Method for Assigning Railway Traffic Managers to Tasks along with Models for Evaluating and Classifying

Abstract

The occurrence of incidences in railway systems leads to impediments and often delays. Because the railway is an anthropotechnical system, two factors are considered as the source of incidents: technical and human. Minimizing adverse incidents in the railway system is the subject of much discussion and research. One of the areas affecting the performance of railway systems is employees. This article presents a method for assigning railway employees to tasks and models for evaluating and classifying railway employees, consisting of two stages. The first stage involves using a survey method and a fuzzy logic model. Each type of service is assigned feature values, obtaining three parameterized employee-role profiles for the train traffic officer. In the second stage, the participant goes through two of the three available evaluation scenarios, during which errors made during the tasks are counted. Validation results of the proposed approach indicate that the method is 87% effective.

1. Introduction

The most common undesirable incidents on the railway system lead to delays and subsequent disruptions. Delays or disruptions usually prevent the planned schedule from being carried out due to discrepancies in time and space. As a result, trains may arrive late and travel on different tracks, requiring changes in the train crew or the train itself. These changes in operations disrupt the planned schedule. Ultimately, this results in the loss of planned, collision-free train traffic and negatively affects the resilience of the railway system.

One area connected to train disruption is a group of factors related to human error. Traffic disruptions can lead to behavior consisting of many minor failures, which can eventually lead to an accident. Overspeeding and maintenance errors, as forms of human error, can lead to train delays. Delays can also be caused by incorrect decisions made by signal box operators. In addition, other effects of traffic disruptions on personnel can be distinguished, such as the impact on personnel of decisions made under stress and lack of time.

In the case of controlling train traffic, there is a wide variation in the type of work to be undertaken. It is possible to have situations with heavy traffic in a small area or a need to control traffic in a larger area that is remote. The variety of tasks in the position of the traffic controller can be contrasted with the diverse predispositions of employees. Some people get tired quickly when they experience too much stimulus. Others get tired quickly when there is too little going on and they can then become distracted. There are also people who perform their tasks accurately and without errors, but only when they do not have to communicate with another person.

The authors have already classified the workforce in previous studies and have demonstrated the links between these categories and the safety of the rail system [1]. Given the variation in people’s predispositions, it is possible to hypothesize that there is a best-fit group of people for a given type of task.

Accordingly, the following research questions can be formulated to be answered in this paper:

• Is it possible to match the developed types of employees with specific types of tasks?
• Is it possible to build a model to categorize an employee based on a simple questionnaire?
• Is it possible to verify the method experimentally?

The article is divided into five sections. Section 2 presents a literature review that defines the research gap. Section 3 introduces a method based on fuzzy logic, which aims to assign an employee status category based on the answers given in a simplified questionnaire. This section concludes with a description of an experiment involving eye-tracking. Section 4 presents the results of the implemented research and a discussion on them. The article ends with conclusions, which consider account the research’s critical aspects and indicate directions for further research.

2. Literature Review

In 2022, 1615 rail accidents were reported in the European Union [2]. Railway traffic safety depends on technical equipment and administrative arrangements [3]. According to a report compiled by the International Union of Railways (UIC), the human factor is responsible for 10.2% of railway accidents [4]. Among these causes, the following were identified: train drivers (3.0%), traffic operating and signaling staff (2.3%), track and turnout maintenance staff (1.5%), and other or not specified (3.5%). At the same time, people are exposed to the influence of other factors that, being in their environment, can affect their actions, decisions, and behavior. Researchers in numerous studies point to various causes of errors on the part of human mistakes. Safety is also ensured by personnel supported by signaling, safety, and communications equipment. Employees must follow local railway operation laws, adhere to international standards and protocols, and refer to technical and operational documents.

One of the pillars of railway system operation is effective human traffic control [5], so railways are recognized as systems with a high degree of human influence. Thus, the human factor is, in most cases, the cause of undesirable events. Railway operators (of the train or traffic systems [6]), in most cases work correctly until conditions change. Such changes in the situation are threats resulting from undesirable events related to machines, materials, environment, and management [7,8,9]. Then, together with human error, failures can lead to accidents [10]. Human errors can occur at the last stage of an accident as well as during the first stages of an accident chain such as design errors or errors in maintenance [11]. When considering the human factor layer in the root causes of an accident, a lack of experience of the personnel is frequently pointed to as the main cause [12].

In addition to changing conditions causing threats, there are many factors that influence the probability of human failures. In ref. [13], the following factors were identified: mental workload, task time, punctuality, stress vulnerability, the quality and quantity of sleep, interface ergonomics, and automation level. The mental workload assessment of railroad pro-access workers has already been analyzed [14]. In ref. [15], the signaler workload exploration and assessment tool (SWEAT) was used for this purpose. The explicit and situational awareness of train traffic controllers was also studied [16]. When considering these factors, employees can negatively impact railway safety.

