Dispatchers are responsible for emergency decision-making in urban rail transit emergencies, and ensuring the safe operation of trains. When an emergency occurs, the dispatcher is required to obtain on-site information in a timely manner, and make the corresponding handling and operational adjustments based on the understanding of the rules and on-site information. Although rail transit enterprises in every city have a relatively complete emergency response planning system, the handling requirements and program content in the emergency response plan are relatively scattered and involve multiple post operations. Secondly, during the process of handling, dispatchers are often under a great amount of decision-making pressure. Wrong decision-making and handling could exacerbate the emergency and even lead to the occurrence of secondary disasters. And although dispatchers are quite familiar with the plan in an emergency environment, cognitive call difficulties and operational errors may occur. Owing to high-stress situations of major events that require rapid decision-making, the reliance on memory may lead to the omission of critical handling steps [1]. Therefore, it is necessary to study the decision-making support for rail transit emergencies. In addition, with the digital transformation of the rail transportation industry, the study on decision-making support for emergencies can provide a preliminary research basis for developing an intelligent adaptive dispatching platform and a practical platform for dispatcher assessment and training. Furthermore, such research can also provide a new implementation method for enhancing dispatchers' emergency disposal capabilities.

Emergency decision-making has been a very active research field in recent years. Some studies have improved the efficiency of decision-makers in dealing with emergencies by providing effective decision-making information. For instance, Liu [2] proposed and introduced in detail the spatiotemporal scenario model (STSM) based on the concept of “scenario.” This model was able to dynamically describe the disaster state and help decision-makers in obtaining a more comprehensive representation of the disaster evolution in certain time and space. In another work, Van de Walle [3] observed through experiments that providing a wealth of information and centrally sharing the information with the team through a coordinator could improve the situational awareness of teams in rapid response to crises. Botega [4] improved the situational awareness (SAW) of decision-makers when analyzing emergency reports through information quality assessment and enriching the situational knowledge with reliable metadata. Furthermore, Li [5] used an interactive card game to assess the information needs of first responders and test their situational awareness in the case of building fire emergencies. Barbarosoǧlu [6] proposed a scenario-based stochastic programming (SP) model to represent a multi-commodity, multimodal network flow problem in a disaster.

Other scholars have built emergency decision-making models by studying the evolutionary mechanisms of emergency scenarios. Wu [7] proposed a novel Bayesian network-based emergency decision-making model for the consequent reduction in individual ship–ship collisions in the Yangtze River. Wang [8] analyzed the scenario evolution mechanism of environmental emergencies and constructed an emergency decision-making model for environmental emergencies using case-based reasoning. Li [9] explained the state characteristics of unconventional emergency systems and analyzed the time frame of unconventional emergency decision-making. Using emergency knowledge elements, Han [10] constructed an assertive reasoning selection methodology by improving the acquisition of the balance coefficient in the C4.5 algorithm. In addition, Wu [11] proposed a quantitative decision-making model for the early emergency response to a ship oil spill, where they established a hierarchical decision-making framework after identifying and quantifying the influencing factors and schemes reported in previous works.

Additionally, some scholars have assisted decision-makers in making corresponding decisions by improving existing models. For example, Liu [12] proposed a fault tree analysis (FTA) method to solve the risk decision problem in the emergency response. Wang [13] proposed a prospect theory-based interval dynamic reference point method for emergency decision-making (EDM) while considering the psychological behavior of decision-makers. Kafoutis [14] attempted to identify emergency plan loopholes and provide a numerical value that indicated the “distance” between what is planned and what should have been planned. Moreover, Chen [15] used hybrid Petri nets to model emergency scenarios and responses, comprehensively and systematically. Ba [16] proposed a multi-disaster scenario method based on experimental–simulation–field data, to study the coupling effects of large-scale, high-intensity, and multi-hazard effects.

Case similarity matching has been widely used in prediction technology in the field of emergency decision-making [17]. Leikola [18] constructed a material description of ore based on case-based reasoning, which was able to retrieve similar materials from a case database. Song [19] used information entropy and cosine similarity algorithms to establish a similarity matching model for formulating an emergency plan for food safety. Huang [20] proposed a similarity-based method for the early detection of emergencies from social media, which clustered social media texts based on the content, time, and place of events. Li [21] proposed a new similar judicial case matching model based on hash learning to obtain the weight of judicial features, and realized a rapid similar matching method using binary codes. Chen [22] considered the similarity measure between heterogeneous multi-attribute cases from the perspective of system space, and designed a grey wolf optimization (GWO)-based relative entropy method considering the dual information correlation for a more reasonable weight allocation. Chazara [23] used abstraction, conceptualization, and reasoning mechanisms to select source cases and set up structures based on connectivity models to flexibly represent the cases. Based on the definition of a hierarchical case tree, Wu [24] established a similar evaluation model while comprehensively considering node structure, concept, weight, value, and other information. Xiong [25] proposed a similarity model based on linguistic fuzzy rules as a knowledge container which could flexibly express the knowledge and criteria of similarity evaluation. Fan [26] proposed a new hybrid similarity measurement method and established a formula to measure the attribute similarity according to the attribute values of each format.

