate is generally greater than 85%. However, as shown in Table 6, none of the last two of the 7 principal components exceeds 0.5, and the cumulative contribution rate of the first 5 principal components reaches 85.849%, which also meets the requirement of greater than 85%. Therefore, this study chose the first 5 principal components as input units of the final BP neural network. Although the unit orthogonal eigenvector could not be directly obtained by the factor analysis tool of SPSS software, it could be calculated by the relationship formula (12) between the unit orthogonal eigenvector and the factor load:where L_t is the load of principal component P_t. So the coefficients of the five principal component linear expressions were known (Table 7).

The contribution rates of the 5 principal components are weights, and the normalized weights are shown in Table 8. The specific values of the 5 principal components in the 11 rescue teams were obtained through Table 7 and formula (9), and the comprehensive scores of the emergency capabilities of the 11 rescue teams were finally calculated through formula (10), as shown in Table 9:

As can be seen from Table 9, among the 11 rescue teams, team 3 gets the highest score of 12.441 points, while team 8 gets the lowest score of only 7.541 points. The difference between the highest and lowest scores is 4.9 points. Compared with the original scoring Table 4, it can be found that team 3 scores 4 or 5 points except D23 and D24. The difference is that for team 8, except for D22, all other indicators are 1 or 3 points. Therefore, the comprehensive scores obtained by the principal component analysis method are consistent with actual situations and can effectively reflect the emergency response status of 11 rescue teams.

4. Implementation of BP Neural Network

4.1. Sample Data Normalization

Normalization refers to limiting the input and output data of the network to [0, 1] or [−1, 1] through the processing of variables, which can improve the efficiency of the transfer function and the accuracy of the output of the neural network. There is a maximum-minimum method to limit the data between [0, 1], and the function form is as follows:

The formula that limits the data to [−1, 1] is

In this paper, the mapminmax function provided by MATLAB was used to obtain the normalized input and output data between [−1, 1]. The normalized data are shown in Table 10:

Teams 1 to 8 were used as training samples for the BP neural network, and teams 9 to 11 were used as prediction samples for the BP neural network.

4.2. Determination of the Transfer Function

In this study, the data were limited to [−1, 1] during normalization. It could be known from the above description of the transfer function that the tansig function should be used. Therefore, it was determined that the transfer function of the hidden layer was the tansig function, and the transfer function of the output layer was selected as the purelin function.

4.3. Selection of Training Function

In order to determine the fast and accurate training function, this research used BP neural network toolbox of MATLAB software to experiment the above five training functions and then compared the training results to choose. According to the previous principal component analysis (PCA), five input neurons have been identified, named , , , , and , and only one output neuron which was “comprehensive score.” From the above empirical formula, it could be known that the number of hidden layer neurons should be selected between [4, 13], and it was temporarily determined to be 9. The number of iterations and convergence accuracy were used as the evaluation indicators for the training function selection. Before using the above five training functions to perform prediction fitting on 8 training samples, the maximum number of iterations was set to 2000 and the target convergence accuracy was set to 0. The results are shown in Table 11:

From the above training results, it can be seen that the trainlm training function achieves high accuracy in only 5 steps. The trainlm training function has the fastest convergence speed, but it is easy to fall into a local minimum. The traingd function and trackingdm function converge slowly in practical applications, and the convergence accuracy is not as high as the other three training functions. The traingda function and traingdx function have greatly improved the convergence accuracy of training, but the traingdx function converges faster than the traingda function and the traingdx function can avoid falling into local minima due to the additional momentum term and adaptive learning rate. At the same time, the training precision of the traingdx function reaches 8.3968  10⁻¹⁰, which is consistent with the convergence accuracy in general cases. Therefore, the BP neural network model established in this paper used the traingdx function as the training function.

4.4. Setting of Training Parameters

In this study, two parameters of the BP neural network model were set. The maximum allowable error was set to 0.00001, the maximum number of learning times was set to 1000, and the remaining parameters adopted default values.

4.5. Determination of the Number of Hidden Layer Neurons

The value range of hidden layer neurons has been obtained through the empirical formula in the previous article. However, if the number of hidden layer neurons is too small, the ability of the neural network to obtain information from the sample is poor, and it is impossible to generalize and reflect the sample law. At the same time, if there are too many neurons, the irregular content in the sample may be learned and the phenomenon of “overfitting” may occur. Therefore, it is necessary to determine an optimal number of hidden layer neurons.

The number of hidden layer neurons can be determined one by one [4, 13] through experiments. The number of training sessions was set to 1000, and the target accuracy was set to 0.00001. The number of iterations and the convergence accuracy of 10 trainings were compared to determine the appropriate number of hidden layer neurons. The training results are shown in Table 12:

As shown in Table 12, the convergence accuracy is on an order of magnitude, but there is some gap in the number of iterations. When the number of hidden layer neurons is 8, 9, 12, and 13, the number of iterations is small. From the perspective of simplifying the structure of the BP neural network, this study determined that the number of hidden layer neurons was 8. The resulting structure of BP neural network is shown in Figure 2:

Figure 2 

The structure of the BP neural network model.

4.6. Computational Complexity

The number of neurons in a neural network has an important impact on the computational complexity. When the number of neurons increases, the network calculation becomes more complex. The complexity of calculating the gradient of a certain layer is , assuming that the number of neurons in this layer is D. This study uses principal component analysis to reduce the number of neurons in the input layer of the neural network and obtained the final structure of neural network: 5 neurons in the input layer, 8 neurons in the hidden layer, and 1 neuron in the output layer. Therefore, the final calculation of the entire neural network is . In contrast, the computational complexity of the input layer of the neural network that directly uses all the evaluation indicators as the input layer neurons is , which is far greater than the computational complexity of the optimized neural network.

5. Simulation and Result Analysis of BP Neural Network Evaluation Model

5.1. Training of BP Neural Network

After the above discussion, the parameters of the BP neural network evaluation model had been determined. The first 8 teams were now used as training samples to train the BP neural network. The training result is shown in Figure 3. After 150 iterations, the mean-squared error of the neural network reaches 9.688e − 6. As shown in Table 13, the maximum relative error between the training results and the comprehensive score is only 0.146%, indicating that the BP neural network has reached the training requirements.

Figure 3 

Mean square error trace.

5.2. Simulation of BP Neural Network

The trained BP neural network evaluation model was used to predict the comprehensive scores of the remaining three rescue teams. The relative errors between the predicted results and the theoretical values are shown in Table 14. It can be known from Table 14 that the maximum relative error is 6.658%, which indicates that the generalized ability of the trained BP neural network evaluation model can meet the emergency rescue capability assessment needs of rescue teams.

6. Conclusion

The aim of this paper is to assess the capability of the emergency rescue team for hazardous chemical accidents in order to better understand the current situation of these rescue teams. This understanding will support the improvement the capability of these teams in a targeted way. In addition, competent authorities could use such assessments to improve their management level. The above research results showed that it is feasible to use the PCA-BP neural network-based evaluation model to evaluate the capability of emergency rescue teams for hazardous chemical accidents, which provides a new idea for emergency rescue capability assessment.

Data Availability

All the data used to support this study have been included within the article.

Conflicts of Interest

The authors declare that there are no conflicts of interest regarding the publication of this paper.

Acknowledgments

This research was financially supported by the Sixth Batch of Experimental Major for Applied Undergraduate Programs in Shanghai, a Key Project of Shanghai Work Safety Administration and Project (2019SYSZD02) of China Association of Higher Education.