The works of refs. [17,18] identified the causes of hazardous incidents in relation to the state of employees in terms of physical and mental issues, and personal abilities as well as gathered skills. Based on the modified human factor analysis and classification system (HFACS), presented in ref. [19], as well as an analysis of unsafe situations and human-caused situations [7,20], the following groups of accidents/incidents influencing factors can be identified: operator acts, preconditions for operator acts, supervisory factors, organizational factors, outside factors, recognizing information, assessing circumstances, decision-making processes, execution processes, miscommunication, violation of procedures/regulations, alcohol, illness, congestion, and disabilities. Sub-causes of accidents resulting from traffic controller’s failures, which have been especially analyzed in ref. [6], are given as follows: failure of equipment, communication device failure, track geometry failure, and those caused by signal misunderstanding, the physical condition of the personnel, and general rules and procedures.

In a study on rail transport of dangerous goods [21], the classification of human error included eight factors: warehouse keepers’ operation error, freight forwarder’s operation error, human-induced factors, negligence of the escort, concealment, false alarm, low-skill level, examiner’s operation error, and incorrect handling. Of the three risks identified in ref. [22], the human factor included but was not limited to non-compliance with regulations, improper adherence to operational/technical recommendations, improper shunting of traffic, poor route preparation, and exceeding the maximum speed limit. Poor self-discipline, inadequate supervision, inadequate rest, and inappropriate attitudes to work are a group of factors identified in the ref. [23].

Previous work has mainly focused on the work of train drivers. The research carried out included train driver fatigue [24] and ergonomic evaluation of the “drivers-tram-traffic environment” system [25]. Much work has been undertaken on train drivers [24,25], as well as on characterizing driving behavior using decision trees [26]. On the other hand, there is a lack of work dealing with the connection of skills and abilities of the traffic control personnel with safety, depending on operational issues. Some works have identified the influence of traffic control system type on traffic safety [27,28]. In addition, the relationship between sleep deprivation and alertness of railway signalers has already been verified [29], as well as the effect of time of day on the risk of injury to a railway worker [30].

Research related to the human factor in the railway system covers many areas. The need to assess human performance in the railway system has already been highlighted in ref. [31]. Considerations that have been carried out also relate to railway workers’ alertness [28], stress [32], and job satisfaction [33]. The original contribution of the work is related to developing a method for assigning railway workers to a position. The literature identifies work on theories and concepts relating to matching employees to positions [34,35,36]. In terms of matching an employee to a position, the railway system also requires consideration of safety issues, especially in the area responsible for traffic control. In the literature, one can find works that evaluate the human factor in the railway system. For example, in ref. [37], a model for evaluating human factors safety in a railway system based on information extraction was proposed. The model considered the human factor safety evaluation system and addressed employee monitoring, human error control, human factors management systems, and human–machine interfaces. Questionnaires relating to the identification of human errors can be helpful. In ref. [38], a tool for assessing the probability of human error based on a questionnaire containing 67 questions was developed to determine the monthly errors made by railway traffic control room employees. Work also deals with the use of fuzzy sets. The developed model for dynamic evaluation of a person for a position is based on fuzzy rating matrices, focusing on job satisfaction [39]. In addition, ref. [1] presents the factors affecting safety in traffic control. Three factors were introduced that directly impact the number of incidents and accidents. The worker’s perception was analyzed: primary duties, secondary duties, and the essential difficulties at work.

3. Material and Methods

3.1. Survey and Interviews of Train Traffic Controllers—Previous Research

In prior studies conducted by our research team [1], an extensive survey was administered to 79 train traffic controllers. The survey collected detailed information regarding their job responsibilities, routine activities, and the challenges they face in their work environment.

Respondents were asked specifically about their main and additional responsibilities, including tasks they performed daily, periodically, and occasionally. The focus was on activities such as controlling train operations, supervising traffic safety, inspecting rail infrastructure, managing rail crossings, and conducting training sessions. Moreover, we requested our experts to provide detailed descriptions of fundamental activities, which were broken down into their individual components.

In the subsequent part of the survey, respondents were asked about their feelings and thoughts regarding the work they perform. This section aimed to uncover the personal perceptions of the respondents concerning their duties. They were requested to define what constitutes success in their job, identify the most essential and challenging aspects of their work, and describe the necessary qualifications and education required. Furthermore, respondents detailed the training needed to perform their job effectively, tasks performed that are not officially listed in the job description, and how downtime is managed. They also discussed the largest problem or challenge faced in the past year, provided a description of their supervisor’s role, and indicated whether they have access to confidential information. Additionally, they described any interaction with high-value items or money, the nature and extent of their interaction with colleagues, the support necessary to conduct their activities, the hierarchical structure, and the presence of subordinate employees and menial tasks.

