Inverse Modeling of Dielectric Response in Time Domain using Physics-Informed Neural Networks
Authors:
Emir Esenov,
Olof Hjortstam,
Yuriy Serdyuk,
Thomas Hammarström,
Christian Häger
Abstract:
Dielectric response (DR) of insulating materials is key input information for designing electrical insulation systems and defining safe operating conditions of various HV devices. In dielectric materials, different polarization and conduction processes occur at different time scales, making it challenging to physically interpret raw measured data. To analyze DR measurement results, equivalent circ…
▽ More
Dielectric response (DR) of insulating materials is key input information for designing electrical insulation systems and defining safe operating conditions of various HV devices. In dielectric materials, different polarization and conduction processes occur at different time scales, making it challenging to physically interpret raw measured data. To analyze DR measurement results, equivalent circuit models (ECMs) are commonly used, reducing the complexity of the physical system to a number of circuit elements that capture the dominant response. This paper examines the use of physics-informed neural networks (PINNs) for inverse modeling of DR in time domain using parallel RC circuits. To assess their performance, we test PINNs on synthetic data generated from analytical solutions of corresponding ECMs, incorporating Gaussian noise to simulate measurement errors. Our results show that PINNs are highly effective at solving well-conditioned inverse problems, accurately estimating up to five unknown RC parameters with minimal requirements on neural network size, training duration, and hyperparameter tuning. Furthermore, we extend the ECMs to incorporate temperature dependence and demonstrate that PINNs can accurately recover embedded, nonlinear temperature functions from noisy DR data sampled at different temperatures. This case study in modeling DR in time domain presents a solution with wide-ranging potential applications in disciplines relying on ECMs, utilizing the latest technology in machine learning for scientific computation.
△ Less
Submitted 28 April, 2025;
originally announced May 2025.
Cutup and Detect: Human Fall Detection on Cutup Untrimmed Videos Using a Large Foundational Video Understanding Model
Authors:
Till Grutschus,
Ola Karrar,
Emir Esenov,
Ekta Vats
Abstract:
This work explores the performance of a large video understanding foundation model on the downstream task of human fall detection on untrimmed video and leverages a pretrained vision transformer for multi-class action detection, with classes: "Fall", "Lying" and "Other/Activities of daily living (ADL)". A method for temporal action localization that relies on a simple cutup of untrimmed videos is…
▽ More
This work explores the performance of a large video understanding foundation model on the downstream task of human fall detection on untrimmed video and leverages a pretrained vision transformer for multi-class action detection, with classes: "Fall", "Lying" and "Other/Activities of daily living (ADL)". A method for temporal action localization that relies on a simple cutup of untrimmed videos is demonstrated. The methodology includes a preprocessing pipeline that converts datasets with timestamp action annotations into labeled datasets of short action clips. Simple and effective clip-sampling strategies are introduced. The effectiveness of the proposed method has been empirically evaluated on the publicly available High-Quality Fall Simulation Dataset (HQFSD). The experimental results validate the performance of the proposed pipeline. The results are promising for real-time application, and the falls are detected on video level with a state-of-the-art 0.96 F1 score on the HQFSD dataset under the given experimental settings. The source code will be made available on GitHub.
△ Less
Submitted 29 January, 2024;
originally announced January 2024.
