Research

CURRENT PROJECT

Non-invasive ECG image classification and detection based on deep learning algorithm and Jetson microcontroller

Figure 1: A propose software and hardware interface for online ECG image detection.

Aims & Objectives

In this project, we aim to investigate an appropriate one of deep learning methods for ECG image classification and detection. Once the selected algorithm determined we will develop the algorithm and embed it into the Jetson microcontroller for further training and testing process. This system capable of performing online ECG image detection that associated to one of cardiology diseases. A proposed framework of the online ECG image detection is presented in Figure 1. The following are the project objectives:

  1. Development of one of deep learning methods for finding a reliable algorithm
    • To investigate an existing deep learning methods that will be suitable for online ECG image classification and detection.
    • To collect a substantial open source datasets of ECG images for training and testing purpose of the selected algorithm.
    • To develop and to refine the selected algorithm to be applicable for software and hardware connection and interface in term of time computation and complexity.
  2. Development of a hardware framework as an online ECG image detection of a certain cardiology diseases.
    • To investigate and determine an appropriate microcontroller for software and hardware connection and interface.
    • To build a hardware framework that capable of detecting one of cardiology disease based on the deep learning model obtain from deep learning process.
    • To evaluate of the performance of the hardware framework for online ECG image detection system.

Research Project Publications and Links

  1. Caesarendra, W., Hishamuddin, T.A., Lai, D.T.C., Husaini, A., Emran, M.E., “ECG Signal Classification and Prediction based on Convolutional Neural Network (CNN) Method”, to be published in Proceedings of the 2021 2nd International Conference on Information System, Computer Science and Engineering (ICONISCSE), CEC 2021, Indonesia.

Project Members

 Start date and duration

1st July 2020, 24 months

Application of Multi-objective Evolutionary Computation in Human Activity Discovery

Introduction
Evolutionary Computation (EC) is a rapidly growing and exciting field of research which has recently gained popularity due to its achievement in outperforming deep learning (a prominent area of Artificial Intelligence research) in playing Atari games [1] in June 2018. Its origin dates back to the 1950s and is an important subtopic of AI [2]. EC has been applied to complex computational problems in a wide range of domains, particularly in producing human-competitive machine intelligence [3], which encompasses problem categories such as classification, clustering, regression, design optimisation, planning and automatic computer program generation.
While Deep Learning have been popularly used to solve problems in Computer Vision such as Human Activity Analysis (HAA), there is less work done in exploring EC techniques to solve such problems despite its widespread application in the problem categories stated above. Human Activity Analysis (HAA) can be seen as comprising of a few stages as shown in Fig.1, starting from collection of data through observation, discover the various activities within the observations, learning a model for each of the activities discovered and finally use these models to recognize human activities. Majority of the works in HAA are done in the learning and recognition, while there are less works done in the discovery stage. This work is to apply EC in the discovery stage.

Figure 1: A pipeline with different processes in a broader perspective of human activity analysis [4]

Aims & Objectives
In this project, we aim to develop an evolutionary algorithm which is capable of performing human activities discovery (HAC) for an autonomous robot. The following are the project objectives:

  1. Development of multi-objective evolutionary algorithms (MOEA) for solving HAC problems
    a. To build and investigate on an existing framework (unsupervised machine learning and/or deep learning) for HAC and test on available daily living activities data.
    b. To refine existing framework using MOEA methods on activities data.
  2. Development of an autonomous robot using the algorithms developed in 1.
    a. To investigate and determine an optimal set of features for clustering of human actions using machine learning technique, i.e. autonomous discovery of human actions from observations by a mobile robot.
    b. To collect a substantial dataset of Activities of Daily Living (ADL) for current and future research in human activity recognition.
    c. To evaluate the performance of the human action discovery on available datasets of activities of daily living (ADL), and the in-house dataset

Research Project Publications and Links

  1. Lai, D. T. C., Sato, Y., “Hybrid Multiobjective Evolutionary Algorithms for Unsupervised QPSO, BBPSO and Fuzzy clustering”, to be published in Proceedings of the IEEE Congress on Evolutionary Computation, CEC 2021, Poland.
  2. https://ailab.space/projects/unsupervised-human-activities-recognition/ [research webpage 1]
  3. https://www.researchgate.net/project/Application-of-Multi-objective-Evolutionary-Computation-in-Human-Activity-Discovery [research webpage 2]