In the final part of the survey, which aimed to gain a deeper understanding of the infrastructure, managers and respondents were asked about their current work history. Specifically, they were requested to provide their age, length of service, and the number of incidents and adverse events, including rail accidents and other safety violations, in which they had been involved due to their profession and service. Additionally, respondents were asked to assess their overall job satisfaction in relation to their employer’s requirements and expectations. To ensure the integrity of the responses, the survey was fully voluntary and anonymous, and data analyses were generalized to the greatest extent possible.

Based on an analysis of the responses, three areas were identified as being associated with the occurrence of accidents or incidents. The responses from train traffic controllers regarding their essential job duties enabled the classification of three primary categories. The respondents indicated that their essential job duties included the following:

• Working as a team;
• Ensuring the safety of traffic and travelers;
• Completing formalities, understood as completing documentation or following strict instructions.

The results obtained from the questionnaires enabled the research team to identify three distinct employee profiles within each group of questions, encompassing most responses. By collating these responses and plotting them on axes representing the number of incidents and accidents (related to their supervisory and operational duties) and seniority, we derived linear regression equations for each identified staff profile. The linear regression equation takes the form $\mathrm{F}\left(\mathrm{x}\right)=\mathrm{A}\mathrm{x}+\mathrm{B}$. Given that the data indicate zero incidents or accidents at the start of an employee’s career, the value of coefficient $\mathrm{B}$ is 0. Consequently, coefficient $\mathrm{A}$ represents the failure intensity, expressed as the number of accidents and incidents per working year.

For the group of questions concerning perceptions of the scope of main job responsibilities, we identified three types of employees as follows:

• Those who believe that teamwork is most important—A₁;
• Those who prioritize completing formal matters—A₂;
• Those who focus on supervising and ensuring safety—A₃.

The directional coefficients of the linear regression line were as follows: ${\mathrm{A}}_1=0.3188$, ${\mathrm{A}}_2=0.3533$, and ${\mathrm{A}}_3=0.2118$. The lowest accident rate was observed among employees who consider ensuring safety as their primary duty.

For periodic and additional duties, the responses obtained from traffic officers were categorized into three groups. The respondents indicated that their periodic duties included the following:

• Maintaining order in the workplace, such as cleaning the stove and smoking in it, washing windows, and cleaning a railway cradle at the railway crossing—A₁;
• Improving their qualifications by attending periodic training courses and reading railway manuals—A₂;
• Completing paperwork—A₃.

The directional coefficients of the linear regression line for these groups were as follows: ${\mathrm{A}}_1=0.3722$, ${\mathrm{A}}_2=0.2212$, and ${\mathrm{A}}_3=0.2691$.

The third area in which variation was observed in the gradient of the cumulative number of incidents was the factors causing difficulties on the job. It was possible to identify three groups of the most frequent responses as follows:

• High levels of stress—A₁;
• Problems resulting from the need to cooperate and communicate with other employees—A₂;
• The volume of duties resulting from traffic disruptions and the mere occurrence of deviations from the planned schedule—A₃.

The respondents’ responses provided the following values for the linear coefficient of the regression simplexes (in sequence): ${\mathrm{A}}_1=0.9537$, ${\mathrm{A}}_2=0.5152$, and ${\mathrm{A}}_3=0.2956$.

Based on this knowledge, in this article, we aim to extend the scientific literature by applying the obtained profiles through their implementation into a fuzzy logic model and, thus, extend the inference of workers involved in running safe rail traffic.

3.2. Job Analysis of Train Traffic Controllers—Current Research

Based on the above regression models [1], hypotheses were made that the indicated three characteristics describe diverse types of employees, so they can be used to parameterize employee profiles. To verify the hypotheses, a non-parametric Kolmogorov test was conducted at a significance level of 0.1 for pairs of distributions for values of the traits. As a result, it was shown that at the given significance level, the null hypotheses of conformity of the runs of feature values were rejected, and the alternative incompatibility hypotheses were accepted. This led to the conclusion that the accepted feature values describe different types of traffic officers.

Next, the position of the traffic officer on duty was analyzed based on the collected materials. As a result, three different types of service were identified that corresponded to employee characteristics as follows: in a local steering center, at a low-load classic station, and at a loaded station.

