Machine Learning Model for Predictive Maintenance of Lifts/Elevators

Project Reference :

AISG-100E-2018-013

Institution :

Singapore Management University (SMU)

Principal Investigator :

Assoc Professor Tan Hwee Pink

Technology Readiness :

5 ( Technology validated in relevant environment)

Technology Categories :

[display_term_name term_name=”technology-offer-category”]

Background/Problem Statement

The rising investment in commercial and residential infrastructure projects in the developing and developed economies has led to an increase in the growth of elevator and escalator markets. The global elevator and escalator market is projected to grow from USD 83.86 billion in 2022 to USD 132.08 billion by 2029, exhibiting a CAGR of 6.7% during the forecast period. 

Elevator maintenance companies currently adopt a reactive lift maintenance approach whereby the crew is notified of a specific fault after it has occurred and the crew rectifies the fault based on fault codes obtained from the lift monitoring device.

Solution

A baseline AI prediction model was trained on real telemetry data, operational data, as well as historical data from two lift brands. Various re-sampling techniques were applied to address data imbalances. 

A data pipeline has also been built such that (i) data preprocessing, feature extraction, training, test data creation, model training, prediction are all in place, (ii) output data from each step flows seamlessly into the next step, and (iii) new features or models can be added easily in future when required. The design of the feature extraction process enables the dependencies for each feature to be resolved smartly to minimize duplicate computations. The data pipeline allows more focus on improvement of the prediction model when more data comes in and facilitates the deployment of the whole package in the premises relatively easily.

The baseline model achieved 77.8% recall with an overall accuracy of 59.3% with random under-sampling. The corresponding recall and accuracy achieved with balanced random forests are 66.7% and 61.4% respectively. 

The above results demonstrate that using AI techniques with resampling is able to predict elevator breakdowns more accurately ( recall >50%) than a random guess. The machine learning model can predict lift breakdowns one week in advance.

The technology has the potential to be developed into a brand-agnostic lift predictive maintenance system, by training the AI model with a sizable data set comprising multiple lift brands.

Benefits

  • Reduce lift downtime
  • Increase lift lifetime
  • Increase in cost-saving for lift maintenance providers
  • Improved passenger experience when using the lifts

Potential Application(s)

Lift / Elevator maintenance in residential and commercial buildings

We welcome interest from the industry for collaboration/ co-development / customisation of the technology into a new product or service. If you have any enquiries or are keen to collaborate, please contact us.