Project Members

Daphne Lai [Researchgate] [Scholar]
Ong Wee Hong [Researchgate] [Scholar] [Research]
Parham Hadikhani [Scholar]
• Adli Mahari (RA)
• Norsyamimi Razak (RA)

Start date and duration
1st January 2020, 24 months

References

[1] Wilson, D. G., Cussat-Blanc, S., Luga, H., & Miller, J. F. (2018). Evolving simple programs for playing Atari games. arXiv preprint arXiv:1806.05695. Available: https://arxiv.org/abs/1806.05695
[2] Bäck, T., Hammel, U., & Schwefel, H. P. (1997). Evolutionary computation: Comments on the history and current state. IEEE transactions on Evolutionary Computation, 1(1), 3-17. Available: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.391.802&rep=rep1&type=pdf
[3] Sipper, M., Olson, R. S., & Moore, J. H. (2017). Evolutionary computation: the next major transition of artificial intelligence?. Available: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5534026/
[4] Ong, W., Palafox , L., Koseki, T., “Autonomous Learning and Recognition of Human Action based on An Incremental Approach of Clustering”, IEEJ Transactions on Electronics, Information and Systems, Vol 135, No.9, 2015 (pdf)

Previous Research Projects

Optimal Dispatch Solution for Power Grids integrated with Renewable Energy Sources

Proposed System for Econo-Environmental Dispatch (SEED)

Under this project, we are developing an Econo-Environmental dispatch system for power grids with substantial renewable penetration. The proposed system predicts the fluctuations in the demand as well as generation through machine learning. The grid system is also modeled for its capabilities and constraints. Then the generation is optimally matched with the demand for minimum cost of energy as well as the lowest level of emissions.  

The temporal variations in the output from potential renewable sources like solar and wind are considered in the generation side. Ramping behavior of these renewable systems are modeled based on the long term performance data using Artificial Intelligence techniques. These models can give an insight to the contributions that can be expected from renewable generation in different time scales. To predict the load in a given time, a load forecasting system based on the novel “Environmental Sensitivity Index (ESI) approach has been developed. Once the load and generation requirements are quantified, the optimal dispatch solution could be suggested by the optimization engine. 


Demand side management of energy consumption in the domestic sectors in Brunei Darussalam

Solar PV systems installed in a house

Brunei Darussalam is targeting to reduce 63% of the National energy consumption by 2035. Besides, at least 10% of the power generated would be from New and Renewables sources by this target year. Hence, energy efficiency and renewable energy use plays key roles in Brunei’s long term energy scenario. Household demand accounts for 38.2% of the total energy consumed in Brunei. Hence, the domestic sector should be a prime target for energy conservation and renewable energy utilization initiatives.

With these in mind, this project aims at identifying efficient demand side management of energy use in Bruneian homes.  Long term energy demand in Brunei has been modelled using the LEAP. Energy demand for various sectors, under this 63% targeted reduction, has been modelled.  Possible energy conservation measures in the household activities are identified based on the data generated from the previous pilot demonstration project carried out by the UBD|IBM Centre. Field level experiments are also conducted.

With an average daily solar insolation of 5.43 kWh/m2/day, solar energy is the most viable renewables energy for Brunei. Possibilities of roof top solar PV systems  in Bruneian homes are also being investigated. Along with the technical aspects, a detailed cost benefit analysis following the present worth approach has also been carried out.


Short term wind power forecasting system combining Physical models with artificial intelligence techniques

Wake propagation in a wind farm (left) and results of the support vector machine model for the wake losses (right). 

Wind being a stochastic phenomenon, the power output from wind energy systems may significantly vary from time to time, in tune with the temporal changes in the strength and direction of available wind resource. These frequent changes in output may pose significant challenges in grid integration of wind farms as well as the efficient dispatch of generated wind power.  Hence, reliable forecasts of the power output from the farms, at different time intervals, are essential for the efficient management of wind energy projects.

Keeping in this view, a reliable wind farm power model has been developed employing artificial intelligence techniques. Models based on different machine learning methods are developed and their respective accuracies are estimated. These models can be trained with historic data on power output from a wind farm at a given wind speed, direction and stratification.  A representative result, showing the performance of Support Vector Machines in forecasting the power output from an offshore wind farm, is shown in the figure. These models can further be integrated with the Numerical Weather Prediction (NWP) systems to forecast the wind farm output at different time scales.