In the history of London’s Underground rail network (popularly known as Tube), the underground fire hazard incident of 1987 at the King’s Cross station was one the biggest disasters. It claimed 31 lives and also severely scarred the reputation of London Underground. After multiple inquiries, it was concluded that the fire had started because someone dropped a lit match onto a wooden escalator. But how did such a trivial incident snowball into a full-blown disaster?
Pulitzer-prize winning author and journalist, Charles Duhigg has analyzed this question in his international bestseller, “The Power of Habit”. Among many other factors that contributed to the disaster, the following chain of events in the run-up to the accident caught our eye:
- A daily commuter noticed a burning tissue paper at the bottom of an escalator and immediately informed the on-duty Ticket Collector (TC). The TC beat out the fire but didn’t investigate further. Moments later, another passenger noticed a wisp of smoke and hit the emergency stop button.
- A policeman also saw a smoky haze. He tried to inform his department but failed to contact the headquarters due to lack of signal strength on his radio in the underground area. So, he went upstairs and called the headquarters.
- The tube’s safety inspector didn’t inform the fire brigade because he was instructed not to call them unless absolutely necessary. Besides, he felt that shouting out ‘fire’ would have caused panic among the commuters.
- A good while after the first alert was made, fire trucks arrived but they couldn’t connect their hoses to the hydrants. Nobody knew the whereabouts of the underground hydrants and no one had the blueprint of the station layout.
Charles concluded that it was a total failure of communication within the organization that led to this accident. He suggests that, like individuals, organizations also have habits which define how they function and like any habit these can also be changed by taking the right steps. We agree with Charles’ insights about how altering the routine habits of people can have a wide-ranging impact on an organization. However, there is one constraint that acts as a glass ceiling on the level of performance: “Reliance on humans”.
Humans have limitations
In the words of Daniel Kahneman, humans are very susceptible to “biases, shortcuts and cognitive illusions.” Constant undivided attention requires a lot of mental effort and is difficult for humans to sustain for a long time. Also, people are emotional and their decisions in an organization are impacted by many dynamics related to personal mental state, relationship with co-workers, office politics. The understanding of what could be an emergency is also not consistent across people. We see a strong case for an Artificial Intelligence system to step in and aid humans in decision-making relating to safety.
The Times: They Are Changing
With the advent of smartphones, wireless data, and social media, the times are changing. Users of public transport now actively share their feedback to network operators via social media. Even a technologically laggard network like the Indian railways now gets ~5,000 tweets per day from customers. India’s railway minister has even made it mandatory for all his senior officers to be active on Twitter. All the stars now seem to have aligned for AI to enter the emergency response and customer service space.
The Solution: AI-Driven Emergency Response System
We propose a solution where an Artificial Intelligence algorithm will monitor all the incoming customer communication via Twitter, Facebook, online chats and identify whether they relate to some critical emergency situation (like fire, faulty equipment, crime). Once identified as an emergency case, the system will figure out which department and location (e.g. safety department of king’s cross station) are best equipped to handle the situation and automatically push the customer complaint to mobile phones of all the relevant stakeholders. By getting the first-hand information directly from the customer and facilitating quick communication, response to the emergency will become more effective.
To drive home the point, we can see how the above-proposed system could have potentially averted the King Cross’ fire on multiple levels. The commuter who first saw the burning tissue paper could have registered an online complaint via online chat or social media. By analyzing the commuter’s message, our AI would identify it as a fire-related emergency situation, figure out that this is a case for the fire department and understand where the complaint has originated from.
Safety personnel at the King’s Cross station and the nearest police/fire stations would have immediately received the alert and would have got added to a group chat on their AI-driven (emergency response) app. The problem in locating the station blueprint could also have been avoided as all the stakeholders would now be involved in a group chat thereby making communication easier. The response time of the incident could have been drastically reduced and smooth communication could have saved the day.
The machine learning algorithm that sits at the heart of this solution has two main tasks to accomplish: 1) Identifying an emergency situation; 2) Notifying the right people about the situation.
Identifying the emergency situation:
- Identifying the emergency situation is a problem of finding a needle in a haystack. Very few tweets talk about emergency situations, hence the Artificial Intelligence problem to solve here is to write a Machine Learning (or Deep Learning if we are using deepness) algorithm which can fit on a very unbalanced dataset. When you train an algorithm in a situation where a small fraction of data is interesting and most of it is not, the algorithm tends to start identifying everything as uninteresting. You will even get a high accuracy if you use traditional accuracy measures, but it would be useless in real life (calling all 10 tweets non-emergency when in reality 1 of them was emergency tweet gives 90% accuracy). To attain the objective we are talking of here, you need special loss functions, sampling at training time and methods like building a stack of multiple classifiers each refining the results of previous one to solve this problem. With real-world datasets, the task of writing a classifier has become a piece of cake with newer, cooler and smarter libraries coming in, fitting them on a dataset as unbalanced as this one would require serious experience.
Identifying the right people:
- The problem of communicating to the right people can be further broken down into two aspects; identifying the relevant department (police, safety, control room) and location (name of the train, station, crossing, tracks). For identifying the location, we can deploy our NER (Name Entity Recognition) APIs that can extract entities from natural language texts like tweets and comments. For identifying the department, we can create a dataset of issues (fire, crime, disruption etc.) mapped to the relevant department and use a custom trained text classifier to route tweets classified as an emergency to the relevant department. At ParallelDots, we help enterprises build such custom classifiers based on their category list or they use our standard Taxonomy API to categorize real-time feeds of user-generated content to generate actionable insights.
Besides, in the case of a high influx of customer complaints related to a single incident (like many people tweeting about the smell of smoke in the air at an underground station), all the tweets can be grouped together into a single issue using our Feed Clustering algorithm and make the communication more streamlined and manageable. Such clustering can be achieved by leveraging our Semantic Search API which can identify the similarity between two texts even when keywords between them are not common.