Dynamic Hypergraph Neural Networks for Flotation Condition Recognition Based on Group-View Composite Features

Metal mineral flotation plays a critical role in the mining industry, where accurate recognition of flotation conditions is essential for optimizing operations and enhancing mineral recovery rates. Although extensive research has focused on observable froth characteristics, a substantial gap remains...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:IEEE transactions on instrumentation and measurement Jg. 74; S. 1 - 13
Hauptverfasser: Wang, Shuai, Wang, Kang, Li, Xiaoli
Format: Journal Article
Sprache:Englisch
Veröffentlicht: New York IEEE 2025
The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
Schlagworte:
Ablation
Accuracy
Algorithms
Artificial neural networks
Data models
Dynamic hypergraph neural network (HGNN)
Feature extraction
Flotation
flotation condition recognition
Graph theory
Graphs
group-view composite features
Image color analysis
Image processing systems
Image sequences
manifold structure
Manifolds
Mineral processing
Minerals
Mining industry
Multilayer perceptrons
Neural networks
Recurrent neural networks
Representations
Robustness
Static models
Training
Visualization
ISSN:0018-9456, 1557-9662
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Metal mineral flotation plays a critical role in the mining industry, where accurate recognition of flotation conditions is essential for optimizing operations and enhancing mineral recovery rates. Although extensive research has focused on observable froth characteristics, a substantial gap remains in understanding the underlying high-dimensional manifold structures within features extracted from flotation video sequences. Moreover, conventional single-view feature extraction techniques often suffer from noise caused by mutual interference, reducing the robustness and reliability of extracted features. To address these challenges, we propose group-view multipath residual neural network (GVResNet)-EvolveHGNN, a novel framework for robust feature extraction and spatial-temporal modeling in flotation video sequences. The GVResNet module adopts a GVResNet structure with a cross-attention mechanism to enhance feature stability and suppress noise. GVResNet uses average pooling (AP) of sequential images and cross-attention to compute the significance of each image feature representation, and groups features based on their relevance and distinctions. In addition, to capture evolving relationships within high-dimensional feature spaces, we introduce EvolveHGNN-an evolving hypergraph neural network (HGNN). Unlike static models, EvolveHGNN adopts the recurrent neural network (RNN) to dynamically adjust hypergraph parameters over time, effectively capturing the temporal evolution of flotation sequences. Ablation studies reveal that our dynamic hypergraph model significantly outperforms conventional multilayer perceptron (MLP) models, which fail to account for spatial structures, and also surpasses static hypergraph algorithms, achieving respective accuracy improvements of 1.41% and 1.35%. This proposed framework offers more robust feature representations, serving as a powerful tool for the accurate identification of the working conditions in mineral processing.
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0018-9456
1557-9662
DOI:10.1109/TIM.2025.3566845