Working in the local steering center (LSC) involves monitoring and controlling the operation of the automatic system and guiding traffic in case of faults. The workstation is equipped with LCD monitors. The LSC serves as a control station for the remote control of railway traffic on a section of a railway line or station. It allows for the centralization of train traffic control at multiple stations. The train traffic controller on duty intervenes in equipment operation only in emergencies, e.g., speed limits and failures of crossing horns. In a normal operation mode, everything is performed automatically. The workstation is equipped with several monitors with the following systems:

• System of rolling stock emergency state detection;
• Control of railway turnout heating and lighting;
• Radio control;
• Video monitoring of railway–road crossings;
• Remote railway crossing control equipment;
• Alarm control at railway stations;
• The condition of railway stations under the supervision of the LSC;
• A system to assist the train traffic controller on duty, which shows, among other things, the timetable;
• Computerized telephone switchboard.

The train traffic controller on duty can, among other things, determine the position of a vehicle equipped with a European rail traffic management system (ERTMS/European train control system (ETCS) onboard equipment, deregister the vehicle, activate or revoke speed restrictions, issue an emergency stop order for a train, send or receive messages from train drivers, verify train numbers, and check information on train paths.

Traffic control posts can be divided into those that manage heavy traffic and those through which fewer trains pass. Working at a high-traffic train station is associated with increased repetitive activities that the on-duty train traffic controller must perform. High traffic volume is associated with trail capacity utilization of more than 70%, at least 3 tracks, and more than 100 trains (line or shunting) daily. The train traffic controller working in the signal box determines the appropriate course of the train by the arrangements recorded in the timetable. A control panel is used for this, on which the track layout of the entire area under the control panel is drawn schematically. The tracks light up in the appropriate color depending on their occupation status. Information on track occupancy status is sent to the signal boxes from trackside devices, detecting the presence of trains on individual track sections. Information about an approaching train to the station is signaled via telephone from a train traffic control post. Depending on whether it is a signaling box serving heavy traffic, the position of an assistant train traffic controller is next to the control desk. He or she assists in maintaining the station’s traffic records, operates communications equipment with neighboring traffic stations, runs the radio communications equipment with train drivers and other communications equipment with other stations of his station, and notifies the crossing keeper of the departure of trains. When the assistant train traffic controller performs the described duties, the disposing train traffic controller conducts train traffic by operating the desktop. From the desktop, the turnout and semaphores settings are examined. The traffic controller changes the position of the switch on the desktop and excites the operation of relays, which activates the control devices of points and signaling. Train traffic controllers (assistant and disposing) communicate with each other regarding the information transmitted by telephone from the previous posts and the actions to be performed by the dispatching train traffic controllers regarding the train’s passage. The position of an assistant duty train controller is not required at a low-traffic signal box, as this is managed by one employee in his spare time between train runs.

Finally, feature values were assigned to each service type, obtaining three parameterized employee–station profiles for the on-duty train traffic controller, as shown in

Table 1. Employee–station profiles for the train traffic controller on duty.

No.	Employee-Station Profiles	Feature 1 Perceived Primary Duties Outside of Running Rail Traffic	Feature 2 Perceived Periodic Responsibilities	Feature 3 Perceived Difficulties in the Position
L	local steering center (T1)	Completing the formalities (C1.1)	Maintaining order at the workplace (C2.1) Completing the formalities (C2.2)	Problems with cooperation (C3.1)
2	Classic post weakly loaded (T2)	Teamwork (C1.2)	Completing the formalities (C2.2)	Stress (C3.2) A flurry of duties and train traffic disruptions (C3.3)
3	Classic post loaded (T3)	Ensuring safety (C1.3)	Maintaining order at the workplace (C2.1) Upgrading qualifications (C2.3)	A flurry of duties and train traffic disruptions (C3.3)

.

3.3. Method for Assigning Railway Workers to Tasks along with Models for Evaluating and Classifying Railway Workers

This method consisted of two stages. First, a survey was conducted on either the employee’s or the adept’s perception of the job. In this stage, inference was made based on the answers and assigned to a particular group using fuzzy models.

In the second stage, the trainee underwent two scenarios (out of three evaluations) in which incorrectly performed activities were counted. If the first categorization stage’s evaluation was ambiguous, the second stage resolved the issue.

For the first stage, the Mamdani-type fuzzy modeling was used. A minimum approach was adopted for the “and” relationship and for “or” a maximum. The implication was based on minima, while aggregation was based on maxima. Defuzzification is centroidal. The model was implemented and used in the framework of the MATLAB R2024a software.

Using the knowledge from Section 3.1 and Section 3.2, a model was developed to classify the train traffic controllers based on three characteristics as follows:

• C1: perceived primary duties outside of running rail traffic;
• C2: perceived periodic responsibilities;
• C3: perceived difficulties in the position.

