This article presented an innovation and technology project of the Electrical and Mechanical Services Department (EMSD) of the HKSAR Government for the enhancement of lift safety and reliability. As at the end of 2020, there are about 70,000 lifts in Hong Kong with about one-third of them aged more than 30 years old. In general, these lifts have relatively more frequent breakdown due to components aging. Equipping these aged lifts with remote monitoring systems could instantly check the operating condition, and alert in the early stage of failure or upon detection of fault symptoms so as to carry out timely maintenance, reduce breakdown time and even prevent accidents.
The rapid advancement of Artificial Intelligence (AI) in recent years is now driving the innovation and technology development in the world. While the use of AI on lift monitoring is still at the initial stage as of today, the EMSD started developing a smart lift monitoring system (hereafter referred to as “the system”) in 2018 using AI and remote monitoring technology, and working in collaboration with the practitioners from the industry, academic and research sector.
The system facilitates lift owners, building management offices, lift manufacturers and lift maintenance contractors to have collaborative analysis and diagnosis of the operating condition of the lifts. The data of real-time lift operation status and alerts can be viewed and downloaded from a common platform (Figure 1). The system is designed for keeping aged lifts as well as modern lifts away from breakdowns and accidents, by informing responsible maintenance parties via the common platform to take immediate corrective actions and further conduct predictive maintenance, thereby enhancing the lift maintenance efficiency and performance.
Figure 1: Common platform for lift monitoring
Invention inspired by pulse diagnosis of traditional Chinese medicine
The pulse diagnosis of the traditional Chinese medicine inspired the invention of the system. Through pulse touching, the health condition of a person can be identified generally. This type of non-intrusive sensing and monitoring technique can avoid detrimental side effects to bodies. This is the analogous to the system adopting non-intrusive means to detect the operating condition of lifts.
The AI, acting as the Chinese medicine practitioners, examines lift operation with non-intrusive sensing and monitoring. The system incorporating AI analysis models can provide fault detection (eg safety circuity tripped) and potential fault prediction for multi-brand lifts without intervening the existing circuits of the lifts. The system can give alerts of lift breakdowns for corrective maintenance and potential faults for predictive maintenance based on electric current signals of lift traction motors, door motors, brake coils and safety circuits, as well as lift car vertical positions.
Simple, low cost and non-intrusive design
The system makes use of remote monitoring units (RMU) for signal acquisition and backend servers for data storage and AI analysis (Figure 2). The RMU consists of non-intrusive electric current sensors which are configured to acquire real-time electric current data of the traction motor, brake coil, door motor and safety circuit of lifts. A LiDAR for lift car position monitoring may also be added as an optional item to the RMU. The electric current sensors are clamped on the power wires of the said lift components, without intervening the circuits of the lifts. The RMU will send the acquired signals to the backend servers for data storage and AI analysis.
Figure 2: System diagram
The capital cost and services down time for on-site installation sometimes are the hindrance of equipping remote monitoring systems for lifts, especially for aged lifts. As such, the system is all along developed according to the guiding principles of simplicity, low cost and non-intrusiveness without sacrificing the monitoring capability at all. The installation of RMU could be completed within half-day and the cost of RMU starts at HK$5,000 which should be affordable to most lift owners.
AI analysis models
The fault detection and prediction are achieved by analysing the electric current data and the lift car position by AI analysis models which utilise both the expert system models and the deep learning models, complementing each other. Through continuous AI learning of the relationships between the electric current data and faults, the AI analysis models can give alerts for lift breakdowns and potential faults.
The expert system can be regarded as a computer system simulating the decision-making ability of a human expert. It can determine and predict faults based on statistical features and predetermined rules. Each expert system model can be applied to perform signal processing or classification tasks for a specific brand or model of lift. The expert system model can extract statistical features of the acquired electric current data from the different segments of lift operating cycles (Figure 3). Those segments can be used to determine whether the lift being monitored is in normal operation or not, and detect specific abnormalities (eg excessive door opening/closing time, abnormal motor/brake restart, improper brake open when door opens, etc.). To enable accurate fault prediction for individual lift, it is necessary to spend up to four weeks after the RMU is installed for parameter configuration towards the predetermined rules of the expert system models.
