| I&T Solution |
Development of AI Models for Predictive Maintenance of Lifts
(REF: S-1138) |
| Trial Project |
|
| Solution Feature |
- A project team with relevant fields of specialization in data science, enterprise asset management, risk management and emerging technologies (including AI, mobile technologies, etc.). Our firm have won several relevant awards globally and locally.
- System architecture is 1. flexible (can be on-premise or cloud infrastructure), 2. scalable (to support large number of lifts and additional sensor data), 3. support live data ingestion and timely processing, 4. open source (for easy adoption) and 5. secured and governed (security controls and data governance applied)
- Systematic AI modelling approach over 15 months: 1. Environment setup and industrial analysis, 2. Data preparation (internal and external, where applicable), 3. Model design, 4. Implementation and finetune, 5. Evaluation and feedback and 6. Reporting (Roadmap and handover). We will adopt Cross-Industry Standard Process for Data Mining (CRISP-DM) approach.
- With inputs from lift engineers, analyze sensor data and also other selected potential external factors (to be agreed with EMSD) to identify correlation and predict maintenance requirements.
- Various machine learning algorithms, including Random Forest (Classification) and ARIMA (Time Series) will be considered and applied in the solution. Metrics, e.g. Accuracy, Precision, Root Mean Squared Error (RMSE), will be used to measure the performance of algorithms.
|
| Trial Application and Expected Outcome |
- We will work with lift engineers and develop the AI models to increase the AI predictive maintenance accuracy in 15 months. Each model may take up to two months to develop, enhance and finetune. With more sensor data provided, the models will be finetuned to enhance their predictive capabilities.
- For variables in the dataset, we will perform correlation analysis which will give us an idea about the degree of relationship between two variables in the dataset. Highly correlated variables might be removed for a better performance result.
- The quality of dataset is important to model training, apart from improving the data quality at data collection stage, data transformation also plays an important role. Such techniques include but are not limited to removing noise & outliners, handling missing data, normalization, and converting categorical or timestamp data.
- Different machine learning algorithms will produce different predictive maintenance results. We will test and measure the results of the relevant algorithms and provide recommendations on which algorithms are appropriate.
- We understand ESMD require open source AI models stored on an EMSD server. Given current Covid-19 pandemic and internal network access requirements, we may not be able to gain access to EMSD premise or the server to work. We may propose an alternate approach involving cloud or hybrid computing.
|
| Additional Solution Information |
Predictive Maintenance of Lifts Discussion Document - EMSD v1.0.pdf
|
| Info on I&T Solution Provider |
| Solution Provider | : | Ernst & Young Advisory Services Limited | | Address | : | 27/F, One Taikoo Place, 979 King's Road, Quarry Bay, Hong Kong | | Contact Person | : | Poh Tiong Wee |
| Position | : | Associate Partner | | Tel | : | +852 2629 3288 | | Email | : |
tiong-wee.poh@hk.ey.com | | Webpage | : | https://www.ey.com/en_gl |
|