The general model is shown in Figure 1. Characteristics C1-C3 are the input variables, while the output is represented by the T variable. The input and output variables are described in detail in the following sections.

3.3.1. Input Variable C1—Perceived Primary Duties outside of Running Rail Traffic

For feature C1, the respondent assigned a weighting factor of total 100% for each possible answer (completing the formalities, teamwork, and ensuring safety). Each answer was analyzed during a second study of the survey responses. The aim of the study was to find possible combinations of answers for C1. That is, whether, in practice, the obtained combinations of answers were 1, a mix of answers 1 and 2, answer 2, a mix of answers 2 and 3, or answer 3, while, at the same time, meeting the condition of not having a combination of 1 and 3. Such an order was possible to find the following:

• “Completing the formalities” $\mathrm{C}1.1=1$, where the weight factor assigned by survey participants is denoted as $\mathrm{W}1.1$;
• “Teamwork” $\mathrm{C}1.2=2$, where the weight factor assigned by survey participants is denoted as $\mathrm{W}1.2$;
• “Ensuring safety” $\mathrm{C}1.3=3$, where the weight factor assigned by survey participants is denoted as $\mathrm{W}1.3$.

As a result, the value of variable C1 is calculated according to the formula:

$$\mathrm{C}1=1·\mathrm{W}1.1+2·\mathrm{W}1.2+3·\mathrm{W}1.3$$

(1)

The sum of the indicated products was entered into the fuzzy model in the form of values $[1;3]$ within the input variable C1. The membership functions for the input variable C1 are presented in Figure 2.

3.3.2. Input Variable C2—Perceived Periodic Responsibilities

The procedure for feature C2 was analogous to that of C1. Each answer was analyzed during the second study of the survey responses to find possible combinations of answers for the C2 variable. Like for C1, it was also possible to line up the answers and to confirm no combination of 1 and 3 as follows:

• “Upgrading qualifications” $\mathrm{C}2.1=1$, where the weight factor assigned by survey participants is denoted as $\mathrm{W}2.1$;
• “Maintaining order at the workplace” $Maintaining order at the workplace \mathrm{C}2.2=2$, where the weight factor assigned by survey participants is denoted as $\mathrm{W}2.2$;
• “Completing the formalities” $\mathrm{C}2.3=3$, where the weight factor assigned by survey participants is denoted as $W2.3.$

As a result, the value of variable C2 is calculated according to the formula:

$$\mathrm{C}2=1·\mathrm{W}2.1+2·\mathrm{W}2.2+3·\mathrm{W}2.3$$

(2)

The sum of the indicated products was entered into the fuzzy model as $[1;3]$ values within the C2 input variable. The membership functions for the input variable C2 are presented in Figure 3.

3.3.3. Input Variable C3—Perceived Difficulties in the Position

The procedure for feature C3 is analogous to that of C1 and C2. Each answer of the survey was analyzed during the second study to find possible combinations of answers for the C3 variable. Like for C1 and C2, it was also possible to line up the answers and to confirm no combination of 1 and 3 as follows:

• “Problems with cooperation” $\mathrm{C}3.1=1$, where the weight factor assigned by survey participants is denoted as $\mathrm{W}3.1$;
• “A flurry of duties and train traffic disruptions” $\mathrm{C}3.2=2$, where the weight factor assigned by survey participants is denoted as $\mathrm{W}3.2$;
• $Stress C3.3=3$, where the weight factor assigned by survey participants is denoted as $\mathrm{W}3.3.$

As a result, the value of variable C3 is calculated according to the formula:

$$\mathrm{C}3=1·\mathrm{W}3.1+2·\mathrm{W}3.2+3·\mathrm{W}3.3$$

(3)

The sum of the indicated products was entered into the fuzzy model as $[1;3]$ values within the C3 input variable. The membership functions for the input variable C3 are presented in Figure 4.

3.3.4. Inference Rules and Output Variable T

Inference was conducted based on the rules in

Table 2. Rules for the fuzzy model of assigning the type of employee for the train traffic controller.

