Machine learning has a proven track record of advancing our lives, whether it is automating tasks and processes, gaining valuable insights out of massive quantities of data and enabling us to take the most effective data-driven actions.
But machine learning and IoT also play a crucial role in saving lives from train accidents, as well as identifying and preventing suicides on train tracks. This is the topic Victoria Chudinov, Data Scientist at DSB Digital Labs will present at the virtual Maintenance Analytics Summit. She will explore possible solutions for preventing accidents and suicides with trains and share learnings from a project to prevent suicidal behaviour and accidental collisions on the railway track in Denmark.
We caught up with Victoria to talk to her about this noble application of ML and IoT.
Hyperight: Hi Victoria, we are very glad to have you again with us, this time as a speaker at the virtual Maintenance Analytics Summit 2020. Let’s start with a couple of words about yourself and your role at DBS Digital Labs for our readers.
Victoria Chudinov: Hi and thank you for having me. In a nutshell, I am a data scientist by day and a fencing instructor by night. And we are here to talk about the day job. It says data scientist in the title, but my role at the DSB Digital Labs is quite broader than just that. The goal of our department is to come up with new and innovative solutions to all the problems a transportation company like DSB might have – from predictive maintenance of our rolling stock to creative ways to use the free spaces at stations.
This means that me and my colleagues are involved in design sprints, prototyping, and so on, in addition to the usual data science and machine learning tasks. We have to figure out when and how to use the tools of the trade to solve completely new problems and help the business make decisions. Sometimes this might be as simple as “I don’t think this project is suitable for ML”, other times – we develop major components of the product. We also advise the business on what Data Science is, where it helps and what is needed.
Hyperight: You are going to deliver a presentation on a very crucial topic that requires the attention of everyone working with data – IoT solutions that save lives from train accidents. Could you tell us a bit more about this DSB Digital Labs project?
Victoria Chudinov: “Saving Lives” was a project that began at the Digital Labs as a design sprint to… well, save lives. Train collisions have a major impact on many people beyond the victim and their close ones – our train drivers, witnesses, response teams, support personnel. The entire transportation network is thrown in disarray, as the tracks where the collision happened are closed. We wanted to investigate whether we can do something to prevent such collisions. We consulted psychologists, researched technologies, read publications and got into arguments. And we found out a few important things:
- There are different groups of people that become victims of train collisions. There are those who want to commit suicide, but a substantial proportion of the collisions are accidents – people being in the wrong place at the wrong time.
- We found that physical, dumb, measures are very effective – fences, locks, and so on. And a possible approach would be to completely deny entry to the stations and the tracks. Unfortunately, given the scale, this is very costly. And it does not solve the problem completely. A determined person can get around these, and they offer no notification, no information, no way to prompt someone to take action.
- The vast majority of train tracks cannot be covered by personnel. Even a lot of stations have nobody on them.
This means that we need a solution that can inform relevant people and help them take timely action. And this is where IoT can help us. This could be as simple as a motion sensor that issues an alert when someone is in a tunnel, and as complex as behaviour analysis algorithms that notify us when someone is showing signs of suicidal behaviour or is putting themselves in danger.
Hyperight: You are going to talk about a range of solutions from simple sensor-triggered devices, to complex machine learning-driven solutions that detect and prevent suicidal behaviour. AI and ML have found a vast application in identifying suicidal behaviour, mainly relying on patient data and health care data. What kind of data are you using in your models, how do you collect it, analyse it, and use it to make decisions and take actions?
Victoria Chudinov: I am afraid we don’t have a very clear answer yet. We are aiming at low-hanging fruits – simple sensors that do not need trained models, and can send us information on the basis of a simple trigger. Or even simpler – above mentioned “dumb” solutions like fences and posters. Pre-trained human-detection algorithms are also something suitable for the places where there shouldn’t be any people.
However, things are more complicated at the stations, where there are many people. There are tell-tale signs of suicidal behaviour. But unfortunately, human behaviour is very complex and has a lot of variation. Our research showed that implementing an algorithm that can do this consistently is a major challenge. And tracking someone’s behaviour over multiple locations raises a lot of ethical and privacy questions. Finally – there is simply very little data, that on top of everything is noisy, follows different formats and overall too difficult to work with. And even if we had more of it, someone will have to go through it, label it, prepare it, which given the nature of the assignment is going to have a major psychological toll. So we then chose the path to rely on more common statistics to identify hotspots, times and patterns and use these to take preventive measures.
Another problem we faced was with the human component here. You might have a great system that detects a person in risk with 100% accuracy, but who is going to take action on that, and what are they going to do? Alerting the train drivers and the Control Center is an obvious first step. But some locations are remote, and it might take a while for someone to get there. And in the cases of suicidal behaviour, that person will have to be trained to deal with the situation. Then, of course, the question is how much accuracy do you need in order to take action. Stopping the trains from running is a huge cost, but so is a human life. So both false positives and false negatives have huge costs and make decision making rather tricky.
Hyperight: Apart from using ML to prevent train accidents and save lives, where elsewhere do you apply machine learning at DSB? And how machine learning, in general, can help optimise railway operational performance?
Victoria Chudinov: Most recently, we are developing a solution to inform people of how crowded trains are and which trains to take, as we tackle the post-COVID-19 world. This “crisis innovation” project has already sparked great interest in the company, and we hope this solution will become a more permanent thing and will see a wider adoption once it’s ready and validated. We have implemented projects for predictive maintenance, price-formation, and a few others. The potential is really big – scheduling, optimisation, prediction and forecasting problems are everywhere in a transportation company.
Hyperight: And lastly, despite all the benefits of machine learning, what are some areas where it can be improved so we can have better outcomes from the models? What are the difficulties you are facing as a Machine Learning Engineer?
Victoria Chudinov: People and data. With people there are two things – on the one hand, it is about educating people what Machine Learning CAN do, helping them realise the potential of the technologies and how it relates to what they do. On the other, it is tempering down expectations, helping people understand what the technology cannot do, and this without killing their vibe.
With data – usually, it is data quality. A lot of big companies that want to take advantage of machine learning and data science have been collecting data, yes, but rarely this has been done with the considerations of ML in mind. As such, a lot of this data might not be suitable for the task at hand, might be inconsistent, and a lot of times just not available. Since we deal with a wide range of topics, we also have come to understand the importance of domain experts, and there is also considerable detective work involved in finding the right people for a given data source.