In summary, the principle of case similarity matching holds that the impact of disasters can be evaluated based on experience with similar types of emergencies, density, and other characteristics [17]. Emergency scenarios go beyond the experience. However, emergency decision-making depends on the dispatchers' application of experience. Case-based reasoning is an effective way to learn the historical classics, and it helps to clarify the evolution of emergency response scenarios. Case matching needs to structure the case, extract the required information, and perform the matching [2]. However, there exist only a few studies on decision-making in the field of rail transit. The main problems of applying the case matching method in emergency decision-making of rail transit dispatchers are listed as follows:

(a) The current record of emergencies mainly relies on the simple written process records. The lack of a structured description easily leads to redundant information, and it is impossible to obtain key information of emergencies directly and effectively.

(b) In most reported studies, the application of case matching is relatively simple; that is, to obtain the element information of the initial stage of emergency, once the overall similarity of the emergency is matched, the case information with higher similarity calculation result is integrated to form a handling program. However, the development of emergencies is inherently uncertain and has multiple development stages. The initial matching result may not indicate the accurate direction of the emergencies' development, and hence, the validity of the information cannot be guaranteed.


Methodological Framework

To solve the problems described in the Introduction, this article conducts research based on the framework provided in Fig. 1.

Fig. 1
figure 1

Methodological framework

Establishing a Scenario Element System

Concept of Scenario

Scenario corresponds to an ordered representation of a set of disaster statuses that emerge on the objects in a certain time-space. Its connotation is to represent the disaster status of each object, consisting of object representation and representation of disaster damage [2]. The basis of a scenario-based decision pattern is the scenario definition and representation [27], and essentially, scenario is a way to help the decision-makers effectively acquire information on disaster status. Analyzing the decision process based on scenario can trigger the specific decisions and raise the situation awareness in decision-makers, since the scenario makes the decision-makers focus on more specific decision problems and helps in projecting the possible future of the current situation [28, 29]. The occurrence, development, and evolution of the emergencies are essentially the result of interrelation and interaction of various scenario elements and objects. Therefore, the determination of scenario elements is the primary task in the study of emergency representation [8].

Disaster Theory

The definition of disaster can be summarized as a phenomenon or process in which the structure and function of the society and ecosystem are destroyed, and human life, property, or the ecological environment are also destroyed due to the imbalance of society and ecosystem or the interference of external factors [30]. The urban rail transit system can be regarded as a complete ecosystem, and its emergency accords with the definition of a disaster. Therefore, it is critical to consider the disaster-inducing factors, disaster-bearing factors, and disaster-pregnant environment in disaster science; i.e., disaster-bearing factors can be transformed into new disaster-inducing factors in the evolution of emergencies and can affect the response of the external environment to events. That is to say, the disaster-pregnant environment will affect the effective implementation of emergency activities [8]. Thus, the extraction and integration of scenario elements from the above three parts (disaster-inducing factors, disaster-bearing factors, disaster-pregnant environment) can systematically depict the emergencies and their evolution, and provide the basis for the extraction of scenario elements of emergencies.

Scenario Element System of Rail Transit Emergencies

This paper relies on the disaster theory to identify the elements of emergency scenario of rail transit. Scenario element objects and their changing statuses are analyzed and extracted from a single historical case, and the scenario elements are classified. While analyzing and processing many cases, the same scenario elements are merged and the value set of the elements is established, thereby constructing a relatively complete emergency scenario element system. The combination results are shown in Fig. 2.

Fig. 2
figure 2

Scenario element system of rail transit emergencies

Case Representation Model of Rail Transit Emergencies

Emergencies Scenario Split

According to the above analysis, this work characterizes the emergency cases based on the scenarios. The cases show the emergence, development, climax, and ending process of the emergencies. These cases are always composed of a single or a group of scenarios, and they express the development of emergencies by establishing the connection between scenarios. Notably, the link between the scenarios is the key decision point information, and the key decision point information advances the event to the next stage. Moreover, different types of emergencies have different key decision point information. That is, the scenario split of a complete case is based on the changes in the value of scenario elements. Because of the change in the value of scenario elements, the dispatcher cannot continue to adopt the predetermined disposal measures, and thus, it is necessary to make new decisions based on the information of key decision points. In other words, the state of the scenario elements remains unchanged in a single scenario, while the state of some scenario elements changes in different scenarios.