Rule	C1	C2	C3	Type	Rule	C1	C2	C3	Type
1	C1.1	C2.1	C3.1	T1	15	C1.2	C2.2	C3.3	T2
2	C1.1	C2.1	C3.2	T1	16	C1.2	C2.3	C3.1	T2
3	C1.1	C2.1	C3.3	T3	17	C1.2	C2.3	C3.2	T2
4	C1.1	C2.2	C3.1	T1	18	C1.2	C2.3	C3.3	T3
5	C1.1	C2.2	C3.2	T2	19	C1.3	C2.1	C3.1	T3
6	C1.1	C2.2	C3.3	T2	20	C1.3	C2.1	C3.2	T3
7	C1.1	C2.3	C3.1	T1	21	C1.3	C2.1	C3.3	T3
8	C1.1	C2.3	C3.2	T1	22	C1.3	C2.2	C3.1	T3
9	C1.1	C2.3	C3.3	T3	23	C1.3	C2.2	C3.2	T2
10	C1.2	C2.1	C3.1	T1	24	C1.3	C2.2	C3.3	T3
11	C1.2	C2.1	C3.2	T2	25	C1.3	C2.3	C3.1	T3
12	C1.2	C2.1	C3.3	T2	26	C1.3	C2.3	C3.2	T3
13	C1.2	C2.2	C3.1	T1	27	C1.3	C2.3	C3.3	T3
14	C1.2	C2.2	C3.2	T2

. As a result, the output variable T was estimated.

Defuzzification and the final result (assignment of a train traffic controller to a category) are determined based on the output membership functions. The meanings of the values are given as follows:

• $T\in [1.00;1.50]$ corresponds to the “local steering center” type;
• $T\in (1.50;1.75)$ is a mixed type “local steering center” and “classic post weakly loaded”;
• $T\in [1.75;2.25]$ corresponds to the “classic post weakly loaded” type;
• $T\in (2.25;2.50)$ is a mixed type “classic post weakly loaded” and “classic post loaded”;
• $T\in [2.50;3.00]$ corresponds to the “classic post loaded” type.

The membership functions for the output variable are presented in Figure 5.

3.4. Method Verification

The experiments were prepared according to three scenarios, which counted the errors and shortcomings committed by traffic officers on duty. Errors were evaluated by assigning a given category as follows: 0 signifies no errors, 1—denotes uncertainty in the implementation of activities, 2—implies error not affecting traffic safety, and 3—indicates critical error. Each of the three profiles is described in detail in the following:

• Local steering center—crossing the mock-up trains with full knowledge of the position of trains in the section under consideration, i.e., in the upper and lower parts of the mock-up. No contact with other traffic stations was required. Therefore, the plates covering the lower section were removed. The spacing between trains in one direction was 5 min;
• Classic post weakly loaded crossing trains through the mock-up with limited knowledge of the position of trains, which requires cooperation with other posts. The lower part of the mock-up was covered. No contact with other traffic train posts was required, so the plates covering the lower part were removed. The interval between trains in one direction was 10 min;
• Classic post loaded—crossing trains through the mock-up with limited knowledge of the position of trains, which requires cooperation with other posts. The lower part of the mock-up was covered. No contact with other traffic train posts was required, so the plates covering the lower part were removed. The interval between trains in one direction was 5 min.

In each scenario, the goal was to correctly serve eight pairs of trains. Correct service involves establishing the correct route from the entry semaphore to the correct station track, and then from the exit semaphore to the correct track. Compliance with the timetable was also important. In addition, eye-tracking glasses were included to verify the actions’ confidence. The research team analyzed the recordings to determine whether the subject first looked for the correct switch or confidently made the choice.

3.4.1. Automating Eye-Tracking Measurement

The main goal of the eye-tracking camera image analysis method used in this study was to gather information from each frame of the video about the location of the turnout switches in the image, their switching status, and the fact that the subject directed his or her gaze to one of them. The primary tool used for this purpose was cascade classifiers, available in the OpenCV library based on machine learning. The stored information was analyzed to look for the occurrence of an assumed sequence of events. A block diagram of the program’s operation is shown in Figure 6.

The cascaded architecture classifier available in the OpenCV library is based on an algorithm developed by Paul Viola and Michael Jones, which initially allowed detection based on only one set of features—Haar wavelets. In later years, the technique was improved, and the possibility of using so-called diagonal features was added. The collections extended by this part are called Haar-type features—another refinement of cascades introduced in cooperation with local binary patterns (LBP).

The mechanism of operation of the classifier using local binary patterns is to divide the source image into rectangular regions and calculate LBP values of pixel intensity levels for them. Based on these, histograms are determined for each region and stored as a vector. These are then compared with the vectors of histograms developed in the training process and, depending on the degree of agreement, the sample is finally classified as positive or negative. The algorithm’s result is finding the object for which the cascade was trained in the image under analysis.

