Classification of particle height in a hopper bin from limited discharge data using convolutional neural network models

Shaohua Chen, Laurent A. Baumes, Aytekin Gel, Manogna Adepu, Heather Emady, Yang Jiao

Research output: Contribution to journalArticle

Abstract

Hopper bin discharge processes are ubiquitous in many industrial applications. The preponderance of previous studies of this system has been focused on the forward discharge process given an initial particle packing configuration. Motivated by the needs in reaction design and optimization, we develop here an inverse reconstruction procedure that enables one to obtain the particle height information in the initial packing configuration in the hopper bin from limited discharge dynamics data and particle characteristics. Our procedure is based on convolutional neural network (CNN) models, which take experimentally measurable particle residence time, diameter and density as input, and provide a classification of particle height in the initial packing as output. Using MFIX-DEM simulations with enhanced physical modeling capabilities, we generate extensive discharge data for hoppers containing two distinct solid phase particles with four distinct classes of initial packing geometries. The CNN reconstruction models are subsequently trained and tested using the discharge data. We find that the reconstruction (i.e., particle height classification) accuracy strongly depends on the initial packing geometry. For hopper-specific CNN models exclusively trained using discharge data from a single hopper, the classification can attain an accuracy up to almost 90%. The accuracy decreases for generic reconstruction models trained using a collection of discharge data for multiple hoppers. Finally, we apply the CNN model to accurately reconstruct a hopper containing four distinct solid phases with a layered configuration to further demonstrate its utility.

Original languageEnglish (US)
Pages (from-to)615-624
Number of pages10
JournalPowder Technology
Volume339
DOIs
StatePublished - Nov 1 2018

Fingerprint

Hoppers
Bins
Neural networks
Geometry
Industrial applications

Keywords

  • Convolutional neural network
  • Discrete Element Method (DEM)
  • Hopper bin discharge
  • MFIX-DEM simulations
  • Reconstruction

ASJC Scopus subject areas

  • Chemical Engineering(all)

Cite this

Classification of particle height in a hopper bin from limited discharge data using convolutional neural network models. / Chen, Shaohua; Baumes, Laurent A.; Gel, Aytekin; Adepu, Manogna; Emady, Heather; Jiao, Yang.

In: Powder Technology, Vol. 339, 01.11.2018, p. 615-624.

Research output: Contribution to journalArticle

@article{19239bd465874fa49af64ea89517a834,
title = "Classification of particle height in a hopper bin from limited discharge data using convolutional neural network models",
abstract = "Hopper bin discharge processes are ubiquitous in many industrial applications. The preponderance of previous studies of this system has been focused on the forward discharge process given an initial particle packing configuration. Motivated by the needs in reaction design and optimization, we develop here an inverse reconstruction procedure that enables one to obtain the particle height information in the initial packing configuration in the hopper bin from limited discharge dynamics data and particle characteristics. Our procedure is based on convolutional neural network (CNN) models, which take experimentally measurable particle residence time, diameter and density as input, and provide a classification of particle height in the initial packing as output. Using MFIX-DEM simulations with enhanced physical modeling capabilities, we generate extensive discharge data for hoppers containing two distinct solid phase particles with four distinct classes of initial packing geometries. The CNN reconstruction models are subsequently trained and tested using the discharge data. We find that the reconstruction (i.e., particle height classification) accuracy strongly depends on the initial packing geometry. For hopper-specific CNN models exclusively trained using discharge data from a single hopper, the classification can attain an accuracy up to almost 90{\%}. The accuracy decreases for generic reconstruction models trained using a collection of discharge data for multiple hoppers. Finally, we apply the CNN model to accurately reconstruct a hopper containing four distinct solid phases with a layered configuration to further demonstrate its utility.",
keywords = "Convolutional neural network, Discrete Element Method (DEM), Hopper bin discharge, MFIX-DEM simulations, Reconstruction",
author = "Shaohua Chen and Baumes, {Laurent A.} and Aytekin Gel and Manogna Adepu and Heather Emady and Yang Jiao",
year = "2018",
month = "11",
day = "1",
doi = "10.1016/j.powtec.2018.08.048",
language = "English (US)",
volume = "339",
pages = "615--624",
journal = "Powder Technology",
issn = "0032-5910",
publisher = "Elsevier",

}

TY - JOUR

T1 - Classification of particle height in a hopper bin from limited discharge data using convolutional neural network models

AU - Chen, Shaohua

AU - Baumes, Laurent A.

AU - Gel, Aytekin

AU - Adepu, Manogna

AU - Emady, Heather

AU - Jiao, Yang

PY - 2018/11/1

Y1 - 2018/11/1

N2 - Hopper bin discharge processes are ubiquitous in many industrial applications. The preponderance of previous studies of this system has been focused on the forward discharge process given an initial particle packing configuration. Motivated by the needs in reaction design and optimization, we develop here an inverse reconstruction procedure that enables one to obtain the particle height information in the initial packing configuration in the hopper bin from limited discharge dynamics data and particle characteristics. Our procedure is based on convolutional neural network (CNN) models, which take experimentally measurable particle residence time, diameter and density as input, and provide a classification of particle height in the initial packing as output. Using MFIX-DEM simulations with enhanced physical modeling capabilities, we generate extensive discharge data for hoppers containing two distinct solid phase particles with four distinct classes of initial packing geometries. The CNN reconstruction models are subsequently trained and tested using the discharge data. We find that the reconstruction (i.e., particle height classification) accuracy strongly depends on the initial packing geometry. For hopper-specific CNN models exclusively trained using discharge data from a single hopper, the classification can attain an accuracy up to almost 90%. The accuracy decreases for generic reconstruction models trained using a collection of discharge data for multiple hoppers. Finally, we apply the CNN model to accurately reconstruct a hopper containing four distinct solid phases with a layered configuration to further demonstrate its utility.

AB - Hopper bin discharge processes are ubiquitous in many industrial applications. The preponderance of previous studies of this system has been focused on the forward discharge process given an initial particle packing configuration. Motivated by the needs in reaction design and optimization, we develop here an inverse reconstruction procedure that enables one to obtain the particle height information in the initial packing configuration in the hopper bin from limited discharge dynamics data and particle characteristics. Our procedure is based on convolutional neural network (CNN) models, which take experimentally measurable particle residence time, diameter and density as input, and provide a classification of particle height in the initial packing as output. Using MFIX-DEM simulations with enhanced physical modeling capabilities, we generate extensive discharge data for hoppers containing two distinct solid phase particles with four distinct classes of initial packing geometries. The CNN reconstruction models are subsequently trained and tested using the discharge data. We find that the reconstruction (i.e., particle height classification) accuracy strongly depends on the initial packing geometry. For hopper-specific CNN models exclusively trained using discharge data from a single hopper, the classification can attain an accuracy up to almost 90%. The accuracy decreases for generic reconstruction models trained using a collection of discharge data for multiple hoppers. Finally, we apply the CNN model to accurately reconstruct a hopper containing four distinct solid phases with a layered configuration to further demonstrate its utility.

KW - Convolutional neural network

KW - Discrete Element Method (DEM)

KW - Hopper bin discharge

KW - MFIX-DEM simulations

KW - Reconstruction

UR - http://www.scopus.com/inward/record.url?scp=85052447428&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85052447428&partnerID=8YFLogxK

U2 - 10.1016/j.powtec.2018.08.048

DO - 10.1016/j.powtec.2018.08.048

M3 - Article

VL - 339

SP - 615

EP - 624

JO - Powder Technology

JF - Powder Technology

SN - 0032-5910

ER -