Establishing an Emergency Case Representation Model

The purpose of emergency case representation is to show the key information of the case more intuitively and effectively, and to express the case uniformly so that it can be described in a unified manner while highlighting its characteristics, Moreover, it provides basic data for the establishment of the case base, and assists in case handling analysis and further research.

Before establishing the emergency case representation model, it is assumed that the “decision point information” obtained by the dispatcher mentioned in the model is correct. The basis of this assumption is that in the process of emergency response, there is a “confirmation” system for the information transfer from each post, such as recitation. In addition, the dispatcher has experience in response and can check the information passed by the field personnel such as the status of faulty equipment through the ATS panel. Most of this information is an accurate description, and there will be no fuzzy description.

The emergency case representation model is constructed as shown in Fig. 3. Here, the scenarios are connected by the key decision point information Di. In the constructed model, scenarios are extracted from the cases, and the cases are composed of multiple scenarios, where different cases can contain the same scenario. Essentially, the value of scenario element reflects the change in the state of emergency. Similarly, the scenario elements are extracted, and the scenario is composed of the multiple elements and the processing measures taken by the decision-makers in current scenario, as shown in Eqs. 1, 2, and 3.

$$S = \{ S_{1} ,S_{2} \ldots S_{i} \}$$
$$S_{i} = \{ F,A\}$$
$$F = \{ F_{1} ,F_{2} \ldots F_{m} \} = \{ X,Y,T\}$$

Among them, the variable F represents the scenario elements related to the occurrence and development of emergencies, X represents the set of hazard factors, and Y represents the set of factors related to hazards. Meanwhile, T represents a collection of elements related to the disaster-pregnant environment, and A represents a collection of detailed disposal measures taken by the decision-makers. In summary, a complete emergency S is composed of several scenarios Si. Each scenario contains the relevant scenario elements F that lead to the occurrence and development of emergency and the corresponding taken measures A.

Fig. 3
figure 3

a Emergency case representation model, b Scenario composition

Screening of Emergency Cases

Criteria for the case screening:

  1. (a)

    The impact on the normal operation of the train is firstly considered.

    The handling of some emergencies is relatively simple and has a little impact on the normal operation of the train, and providing the decision-making support information for such cases is of small significance.

  2. (b)

    No violations in the disposal process.

    The purpose of constructing the case base is to provide the dispatchers with decision-making information during different emergencies. It is necessary to ensure that the dispatchers' handling of the selected cases is accurate and compliant, to ensure the effectiveness of the provided information.

  3. (c)

    Details of emergency information and historical records required for supporting decision-making.

    The cases in the case base must provide sufficient scenario element information and the dispatcher's handling process, to achieve scenario splitting and scenario element value extraction.

Based on the above requirements, this work collects 161 emergency cases in rail transit in a city over the past 8 years. After screening, a total of 65 cases that meet the requirements are obtained, which initially constituted a case base for the retrieval.

Decision-Making Support Model Based on Similarity Matching

Determination of the Weight of Scenario Elements

Each scenario element has a different degree of impact on the emergency, and the dispatcher needs to focus on the state of the scenario element that has a greater impact on the emergency during the process of handling. To realize a more reasonable emergency scenario element system, the weight of each key element must be determined. Currently, there exist various types of weight determination methods. Among them, the subjective method of weighting has greater volatility and poor objectivity [31]. Accordingly, the weights of scenario elements should not be determined by subjective methods such as expert scoring. Alternatively, the information weight method corresponds to an objective weighting method, and it determines the weight based on the discrimination information contained in the index [32].

Assuming that the constructed case library contains \(m\) key scenario elements and \(N\) cases, then \(\overline{x}_{j}\) is the average value of indicator \(x_{j}\), and \(s_{j}\) represents the standard deviation of indicator \(x_{j}\). After normalizing the coefficient of variation \(C_{v}^{j}\) , each scenario can be obtained. Moreover, the weight value of the element is \(\omega_{{\text{j}}}\), as shown in Eqs. 4, 5, and 6 [33].

$$\overline{x}_{j} { = }\sum\limits_{i = 1}^{N} {x_{ij} /N}$$
$$C_{v}^{j} = \frac{{s_{j} }}{{\overline{x}_{j} }}$$
$$\omega_{j} = \frac{{C_{v}^{j} }}{{\sum\nolimits_{j = 1}^{m} {C_{v}^{j} } }}$$

Similarity Matching

Because of the high complexity of emergencies and huge differences between individuals, it is difficult to match valid cases with sufficient similarity through the overall retrieval of cases. Based on the above emergency representation model, we can infer that the decision-making is essentially based on scenarios rather than cases. In addition, the complexity of the scenario is much lower than that of the case, thereby making the scenario a better retrieval unit to meet the requirements of rapid and accurate decision-making in the case of environmental emergencies.

