)
Analysing Time-Varying Signals
Time series machine learning uses algorithms to analyse temporal patterns and dependencies in sequential data for tasks like classification, clustering, regression, and anomaly detection, enabling predictions or pattern recognition without relying on traditional statistical methods.
Calling in the Wilderness: The Use of Passive Acoustic Monitoring in Biodiversity Surveys
One-shot X-ray Angiography Video Segmentation by Local Matching
Artificial Intelligence Ensemble for Identifying Windfarm's Power Cable Failures
We also have a selection of projects in other themes with relevance to Time Series Analysis:
Time Series Classification
Participants
Summary
Time series classification (TSC) is a type of machine learning where the objective is to develop algorithms that can assign sequences of data points, typically indexed in time, to different categories. Time series data are abundant in many application domains, including health care, finance, speech and audio processing, human activity monitoring, manufacturing, environmental monitoring, and more. Researchers at UEA have focused on both the development of state-of-the-art TSC algorithms and their application to real-world problems. Notable algorithms produced at UEA include the Elastic Ensemble, the Shapelet Transform, and the current state-of-the-art for TSC, HIVE-COTE2. Examples of previous and ongoing applications of TSC at UEA include non-invasive authentication of alcoholic spirits, classification of dangerous mosquito species, TSC with EEG readings, detection of right whales in the North Atlantic Ocean, classification of electrical devices from smart meters, and various projects involving human activity recognition.
A key priority of TSC research at UEA is emphasising reproducible and open science. This commitment led to "The Great Time Series Classification Bakeoff," one of the largest experimental comparisons ever conducted in the field of machine learning to compare leading TSC algorithms across a wide range of problems. Supporting this work, researchers at UEA contribute to several open-source data repositories (timeseriesclassification.com, UEA multivariate time series classification archive) and codebases - for example, UEA researchers were invited to collaborate with the Alan Turing Institute to create open-source tools for time series, with the output of this collaboration continuing with aeon: an open-source toolkit in Python for learning from time series (https://github.com/aeon-toolkit/aeon)
Calling in the wilderness: the use of Passive Acoustic Monitoring in biodiversity surveys
Participants
Jointly hosted by the British Trust for Ornithology, Jen's PhD focusses on developing acoustic classifiers to analyse data collected from a landscape-scale survey carried out in Polesia as part of the Endangered Landscapes Programme. We used convolutional neural networks to develop an acoustic classifier to identify 18 species of European small mammals. We also investigated data augmentation methods to increase the amount of training data which can be difficult to obtain for small mammals.
Our work now focuses on the application of our classifier to analyse the survey data collected in Polesia and the ecological insights we can determine in terms of small mammal distribution. We also focus on the development of novel audio transformer models to identify 10 bush cricket (Orthoptera) species using waveforms and spectrograms as input.
)
Partners
){kind=logo}
){kind=logo}
Publications
[1] Milner, Ben ; Sfeclis, Georgiana-Elena ; Websdale, Danny. / Investigating Imaginary Mask Estimation in Complex Masking for Speech Enhancement. Paper presented at 31st European Signal Processing Conference (EUSIPCO).
One-shot X-ray Angiography Video Segmentation by Local Matching
Participants
Summary
High-quality, densely annotated data serve as a crucial foundation for developing robust X-ray angiography segmentation models. However, obtaining per-object pixel-level annotations in the medical domain is both expensive and time-consuming, often requiring close collaboration between clinical experts and developers. This project aims to reduce the annotation costs of X-ray angiography videos by leveraging one-shot video object segmentation (OSVOS), which separates target objects from the background using only the masks from a single frame during both training and testing. We introduce a novel OSVOS model that employs a local matching strategy, restricting the search field to the most relevant neighbouring pixels. Extensive experiments on both public and private datasets show that our proposed OSVOS method outperforms current state-of-the-art video segmentation methods in terms of segmentation accuracy and generalization capability (i.e., seen and unseen categories). This work presents a robust and data-efficient annotation tool for X-ray angiography segmentation. It offers enhanced flexibility and potential for a wide range of clinical applications.
)
Artificial Intelligence Ensemble for Identifying Windfarm's Power Cable Failures
Participants
Summary
Offshore windfarms have been increasingly constructed around the coast of the UK and other countries to generate more renewable electricity. However, as the generated electricity needs to be transmitted to the power grid onshore via high voltage cables, these cables can fail frequently at the harsh undersea environments. It has been identified as a major issue for offshore windfarm operators and accounted for 75-80% of the total cost of offshore wind insurance claims.
This project sets out to investigate the characteristics and patterns of cable faults through analysing the high voltage (HV) cable fault data recorded by the IED (Intelligent Electronic Device) and then to use the discovered knowledge to develop methods to detect and predict faults as soon as possible.
Our preliminary analyses with different temporal, spectral and statistical features extracted from the data identified some patterns of faults and built some machine learning ensembles. The test results showed that they can predict the faults with an accuracy of 97%. More work should be done with more data to verify our findings.
)
Partners
){kind=logo}
Funding
)