We are international front runner in applying artificial intelligence, machine learning and cloud computing technologies in smart infrastructure solutions. We work with government and industry partners on projects covering utilities, energy, telecommunications, and transport.
Trains are an important means of transportation which serve millions of residents in mega cities. In order to keep the operation running smoothly, the performance of trains, e.g., punctuality and fuel efficiency, need to be reasonably predicted beforehand. Then the operator can carry out timely adjustments on train operation, for example, changing timetable to prevent further delays, planning maintenance to reduce faults, and giving extra time for overcrowded stations, based on the insightful predictions we provided. Modelling such operations, traditionally, is barely possible due to the large varieties of factors, such as complicated train network, unexpected incidents, human activities, etc. Therefore, the outcome of this project aims to build the platform to continuously analyse the train performance and furthermore give reasonable suggestions to the operator by agilely aggregating massive data source.
The solution is built on innovative machine learning techniques including retrieval techniques, ensemble models, monitoring system, and continuous evaluation. The innovated platform can aggregate different types of data sources in an agile manner with minor data format adjustments. It has involved more than 100 factors, from days to seconds, and it will be providing more insights for any hypothesis suggested by experts or knowledge discovered by additional data.
Trains, as an primary transportation mode, play an increasingly important role in keeping a city moving and achieving urban sustainability. The train timetable should be evaluated regarding its punctuality, reliability and robustness before its implementation. However, this is a challenging task due to the train network complexity and patronage fluctuations. This project utilises a wealth of historical train performance, customer and incident data to evaluate a timetable and help create more robust timetables.
We solved the problem by integrating comprehensive data sources, identifying inspiring insights and patterns and developing advanced evaluation models by using cutting-edge machine leaning technique. The solution is based on innovative analytical methods or machine learning techniques, including deep learning, ensemble learning and computer vision. Various breakthroughs from prediction modelling, risk estimation, and complexity learning are presented. The outcomes of this project enable our client evaluate a timetable effectively in several minutes and the evaluation results and insights are presented in a dynamic and interactive manner. These outcomes could help the operation and timetable design teams create more robust timetables.
With the increasing demand for public transportation due to congested highways, trains have become one of the most viable alternatives, especially for daily commuting. While transit agencies are excited with the increasing ridership, they are also challenged with a higher volume of passenger flow and longer queuing lines at the existing stations. Large scale renovation of rail stations to cater for growing passenger demand has been taking place in various sites all over the world. Railway passenger flow data including the arrival and departure of each train, the number of passengers getting on and off at each stop and the number of transferring people in each station can be used for planning facilities, train scheduling and other planning. It is very important to schedule train timetables properly so that they meet the increasing passengers’ demand.
To improve the current situation and plan for the future, we have developed real-time object detection and tracking models based upon advanced computer vision techniques with the help of installed cameras in the stations. The solutions include using advanced deep learning based object detection technology to recognize trains in each frame (image); tracking the passengers flow on the platforms in continuous frames (images); tracking the passengers stepping on and off the escalator in continuous frames (images); tracking the actual train wheel stop/start time in continuous frames (images). Those data-driven solutions help the train controllers and managers to better understand the passenger demand at a given station and better schedule the future train timetables.
Original Equipment Manufactures (OEMS) of electronic products often offer service warranty to their customers. To maintain high-level customer satisfaction, spare parts are required to be stocked at a sufficient level to support the service. As the suppliers may stop producing certain models due to technological/economic reasons, it is important to order one last quantity of spare parts, known as last time buy (LTB), to mitigate the risk of stock shortage. Spare parts by LTB play a significant role in inventory maintenance workload.
This project is a collaboration between a prominent electronic producer and the team, which endeavours to reduce the workload of inventory maintenance. As an initial step, we have built a visualisation tool that closely monitors the engineering KPIs based on producer's install base and failure data, by utilising our machine learning capability. Currently we are developing predictive models that take into account part failure, install base, stock and warranty data, for forecasting the quantity of LTB. The predictive models are expected to provide failure predictions for parts, and hence optimize the stock management.
Routine inspection of rail network is essential for the safety and efficient operation of trains. Train operators usually conduct inspections of the whole network using the inspection vehicle, capturing tens of millions of images of tracks and surrounding environment. These images will be manually reviewed by track inspectors to find out any defects, which requires years of domain expertise and training. With the building of new rails and aging of existing rails, it is increasingly challenging to manage the workload of inspectors while keeping the high safety standard of the track inspection process.
Our team developed deep learning models to detect regions of interests for suspicious defects. The models are trained using labelled defect images and normal inspection images, and able to detect missing clips/bolts from inspection images with high accuracy. As a result, train operators could extend the range of inspection that previously overlooked due to limited resources, assist inspectors in the decision making process, and improve the efficiency of the inspection process while maintaining high maintaining standard.
Understanding customers’ expectation on services is the key for broadband service providers. For new service roll-out, service providers need to allocate suitable amount of workforce and resources in appropriate time and location. For existing service regions, service providers aim to maintain top-tier customer service and sustainable revenue. The predictions of activation demand and utility rate could be very challenging due to the involvement of multiple interacting factors such as service types and local demographics.
