A Hybrid Deep Physics Neural Network Model for Unsupervised Autoencoder to Perform Data Conditioning

Denoising, smoothing, missing data, and outlier issues that occur in many fields involve finding models for noise and data. However, the models often make assumptions of the noise, such as white noise, Gaussian noise, etc. Numerical methods [wavelets, Fourier transforms, Kalman filter, principal com...

Celý popis

Uložené v:
Podrobná bibliografia
Vydané v:IEEE International Symposium on Signal Processing and Information Technology s. 1 - 7
Hlavní autori: Madasu, Srinath, Rangarajan, Keshava P.
Médium: Konferenčný príspevok..
Jazyk:English
Vydavateľské údaje: IEEE 01.12.2019
Predmet:
autoencoder
data conditioning
data denoising
data-driven
deep neural network
hybrid model
ISSN:2641-5542
On-line prístup:Získať plný text
Tagy: Pridať tag
Žiadne tagy, Buďte prvý, kto otaguje tento záznam!
Popis
Shrnutí:Denoising, smoothing, missing data, and outlier issues that occur in many fields involve finding models for noise and data. However, the models often make assumptions of the noise, such as white noise, Gaussian noise, etc. Numerical methods [wavelets, Fourier transforms, Kalman filter, principal component analysis (PCA), etc.] are available for data conditioning. However, they suffer from inaccuracies. This paper presents a new approach, in which a hybrid physics-based model and deep physics neural network (DPNN) model are used to simultaneously build a data conditioning autoencoder (DPNNAE) based on DPNN for enhancing the signal strength, compared to a conventional autoencoder (AE). It necessitates integration of domain-specific physics-based models with a data-driven approach. A one-dimensional (1D) stochastic Burgers equation is used as a prototype for demonstrating the new algorithm but can be easily applied to other systems. This paper provides a new data conditioning approach that performs denoising, smoothing, data imputation, and outlier removal using DPNN and AE. This approach offers an improved and efficient methodology to arrive at the noise-free data and fills the missing data. During this work, the DPNN was applied to predict noise-free velocity in the unsteady viscous stochastic Burgers equation. The loss function is obtained from the physics models produced from domain insight and measured data. Thus, the deep neural network (DNN) framework integrates physics-based models into its framework. The hybrid DPNN model uses measured data from sensors; both measured data and physics are simultaneously satisfied by the model resulting in noise-free measured data. The conditioned data can be further used to build the DPNNAE model for further conditioning, and the DPNN can be used for filling the missing data. The DPNN method outperformed other common denoising algorithms. The new hybrid modeling data conditioning algorithm developed during this work can be applied to any real-time modeling system with real-time measurements for denoising.
ISSN:2641-5542
DOI:10.1109/ISSPIT47144.2019.9001770