Abstract : The proposed mini-symposium is focused on a few novel techniques in IAM, their applications and promising opportunities. The techniques considered rely on artificial intelligence methods to solve problems in engineering, like wind turbine preventive maintenance or predicting molecular weights of industrial polymers using diffusion NMR spectroscopy; to advance in making AI more reliable by enabling it to cope with causality and thereby enhancing its explainability; to promote neural networks capable of directly processing geometric entities and use them for robust deep learning in various domains, including artificial vision; and to tackle with the engineering problems of multiphasic electric power generation.
Organizer(s) : Sebastian Xambó-Descamps, Yolanda Vidal, Eduardo U. Moya Sánchez
[05352] Novel deep learning methodologies in Industrial and Applied Mathematics
Format : Online Talk on Zoom
Author(s) :
Sebastian Xambó-Descamps (IMTech and BSC)
Abstract : This talk, with the same title as the MS, is meant to be the first and it aims at a broad presentation of the most promising novel methodologies in IAM based on deep learning techniques, with a particular attention focused on those pioneered by the MS speakers.
[05354] AI Lifecycle Zero-touch Orchestration within the Edge-to-Cloud Continuum for Industry 5.0
Format : Talk at Waseda University
Author(s) :
Marta Barroso Barroso (Barcelona Supercomputing Center)
Abstract : AI is one of the biggest megatrends towards the 5th industrial revolution. Although these technologies promise
business sustainability as well as product and process quality, it seems that the ever-changing market demands and the complexity of technologies, impede broad application and reuse of Artificial intelligence (AI) models across the
industry. KnowlEdge is an European project funded by the Horizon 2020 (H2020) that aims to develop of a new gene-ration of AI methods, systems, and data management infrastructure in order to break the entry barriers for these tech-nologies and unleash its full potential. In particular, knowlEdge project converges techniques from multiple computing areas, including AI , distributed data analytics, IoT, software engineering, edge and Cloud technologi-es into a unified software architecture. The outcomes of the project not only enable the automated extraction and utilization of data co-ming from multiple and geographically dispersed sources, it also provides a way of reusing and sharing AI models in
an (semi-)automated way in particular companies that are only able to perform the execution of models rather than the training themselves.
[03575] Innovative Models for Explainable Artificial Intelligence
Format : Online Talk on Zoom
Author(s) :
Silvia Franchini (National Research Council of Italy)
Francesco Prinzi (University of Palermo)
Salvatore Vitabile (University of Palermo)
Abstract : Traditional data-driven ML approaches show very interesting performance even if their internal mechanisms are very cryptic (black box). However, in some critical contexts, model interpretability is mandatory to explain the learned functionality, becoming even a legal requirement. Among the benefits of reformulating neural networks through the geometric calculus paradigm, geometric interpretability could potentially serve as a characteristic that improves model transparency. This work proposes the use of higher-dimensional neurons to reduce computational complexity while preserving model accuracy.
[05376] Applications of Quaternion Monogenic Signal ConvNet Layer
Format : Online Talk on Zoom
Author(s) :
E. Ulises Moya-Sanchez (Universidad Autonoma de Guadalajara/Gobierno de Jalisco)
Genaro Paredes (Universidad Autonoma de Guadalajara)
Sebastian Xambó-Descamps (UPC)
Ulises Cortes (BSC)
Abraham Sanchez (Gobierno de Jalisco)
Abstract : The monogenic ConvNet layer is a quaternion bio-inspired input layer. This layer creates a new geometric feature space using the Fourier transform. This new representation assigns a structural and geometrical interpretation to each image point and allows the detection of local symmetry elements (such as line-like or edge-like). Its main strength is that it behaves robustly under a variety of illumination transforms. In this work we present the design details and characteristics of this layer and consider a number of situations in which it can be applied.
[05353] Artificial Intelligence for Wind Turbine Predictive Maintenance
Format : Online Talk on Zoom
Author(s) :
Yolanda Vidal (Universitat Politècnica de Catalunya. Jordi Girona 31. 08034. Barcelona. VAT: ESQ0818003F)
Abstract : This proposal states a data-driven predictive maintenance (PM) strategy for wind turbines that uses artificial neural networks with Bayesian regularization and Levenberg-Marquardt optimization. The proposed strategy aims to address challenges associated with SCADA data such as high dimensionality, low sampling rate, and unbalanced datasets. The strategy will be validated on real SCADA data from a wind farm consisting of 12 wind turbines and is expected to provide reliable predictions with minimum false alarms and early warnings months in advance. This PM approach can help reduce the levelized cost of energy (LCOE) of wind farms and promote renewable energy as a cost-effective solution to achieve energy independence and combat climate change.