Figure 3: Electric current data of a lift operating cycle (blue: door motor current, green: brake current, violet: safety circuit current and red: traction motor current)
With a view to compressing the time for system parameters configuration to less than two weeks, a deep learning model based on a cutting-edge AI architecture, namely “Multivariate Long Short–Term Memory Fully Convolutional Network (MLSTM-FCN)” (Figure 4), was developed to adaptively monitor the lift operating conditions.
Figure 4: Expert system and deep learning monitoring
The MLSTM-FCN leverages the advantages of both multivariate long short-term memory network and convolutional network. The MLSTM-FCN can adaptively learn the relationships between the acquired data and lift faults, and output the prediction results. MLSTM-FCN can therefore automatically classify fault characteristics of individual lift into relevant fault categories.
The deep learning model enables faster and more intelligent fault detection and prediction as compared with the expert system, but massive data is required to train up the deep learning model. The learning results of an individual lift could be applied to the learning processes for other lifts with similar configurations.
Lift fault simulations – Essential to AI analysis models development
Lift fault simulations are useful means to enable the AI analysis models to learn the relationships between electric current data and faults. The simulation of a critical fault on the brake that would lead to fatal accidents is illustrated below.
The drum type electromagnetic brakes (Figure 5) tested in this simulation are widely used in lifts in Hong Kong. When the lift is called to run, the brake coil is energised electrically, thereby creating sufficient magnetic force to the plunger. The plunger exerts a force on the compression springs for driving the brake arms and associated brake pads apart from the brake drum. The traction motor is then allowed to rotate. Upon the brake coil is de-energised by cutting off electricity supply, the brake is applied. The compression springs push the brake arm with associated brake pads. The brake pads then press on the brake drum.
Figure 5: Drum type electromagnetic brake
In this simulation (Figures 6 and 7), the compression springs were deliberately set to more compressed such that magnetic force was insufficient to pull the plungers apart and the brake pad barely touches the brake drum while rotating. This setup caused sluggish brake operation and boosts up the electric current of the traction motor and brake coil. The traction motor current and the brake coil current at normal and faulty brake operation were shown in Table 1. The data was input to the AI analysis models for differentiation of normal and faulty brake operation, thereby enhancing the ability of fault detection and prediction.
Figures 6 and 7: Simulation setup for brake faults
|Traction motor current at mid of travel (A)||Brake coil current at mid of travel (A)|
|Normal brake operation||8.1||1.6|
|Faulty brake operation||46.0||2.4|
Table 1: Results of normal and faulty brake operation
Going forward - Enhancing AI performance
The RMU of the system have been deployed in 15 government lifts till now. The system performance is satisfactory and promising. The deployment scale of the RMU is being expanded gradually now to cover different lift brands and models, allowing the system to acquire more data for training up the deep learning models to achieve more accurate fault detection.
More emerging AI technologies meanwhile are being researched and incorporated into the system by the project team. Focus is now on the use of generative adversarial network and imbalanced learning to enhance the fault detection and prediction capabilities under the scenario of insufficient or limited lift fault data.
Potential smart applications for lift users
The application of the system is not just limited to the enhancement of lift safety and reliability. The information of real-time lift car positions and some lift operating conditions collected by the system may be shown to lift users through mobile apps, with a view to enabling users to better plan their lift journeys, such as notifying lift availability, reducing waiting time in lift lobbies, etc. This would be a good application for smart living.
The system has been recognised both internationally and locally, including being awarded a Silver Medal at the International Exhibition of Inventions of Geneva 2021 and acclaimed at Outstanding Technology (Building) at the Build4Asia Awards 2020. The system is granted a Hong Kong Patent (No.: HK30012023) and pending for the US Patent. A technical paper for the system was published in the IEEE Access Journal (reference link: https://ieeexplore.ieee.org/document/9333567).
Gratitude goes to Ir Charles Wong Kai-hon and his team of Pro-Act Training and Development Centre (Electrical) for their collaborative effort and advice on data analysis and lift fault simulation. The effort of other partners from the industry, academic and research sector is also highly appreciated.
About the authors: Ir CHEUNG Ka Kei, Dr LI Xuran Ivan and Mr Eric C H YIP are from The Electrical and Mechanical Services Department, The HKSAR Government.