Leader: David Camacho (UPM)

Objectives - The main objective of this WP will be to develop a comprehensive framework for advanced time series analysis that leverages transfer learning, pre-trained models and self-supervised learning, while also integrating the neural network techniques with visual analytics to improve pattern discovery, representation learning and anomaly detection in time series data.

• To study the limitations of the current time series analysis techniques when faced with complex industrial problems
• To create novel time series models by incorporating methods such as transfer learning, pre-trained models and self-supervised learning
• To harness the capabilities of visual analytics techniques for creating innovative explanation methods that are applicable to time series analysis and outlier detection challenges

Tasks:

T1 Fusing Neural Networks and Visual Analytics for Explainable Time Series Analysis

This aim of this task will be to create an integrated approach that harnesses the synergy of eXplainable Artificial Intelligence (XAI) and advanced visualisation techniques, thereby effectively merging the feature extraction and representation learning capabilities of neural networks with the intuitive interpretability that is afforded by visual analytics for time series data. The ultimate objective is to offer a holistic solution for comprehending, elucidating and potentially predicting the patterns and anomalies within time series datasets. The principal outcome of this endeavour will be the development of XAI models that will significantly improve the interpretability of the deep learning model. In addition, advanced visualisation techniques will not only bolster result the explanations of the results but will also enable the anomalies and outliers within the time series data to be identified. To underscore the practicality of the implemented models, this project will be conducted across various industrial domains including but not limited to space and industry, thus demonstrating their real-world feasibility and applicability.

T2 Advancing Time Series Analysis via Transfer Learning, Pre-trained Models, and Self-Supervised Learning

This task constitutes a comprehensive endeavour that is designed to confront the multifaceted challenges that have long hindered traditional time series analysis approaches, which chiefly stem from the copiousness and heterogeneity of temporal data. In the attempt of overcoming these obstacles, we are poised to harness the formidable capabilities that are offered by the state-of-the-art deep learning techniques. These encompass the art of transfer learning; whereby pre-trained models are adroitly fine-tuned to cater to related time series tasks. Additionally, we will employ the resourcefulness of pre-trained models—deep learning constructs that have previously been trained on a diverse array of time series data—to facilitate feature extraction and transfer learning in our context. Furthermore, our arsenal will include self-supervised learning, a method that empowers models to autonomously generate labels directly from the data, thereby rendering it exceptionally suitable for scenarios that are characterised by limited availability of labelled data. Our primary goal within this task is to embark on a voyage of exploration and innovation in the realm of time series analysis, all of which will be underpinned by the deployment of these cutting-edge deep learning techniques, with the overarching objective of improving the accuracy, interpretability and adaptability of methodologies.

T3 Deployment and Evaluation of Developed Models and Tools in Real-world Scenarios

This task constitutes a pivotal phase that will be dedicated to the practical deployment and empirical assessment of the diverse models and techniques that had been meticulously crafted in Tasks T3.1 and T3.2. With unwavering attention paid to the translation of the theoretical constructs into tangible solutions, this task takes centre stage in implementing and rigorously validating these innovative methodologies within genuine industrial landscapes. By orchestrating the integration of our cutting-edge developments into authentic, industry-specific contexts, we anticipate not only assessing the feasibility of these techniques but also forging strategic collaborations with the pertinent sectors and industrial stakeholders who share our vision and objectives. This intersection of academia and industry serves as a testament to the pragmatic potential of our models and promises to offer tangible benefits in real-world settings.