Our team designed innovative machine learning model for customers’ demand forecasting in fine-grained spatial and temporal dimensions. The model can handle heterogeneous data, and offers great scalability for large-scale data. The techniques developed in this project could be applied to other customer-centered services. It provides data-driven support for decision making process, which helps the service provider to better manage the life cycle of service roll-out.
Smart parking management based on sensors and real-time data can save drivers' time in finding parking spaces, optimise the usage of parking spaces and support operational staffs to easily monitor and manage parking resource. However, installing sensors at all parking spaces will be expensive and inefficient. This project aims to develop a prediction model for parking space occupancy rates using machine learning techniques. Our solution learns the patterns via historical occupancy records and estimates the occupancy for those spaces without sensors.
Our team has developed an innovative solution to predict occupancy rate of parking spaces based on sensor and transaction records. Our model considered various types of features to learn the patterns of parking behaviour, including location and regulation information, historical sensor parking events, and parking payments. The model can output accurate occupancy rate for parking spaces without sensors at any time.
Water utilities often manages water supply and demand for millions of residential and non-residential properties. It is paramount to guarantee continuous supply, water quality and pressure, as well as to minimise operational costs. In this Intelligent Network Optimisation project, we developed new machine learning models for water demand forecasting, water quality modelling and dosing optimisation, which effectively predict water demand in the near future and define optimal strategies for chlorination. New machine learning models are also being built to optimise energy usage, by balancing pumping regimes and costs while meeting water demand. All these models will work together to optimise water supply network.
We have developed a cash flow forecasting approach based on machine learning techniques. This approach can provide an overview of the company's cash position in a short term in the future. We use machine learning techniques to forecast a company's transactions in the next 8 weeks based on the historical patterns of cash inflows and cash outflows. The cash flow modelling has been divided into three components and we have provided solutions for each component: 1) cash -- predict when and how much cash sales and cash expenses will be received or paid; 2) transfer -- predict when and how much transfer within the company will occur; 3) invoice and bill: predict when payments for invoices and bills will be received or paid.
The cash flow of a company and their bank accounts can be computed by aggregating the predict results for these three components. The outcomes can support businesses to foresee how much cash they will have in their accounts and make informed decisions based on these information.
Sewer corrosion is a serious issue in wastewater systems for Water utilities. The concentration of H2S is a key factor leading to sewer corrosion, but it can be alleviated by dosing of Iron (Ferrous) salts and other chemicals. A chemical dosing profile should be specified to determine the dosing rate based on many dynamic factors, e.g., H2S level, flow volume. In order to improve the dosing effect and save chemicals, the chemical dosing profiles for each Chemical Dosing Unit (CDU) should be optimised according to the dynamic factors.
This project developed a smart toolkit to facilitate the optimization process. The toolkit consists of three major modules, i.e., sewer dosing analysis, H2S estimation and sewer corrosion prediction. It enhances the capability of the Sydney Water teams by improving their performance and investment in controlling H2S, managing corrosion and odour in concrete sewers. The target users of the toolkit are the operators in Water utilities.
There are approximately 800 jack arches underneath the bus lane of the Sydney Harbour Bridge. The jack arches are very difficult to access and are typically inspected at two year intervals according to standard visual inspection practices. RMS required a way to continuously monitor the structural condition of each of the jack arch supports and alert the asset manager and bridge inspector if any issue is detected. An inspection can then be scheduled and preventative maintenance are carried out without disrupting road users.
We have designed and developed a monitoring system including data acquisition, data analytics and a user dashboard to monitor each of the jack arch components in real time using machine learning techniques. Each jack arch has been instrumented with three low-cost tri-axial accelerometers connected to a smart node, resulting in about 2,400 sensors in the system. The acceleration data have been used to train a machine learning model and detect any deviation from the model. This SHM system was awarded "The Most Practical SHM Solutions for Civil / Mechanical Systems Award" at IWSHM - the 10th International Workshop on SHM in 2015.
We have collaborated with a variety of utilities on asset failure predictions since 2011, currently the assets include supply water mains, sewers, and gas pipes. Among the projects, we developed technologies to predict the risk of these assets. The prediction is used for cost-effective maintenance to save cost, reduce uncertainty for asset network, and enhancing the custom experience.
We developed a data-driven solution for predictive maintenance of assets. It is analytics as a service platform that can derive insights from operational data, signal future failure risks, as well as provide decision support for asset owners. It can identify the causal factors of asset failures, prioritize high-risk pipes for renewal and then systematically reduce failure risk over time. The platform is now applied into problems including supply water main failure prediction, sewer choke prediction, and gas pipe leakage prediction. Near doubling of consequence cost saving can be achieved based on the case studies. It is able to recommend high-risk assets, based on failure probabilities, consequence costs, renewal costs, budget constraints, and geographical constraints. The improvement of current prediction results to drive better tactical decisions.