For this reason, in contrast to the traditional case similarity matching algorithms, this paper establishes a decision-making support model based on the constructed case representation model. The main steps involved are given as follows:

Step 1: Establish a scenario element system to obtain the scenario elements describing the emergencies.

Step 2: According to the case representation model, structure the expressions of cases that meet the requirements and split the scenarios to form a search case base.

Step 3: Based on the case information in the case base, use the information weight method to determine the weight value of each scenario element.

Step 4: Use the current scenario of emergency as the retrieval unit to perform the similarity matching with each scenario in the case base and select the top three scenarios for handling measures that will be used as the decision-making support information for the current emergencies.

Step 5: Determine whether the decision-making support information is applicable, and if not, make the appropriate corrections.

Step 6: Repeat steps 4 and 5 until the normal operations are restored.

To perform the calculation of similarity matching between the emergency scenario and the historical case scenario in the case library, it is assumed that the constructed case library contains m number of key scenario elements and N cases, where each case can be divided into \(i\) sub-scenarios. Then, the scene element vector of scenario SN-1 in the historical case is represented as uN-I= {x1, x2xm}. Meanwhile, the scenario element vector of target case Oi is represented as vi = {y1, y2ym}, and the weight of \(j\) scenario element is ωj. Then, the similarity between the target case Oi and the historical case SN-1refers to the normal operation.

$$\begin{aligned} {\text{sim}}(S_{N - i} ,O_{i} ) & = \sum\limits_{j = 1}^{m} {x_{j} w_{j} \cdot y_{j} w_{j} } \\ {{}}^{{}} = \frac{{\sum\nolimits_{j = 1}^{m} {x_{j} w_{j} \cdot y_{j} w_{j} } }}{{\sqrt {\sum\nolimits_{j = 1}^{m} {(x_{j}^{2} \omega_{j}^{2} )} } \cdot \sqrt {\sum\nolimits_{j = 1}^{m} {(y_{j}^{2} \omega_{j}^{2} )} } }} \\ \end{aligned}$$

The specific similarity matching process is illustrated in Fig. 4.

Fig. 4
figure 4

Similarity matching process

Case Analysis

Determination of Case Base and Weight of Scenario Elements

Based on the scenario element system and the emergency case representation model, the 65 cases considered are divided into scenarios, and fixed patterns of information are extracted to obtain the scenario element values of historical cases in the case base and thus form the scenario element vector of each case. The case base information is provided in Table 1.

Table 1 Case base

Equations 4, 5, and 6 are used to calculate the weight value of each scenario element, and the corresponding results are shown in Table 2. Here, the coefficient of variation of each element is \(C_{v}^{j}\)≥0.15 [33], and hence, these scenario elements are very important for case matching and cannot be eliminated.

Table 2 The weight of the scenario elements

Case Calculation

This paper designs a hypothetical case that does not exist in the case base, as the target emergency. The basic information is provided as follows:

During the morning rush hour, due to a rainstorm, the wheel and rail of the xxx train slipped, the mode was lost, and the train stopped and failed to troubleshoot. After the rescue, it was found that not all the doors of the faulty train could be opened when clearing the passengers at the platform. There was a free storage line in front of the train operation direction, and at this stage, the trains were all operating on the mainline, and there was no train in the yard.

According to the case representation model, the target emergency is divided into two scenarios O = {O1, O2}, transforming the information into scenario element vectors. v1= (1,1,3,3,1,1,1,1,1,7,2,2,1,2,3), v2= (1,3,1,1,1,1,1,1,1,1,2,2,2,2,3) D1= the door of the faulty train could not be opened.

MATLAB software is utilized to calculate the similarity between the scenario element vector of the target emergency and the scenario element vector in the case base. Table 3 shows the calculation results for the top three similar scenarios after and without the scenario separation. In addition, Fig. 5 shows all the results of similarity calculations.

