ST-HCSS: Deep Spatio-Temporal Hypergraph Convolutional Neural Network for Soft Sensing

TL;DR

The paper tackles soft sensing in industrial processes under unknown topology by learning a higher-order hypergraph to capture non-Euclidean sensor relationships and temporal dynamics. It introduces ST-HCSS, which combines a multi-view mixer, gated temporal convolution, and spectral hypergraph convolution, with an unsupervised hypergraph structure learning step that builds hyperedges via $k$-nearest neighbors and weights based on Euclidean distances. A key contribution is the end-to-end architecture that maps sliding-window inputs $\mathbf{X}$ to dominant-variable predictions while updating hypergraph representations $\mathbf{X}^{l+1}=\sigma_r(\mathbf{D}_v^{-1/2}\mathbf{H}\mathbf{W}\mathbf{D}_e^{-1}\mathbf{H}^T\mathbf{D}_v^{-1/2}\mathbf{X}^l\boldsymbol{\Theta}^l)$. Empirical results on the Cranfield Multiphase Flow Process show ST-HCSS outperforming eight baselines in NRMSE, NMAE, MAPE, and $R^2$, with the learned hypergraph adjacency mirroring data correlations, highlighting its potential for accurate, scalable soft sensing in complex industrial settings.

Abstract

Higher-order sensor networks are more accurate in characterizing the nonlinear dynamics of sensory time-series data in modern industrial settings by allowing multi-node connections beyond simple pairwise graph edges. In light of this, we propose a deep spatio-temporal hypergraph convolutional neural network for soft sensing (ST-HCSS). In particular, our proposed framework is able to construct and leverage a higher-order graph (hypergraph) to model the complex multi-interactions between sensor nodes in the absence of prior structural knowledge. To capture rich spatio-temporal relationships underlying sensor data, our proposed ST-HCSS incorporates stacked gated temporal and hypergraph convolution layers to effectively aggregate and update hypergraph information across time and nodes. Our results validate the superiority of ST-HCSS compared to existing state-of-the-art soft sensors, and demonstrates that the learned hypergraph feature representations aligns well with the sensor data correlations. The code is available at https://github.com/htew0001/ST-HCSS.git

Figures (3)

• Figure 1: Overview of the designed architecture. It begins with the (A.) raw sensor data preprocessed by (B.) that transforms into a sliding window format. Next, a (C.) multi-view mixer that extracts spatio-temporal data features. Meanwhile, an (D.) unsupervised hypergraph structure learning module to classify related sensor nodes with hyperedges. Subsequently, a (E.) convolutional-based hypergraph representation learning module that produces the final prediction.
• Figure 2: Hyperparameter analysis on ST-HCSS. First row shows kernel size of {3,5,7,9,11}. Second row represents mixer block number of {1,2,3,4,5}.
• Figure 3: Heatmap of normalized weighted hypergraph adjacency and data correlation with respect to each of the process in table I