Elevator-Optimising-Using-Machine-Learning

Project developed for NTU DeepLearning Hackathon

1. Problem Statement and Aim

Today, more and more IoT devices are already changing every aspect of life, and elevators as an important part of life are no exception. For example, KONE uses the Watson IoT Platform running on IBM Cloud to collect streams of incoming data about movement, vibration, loading, and other information.

However, we've noticed that despite more and more elevator companies are concerned about how their elevator's functionality, not all of their products are running with IoT devices embedded.

Therefore, we want to provide a solution that can be retrofitted to any existing elevator and provide companies with a range of data about user experience and elevator performance through data processing and Machine Learning.

2. Problem Elaboration

Every modern elevator will record this information during its runtime: Action (door open, door close), direction (up, down, idle), current floor, etc. These raw data simply record the status of the elevator, they record every event, but they are not used for optimizing the elevator performance.

Our algorithm should use data processing and ML techniques to transform these raw data into something that's related to user experience. For example, wait time. We've drawn 3 sub-problems from this demand:

Is wait time related to the time of the day?
Is wait time related to which floor the user is on?
What factor affects wait time the most?

With these subproblems answered, one would be in a good place to improve the elevator's scheduling algorithm. And that is our ultimate goal. Unfortunately, due to limited time in this 3-day Hackathon, we don't have enough time and resources to keep fine-tune a machine learning-based elevator scheduling algorithm to beat traditional elevator algorithms.

3. Exploratory Data Analysis

The dataset we used is collected from 4 different elevators in an office building in New York City, which we classified as a time-series problem. Through forward-search and backwards-search in the time series, we observe the action taken by the elevator, i.e., door opening and closing, to determine user in-and-out. Through analyzing these data, we answered the 3 subproblems which we formulated earlier.

4. Future Exploration and Development Directions

We aim to improve elevator operation efficiency using Machine Learning Algorithms.

The idea is to build a prediction model that calculates the likelihood of requests in every time instant. To do that, we created an elevator class to simulate a single elevator environment. We also built a random-request generator that would create realistic requests for the elevator. The generator will consider the trends we discovered from EDA.

Our goal is to implement the algorithm and compare it with the SCAN Disk distribution algorithm which is commonly used for elevator services.

The comparison will be of 2 aspects: user experience (wait time), and elevator wear (determined by the distance moved) However, due to limited time and resources in this 3-day Hackathon, we didn't manage to fine-tune the model.

It will be our future research direction for this project.