Table 3 The calculation results for the top three similar scenarios
Fig. 5
figure 5

Distribution of similarity calculation results

Result Analysis

Similarity matching results reported in Table 3 show that for O1 scenario of the emergency, the case with the highest similarity is S26. The cause behind this case is that the mode is lost and the braking cannot be eased. This cause is actually the same as the cause of the target emergency. Here, the next step is to refer to the handling measures for the faulty train in case S26, but this case occurred during the off-peak, and its operation adjustment plan is of low reference. On the other hand, case S46, where the braking cannot be eased due to a traction failure, occurred during a peak period, and thus, its operation adjustment plan has a certain reference significance.

Additionally, for the O2 scenario of the emergency, in case S44-1, the train could not open the door to clear the passengers on the platform, and here, the handling measures can be consulted. The case S34-1 occurred in the morning rush hour, and there was a free storage line in front of the incident location, which was the same as the line resources, time, and train resources of the target emergency, and therefore, its operation adjustment plan possessed a strong reference. In case S31, the train could not move because the high-speed position switched during the morning rush hour, which eventually led to rescue. The operation adjustment plan for such case can be used for reference.

In summary, the handling plan for the target emergency is as follows. Scenario O1: After the train driver reports that they are unable to move the train, the subsequent train is ordered to rescue the faulty train, after clearing the passengers at the platform. Following the successful trial pull, the rescue train is ordered to run to the front platform for clearing the passengers by cutting off the ATP. During the train rescue connection period, the subsequent trains are detained, the number of trains in the faulty section is controlled, and some trains are properly arranged to fill the training gap in the non-faulty section in the fault direction, after the return station is cleared. Scenario O2: The train door cannot be opened, and the station is required to cooperate with each train to open a door and implement the clearing of passengers. After clearing the passengers, connect the train to the front storage line, and the faulty vehicle exits the operation mode. After unhooking, the rescue vehicle passes through the return line to the platform for normal passenger operation. During the continuous operation of the trailer, according to the passenger flow situation, reasonable measures are taken to limit the current at the station.

To verify the effectiveness of the proposed model, the entire emergency event is taken as the retrieval object, the similarity calculation is carried out with the traditional method, and the results obtained by the two methods are analyzed. It can be seen from Table 3 that the result of similarity calculation while using the entire emergency as the retrieval object returns the highest similarity case with the scenario as the retrieval object, which essentially verifies the feasibility of using scenario as the similarity retrieval object. By analyzing Fig. 5, it can be seen that the distribution and trend of the three broken lines are roughly the same. Meanwhile, the similarity curves of O1 and O almost overlap, indicating that the case matching results obtained by the traditional method are suitable for the initial stage of emergency, and the method established in this paper can also obtain the results obtained by the traditional method. The similarity calculation result of O2 is generally greater than O, reflecting that the retrieved information is more targeted after the scenario is split, which relieves the shortcomings of traditional methods to a certain extent and improves the efficiency of information use.

The results of this research show the following:

  1. (a)

    When the similarity calculation results are the same, it is impossible to effectively distinguish which case is more in line with the current emergency. Follow-up studies are needed to evaluate the effectiveness of case handling, such as delays in time and order clearance. That is, when the case similarity calculation results are the same, the case handling with the better evaluation result is preferred as the decision-making support information.

  2. (b)

    There are many factors affecting the rail transit emergencies, and the constructed scenario element system cannot fully cover all the elements. When there are too many elements in the scenario, the values of individual elements cannot be effectively reflected in the similarity calculation results, which is not conducive to the case retrieval.

  3. (c)

    At present, there is no clear regulation on the number of cases that provide the decision-making support information after the similarity calculation, and too much information will surely reduce the dispatcher's handling efficiency. Based on the actual decision-making process, this article subjectively selects the top three similar cases as the decision-making support information, which has some inherent limitations.


Based on the disaster theory, this paper establishes a rail transit emergency scenario element system, constructs a rail transit emergency case representation model, and proposes to use the scenario as the retrieval object to match the target emergency case and assist the decision-makers. Decision-making information is used to verify the feasibility of the proposed model using traditional similarity testing methods.

Compared with the traditional similarity calculation, the decision-making support model established in this paper offers the advantage of using the scenario as the retrieval unit to search for the current scenario in a timely manner according to the development of the event and the change in the value of scenario element, and provide more targeted decision-making information for the handling of complex emergencies, until the handling of the incident is completed. Even in different types of cases, the same handling measures can be used as the information for decision-making under similar scenarios, thereby increasing the information utilization rate. Secondly, there is no clear regulation at the moment on how much case information should be selected as the decision-making support information. If the top-ranked case information is selected as the decision-making support information, this article can obtain results close to those of traditional methods. However, if the case with the greatest similarity is selected, the method constructed in this article obtains a combination of handling plans in different scenarios under emergencies, and provides more comprehensive and targeted decision-making support information.