Because of the need to detect switches that may be in two position states, two classifiers were used, each responsible for finding a switch in each position. Along with the OpenCV library, cascade classifiers trained to detect certain features have been made available. However, they focus on detecting human silhouettes, faces, eyes, and license plates. To recognize other objects, it is necessary to train the cascades on their images. This is undertaken using the opencv_traincascade application, which is available from the OpenCV library. The process of preparing samples and training the cascade classifier available in the OpenCV library can be divided into the following steps:

• A collection of images of the object to be detected, known as positive samples, should contain the object of interest from different perspectives. Next, the coordinates of the rectangular frames bounding the item of interest on each sample must be indicated. To make this laborious and time-consuming step more accessible, we used the opencv_annotation tool, which allows objects in the image to be surrounded by rectangles using a computer mouse.
• A vector output file was created with positive samples using the createsamples tool from the OpenCV library. It takes the objects indicated in the previous step and normalizes them to the desired size. It is possible to generate new positive examples by, among other things, geometric transformations or adding noise to the selected samples. This is a helpful feature when a small amount of training material is available, although it is not as effective as providing unaltered but different images.
• Collecting or generating negative samples are images that do not contain the objects to be detected. The best results originate from data from the same source as the positive samples, such as comparable frames from the analyzed video. It is also possible to use images that do not contain the elements of interest.
• The final stage is training the cascade. It tests approximately 100,000 features in each training window for all positive and negative examples. This process can take many hours or even days, depending on the number of samples provided and the parameters set.

With the classifiers trained, one can analyze the eye-tracking camera image. In the first step, a video frame is taken, on which the detection of switches is made. Based on the coordinates of the found objects, the switches’ mutual position is determined and numbered, starting with the found object lying closest to the upper left corner of the image. Depending on which classifier detected a particular switch, theswitch’s state (lever position) is recorded. The subsequent image frame is taken if no switches are found in the image.

In the next step, the coordinates of the point at which the gaze was directed are determined. The image from the eye-tracking camera contains this information in the form of a red point. Based on the knowledge of the color and size of the point, its localization is performed. The thresholding and contour detection tools cv.findContours from the OpenCV library were used for this purpose. If the coordinates of the point are contained in any of the areas where switches were found, this is noted in memory.

This process is repeated until the end of the recorded material. In this way, an array is created containing frame-by-frame information about the state of each switch and the gaze directed to the switch. These data are then analyzed according to the assumed two scenarios as follows:

• The observed person looked at one or more turnout switches and switched the last one he looked at;
• The observed person looked at one or more turnout switches, then turned his or her gaze again to the switches he or she had already seen and only then made a switch.

In addition, each such event is determined as follows:

• Whether the number of observed switches was more significant than three;
• Whether the observed person looked alternately at the mock-up and the switches more than once.

Based on a time-lapse analysis of the image from the eye-tracking camera and tests of the subsequent performance, the number of frames of the image where the point of gaze is within the switch was selected. With this number set at three, the optimal performance of the program was obtained, that is, the elimination of situations in which the gaze was accidentally directed briefly to the switch and, simultaneously, the actual observations of the switches were not omitted.

3.4.2. Experiment Structure

The scenario resulting from the subject’s categorization was assigned first. In the case of a safe error with either uncertainty of action or unsafe error, the second scenario was inflicted according to the second closest category. Then, the one with the fewest errors was selected as the leading one. Each train path was evaluated. Shots from the eye-tracking camera footage are presented in Figure 7 and Figure 8.

For the train traffic controller, categorization was first based on a questionnaire and method. Then, the experiment was conducted according to the scenario dedicated to the employee category. In case of errors, the experiment was repeated with a second scenario. After these two experiments, the actual category was assigned according to the scenario with fewer errors.

4. Results of the Experiment and Discussion

The validation experiment consisted of 15 participants that already had experience in the job, but not more than 5 years. They were not older than 30 years. The volunteers of the validation experiment did not participate in the previous research, including the survey.

The occurrence of significant errors on the job was assigned if the train traffic controller was characterized as follows:

• Committing safe errors and uncertain performance for at least 25% of the conducted trains, where the uncertainty was identified by the eye-tracking image processing algorithm;
• Committing safe errors in at least 50% of the trains in the absence of unambiguous uncertainty in the performed activities, where the uncertainty was identified by the eye-tracking image processing algorithm;
• Making at least one unsafe error.

The test validation results are presented in

Table 3. Validation test results.

Category—Model Score Based on Survey	Description of Service by Experiment (Category Correction)	Were There any Significant Errors/Deficiencies While Performing the Job Tasks?	Matching in the First Iteration or after Any Correction
LSC	LSC	no	yes
LSC	Classic post weakly loaded	yes	no
LSC	LSC	no	yes
Classic post loaded	LSC	yes	yes
LSC	LSC	no	yes
Classic post weakly loaded	Classic post loaded	yes	no
Classic post weakly loaded	Classic post weakly loaded	no	yes
Classic post weakly loaded	Classic post weakly loaded	no	yes
Classic post weakly loaded	Classic post weakly loaded	no	yes
Classic post weakly loaded	Classic post weakly loaded	no	yes
Classic post weakly loaded	Classic post weakly loaded	no	yes
Classic post loaded	Classic post loaded	no	yes
Classic post loaded	Classic post loaded	no	yes
Classic post loaded	LSC	yes	yes
Classic post loaded	Classic post loaded	no	yes

.

