.png){kind=icon}
.png){kind=icon}
Return
Hierarchical Complementary Enhanced Autoencoder Integrating Spatio-Temporal Interaction Feature for Soft Sensor
.png){kind=icon}
PRE
AI
DOI：10.1002/apj.70131.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
To address the issues of neglecting the spatiotemporal correlations among process variables, low-level features are vulnerable to noise interference, and the gradual loss of key information layer by layer during deep network training in traditional stacked autoencoder-based soft-sensor models, this paper proposes a hierarchical complementary enhanced autoencoder integrating spatio-temporal interaction feature (ST-HCEAE) model. Firstly, a spatio-temporal interaction network (STIN) is constructed to extract the global and local spatio-temporal interaction relationships. Temporal features are captured by using the temporal attention mechanism. The maximum information coefficient (MIC) is adopted to obtain the global spatial interaction relationship across time steps. The gated graph attention (GGAT) mechanism is proposed to obtain the local spatial interaction relationship at each time step, and the spatio-temporal interaction features are obtained through adaptive fusion. Secondly, the hierarchical complementary enhancement (HCE) module is introduced. By adopting the dynamic fusion strategy of upsampling hierarchical complementary enhancement (HCE-up) and downsampling hierarchical complementary enhancement (HCE-down), and leveraging bidirectional information compensation between adjacent hidden layers, the problems of noise interference in the feature extraction process and key information deficiency in deep network training are effectively addressed. Afterwards, multi-layer features are fused through the gating mechanism to establish a regression prediction model. Finally, two industrial cases, namely the debutanizer tower and sulfur recovery, were adopted for experimental verification. The experimental results show that, compared with the other five existing modeling methods, the proposed ST-HCEAE method has higher prediction accuracy.
Keywords:
graph attention network
maximum information coefficient
soft sensor
stacked autoencoder
Journal
IF:
1.6
Papers:
176
Citations:
2.2K