Automatic Train Collision Prevention System

Abstract
Railway systems are entering an era of highly intelligent automation where stability
and safety are becoming increasingly important. However, there is still a lack of
intelligent and effective ways for railway accident prevention, especially active
accident prevention strategies. This paper presents a railway accident prevention
method based on the reinforcement learning model and multi-modal data to achieve
active railway accident prevention strategies. Three metrics are designed to show
the performance of active prevention methods. Based on the three metrics and the
data from Federal Railroad Administration, the effectiveness of the proposed method
is verified in the case study by introducing two methods as baselines. The results
also show that nearly 30% of accidents can be effectively prevented through active
preventive measures with the proposed method. Finally, this paper analyzes the
influence of personal skills on the proposed model and makes relevant suggestions
for improving railway safety based on the analysis of the results.

Automatic Train Collision Prevention System, a groundbreaking solution

designed to enhance railway safety by preventing collisions between
trains. Equipped with infrared sensors, this innovative system detects
approaching trains and automatically halts the engine (motors) upon
detecting any obstruction ahead.
Introduction
Safety is a long-lasting topic for railway systems. The occurrence of railway
accidents may cause substantial property losses and casualties. According to
reports from the Federal Railroad Administration (FRA), the total number of
accidents/incidents has reached 92,000 in the past ten years (2012–2021), resulting
in 60,948 injuries and 6486 fatalities.1 Meanwhile, the total number of
accidents/incidents was 9215 in 2012 and 8035 in 2021(9745 in 2019), while the
total fatalities were 558 in 2012 and 766 in 2021. Although efforts [1], [2], [3] over the
past decade have achieved certain results, the number of accidents is still at a high
level. Traditional accident prevention strategies have reached a bottleneck, and a
more active strategy is needed to prevent the happening of railway accidents.
Active accident prevention is to recognize and eliminate the factors that will trigger
accidents in advance through automatic risk detection technologies and effective
responses. For example, through vision sensors, some possible conflicts can be
detected and used to warn passengers to prevent collision risks [4]. Conversely,
passive prevention refers to the methods that examine as many accident factors as
possible in the hope of ruling out future factors which may have caused accidents
[5]. Recognition of possible hidden risks of accidents is the main characteristic of
 active prevention [6]. Moreover, active accident prevention is a combination of
accident prediction, causal analysis, and accident data analysis.
In railway systems, the implementation of active accident prevention suffers from the
challenges of accident detection and the limitations of insufficient pre-accident
response. First of all, the response to prevent accidents needs a risk predictor with
high reliability to ensure that the response is valid. So, risk prediction is required to
be as accurate and specific as possible. However, many factors, including weather,
geography, equipment, train condition, passenger behavior, and even unobservable
factors, could be the trigger for a railway accident, while many prediction models can
only consider part of them because it is hard to get the information of all the factors.
Secondly, the pre-accident response, which means the response to prevent an
accident after accident risk is detected, requires specific information about the future
accident. This means that only knowing whether an accident will happen is not
enough, and the cause of this accident is also needed. Therefore, many accident
prediction methods cannot satisfy the need for active accident prevention because
they only predict the occurrence or frequency of accidents [7], [8], [9]. Moreover,
response time should also be considered an important factor. Some responses for
railway accidents may need a long time, and thus the prediction model should
ensure the data can be accessed in advance to reserve enough time for active
accident prevention. To sum up, successful implementation of active accident
prevention for railway systems needs to address the following issues: an accurate
prediction of accident occurrence, a specific prediction of accident cause, and full
use of multi-source data to reserve enough time for pre-accident response.
This paper presents a reinforcement learning-based method to predict the causes of
railway operational accidents along with time flows to achieve active accident
prevention for railway systems by addressing the above issues. The main idea of this
method is to achieve a dynamic scenario-response mode to anticipate the cause of a
possible accident and take the corresponding preventive measures in advance to
take the initiative to prevent accidents. The proposed model is mainly based on the
historical information of accidents and related characteristics, so data for prediction
are within broader time constraints, ensuring that possible accident information can
be learned in advance and corresponding prevention and control can be carried out.
Moreover, the multi-modal data, including structured data and unstructured data
(text), are combined through embedding vectors and natural language processing
(NLP) technologies to achieve full use of the information contained in the accident
records.
The main contributions of this paper are as follows:
 (1)
A reinforcement learning (RL) method is introduced to achieve railway active
accident prevention. The RL method achieves an effective way to learn useful
information from accident records and predict the cause of the coming accident
ahead of a period, allowing the staff to make a good preparation for accident
prevention. Three metrics are designed to verify the performance of the proposed
method.
 (2)
Multi-modal data, including text and numerical data, is organized for accident
prediction, representing categorical data by embedding vectors.
 (3)
Randomness is introduced to be an indicator denoting the influence of personal skills
to make the proposed method aware of the impact of human factors. Considering
 that accident prevention highly depends on the experiences of employees, the
randomness makes the proposed method able to reflect the influence of personal
skill level. In the experimental section, this paper presents the results from the
different levels of randomness and advises some measures for improving personal
skills.
The rest of this paper is organized as follows. The implementation of active accident
prevention requires the combination of accident prediction, accident cause analysis,
and full use of railway accident data. So, Section 2 reviews the related works,
including railway accident prediction, cause analysis, and data analysis in railway
safety. Moreover, RL-based methods that are applied to railway systems are also
reviewed to shed light on the advance of RL-based methods in railways. In
Section 3, the data used in this paper is described from point of the motivation of this
research. Section 4 presents the proposed reinforcement learning-based method.
Section 5 shows the experimental results based on the operational accident data.
Finally, some discussions and conclusions are drawn in Section 6 and Section 7.

Highlights
 •
An active accident prevention method for the railway system is proposed.
 •
A reinforcement learning method is introduced to learn the pattern of
accidents.
 •
Multi-modal data, including text and numerical data, is used for accident
prediction.
 •
The influence of human factors is analyzed, and related safety
measures are advised.