Upon analyzing the data in Table 3, it should be noted that the respondents correctly performed the tasks in the experiment in 11 out of 15 cases, which indicates that the effectiveness of the proposed method is $p_1={\displaystyle \frac{m_1}{n_1}}={\displaystyle \frac{11}{15}}=73\%$. After changing the scenario of the experiment to be adequate for the second preference of the post and conducting the experiment again, a total of 13 out of 15 cases showed a correct match to the post. This indicates the effectiveness of the method at $p_2={\displaystyle \frac{m_2}{n_2}}={\displaystyle \frac{13}{15}}=87\%$.

The gathered results have been analyzed statistically to answer the following questions:

• Does the second step improve the effectiveness?
• Is the effectiveness after one or two steps at least 90%?

To answer the questions, two parametric tests were performed. The first was the parametric test for two proportions. The z-value was calculated according to:

$$z={\displaystyle \frac{p_1-p_2}{\sqrt{\frac{\frac{m_1+m_2}{n_1+n_2} · \left(1-\frac{m_1+m_2}{n_1+n_2}\right)}{\frac{n_1·n_2}{n_1+n_2}}}}}$$

(4)

where the critical region for the significance level $\alpha =0.05$ has a range of $(-\infty ;$−1.65). The calculated z-statistics is equal to −0.96.

Second, a parametric test for the proportion was performed, investigating the equality to 90% for the proportion after one step and for the proportion after two steps. For this, the formula was used.

$$z={\displaystyle \frac{p-p_0}{\sqrt{\frac{p_0·(1-p_0)}{n}}}}$$

(5)

where the critical region for the significance level $\alpha =0.05$ has a range of $(-\infty ;$−1.65). The calculated z-statistics are $z_1=-2.19$ and $z_2=-0.39$.

Analyzing the results the following conclusions can be made:

• The second step improves the effectiveness of the method (according to the second parametric test);
• The effectiveness after the first step is statistically less than 90%;
• The effectiveness after two steps is statistically at least 90%.

At the current stage of research, no correlation between method errors and specific employee groups could not be found.

5. Conclusions

The human factor in railway systems is the subject of much research. One of the goals of implementing such studies is to minimize railway accidents. This paper presents the concept of developing job categories and assigning them to the employees who will be best suited to minimize incidents and accidents in the railway system.

The study was carried out with anonymous volunteers. The experiments were performed without any risk to the participants (a certified e-tracking unit, 12V equipment, and short sessions). According to that, the ethical commission with the University had no doubts about the research.

The approach proposed in the paper focuses on two stages. In the first, using a fuzzy logic model, feature values were assigned to each service type, obtaining three parameterized employee-station profiles for the traffic officer on duty: local steering center, classic post weakly loaded, and classic post loaded. The classification model used three features relating to essential duties, periodic duties, and job difficulty.

In the second stage, the participant advanced through two of the three available evaluation scenarios, during which mistakes made during task performance were counted. Each scenario’s main goal was to handle eight pairs of trains correctly. Complementing step two of the proposed approach was the use of eye-tracking goggles, which helped verify the confidence of the actions performed. Validation of the conducted experiments made it possible to determine the errors’ characteristics. The effectiveness of the proposed approach was 87%.

One of the most significant advantages of the proposed solution is that the developed method enables the assignment of railway employees to tasks, along with models for evaluating and classifying railway employees. This solution aims firstly to maximize operational efficiency and, secondly, to minimize the risk of human error by optimally matching personal qualities and skills to the requirements of individual jobs. According to the previous research [1] and Reason’s so-called ‘Swiss Chase’ safety model, it can be stated that the presented approach has a positive influence on the safety of the system. This happens because the method connects operators with operational situations and operated devices that cover the weaknesses of the operators.

The expected correctness of about 90% of the method was statistically reached. Also, the introduction of the second step was reasonable because it increased the accuracy.

Directions for further research will focus on increasing the test sample and to engage the Polish Rail Infrastructure Manager. Studying more areas of rail vehicle operation is also interesting, as well as analyzing the effectiveness of training strategies, and comparing training using virtual reality with traditional desktop methods. In addition, the authors look forward to the possibility of IT implementation of the method to make it more attractive for practical use.