Guanghui Liang

Electrical resistance tomography (ERT) is a non-intrusive method used to determine internal conductivity distributions using measurement voltage acquired from electrodes positioned along a boundary. Although, ERT delivers distinct strengths, reliably generating high-resolution conductivity distributions continues to be demanding owing to unstable and complex non-linear structure of inverse problem. In response, a deep neural network architecture, termed SURE-Net (spatially up-sampled residual encoder network), is proposed to enhance image clarity and shape integrity in ERT image reconstruction. The framework integrates hierarchical feature extraction with electrode-adjacency alignment, residual-based refinement, and a non-learnable feature expansion stage to promote spatially consistent reconstruction. The SURE-Net further maintains stable reconstruction behavior under limited input conditions and preserves fine structural boundaries that are typically blurred in traditional approaches. Through extensive simulation and experimental evaluations, including robustness tests under varying signal-to-noise conditions, SURE-Net demonstrates consistent conductivity reconstruction with improved feature clarity. The model reflects high accuracy under different conductivity distributions and challenging structures. Moreover, effective model adaptation from restricted inputs is supported by its design, ideally suited in practical ERT applications.

Electrical resistance tomography (ERT) uses a network of boundary electrodes to visualize and analyze processes involving multiphase flow effectively. Nonlinearity and ill-posed problems are challenging for ERT reconstruction. However, reconstruction image accuracy is in demand. Due to a lack of suitable training, the existing deep learning network for ERT image reconstruction depends on sparse information flow and gradient flow. It is challenges to represent the structural topological link among internal conductivity characteristics using the usual reconstruction approach. By propagating feature representation across topological connections, the non-Euclidean structure of the ERT measurement encodes the problems. Capturing spatial and topological dependencies requires adaptive learning to handle irregular prominent feature positions and conductivity domains. The adaptive learning is modulated by edge priority for different graph states in a dynamic graph where edge features embed spatial constraints and graph geometry. Therefore, a multi-head attention mechanism is incorporated into the graph convolutional network to maintain the local smoothness while preserving sharp transitions. The dynamic graph is learned through edge weights jointly with node features using the RMSprop optimization method. Adaptive edge learning can identify object geometries and real-time flow patterns. Using extensive numerical simulation and experiment results, the proposed algorithm improves the imaging accuracy compared to the traditional methods without dense connections. It is compared with existing deep learning methods.

Accurate measurement of the solid phase fraction in liquid–solid two-phase flow remains a significant challenge in industrial applications. Traditional methods are often constrained by the prerequisite for precise physical parameters and the inherent complexities of hydrodynamic modeling, which collectively hinder accurate online measurement. To address this, a novel data-driven framework is proposed whose core innovation lies in the unique combination of the empirical wavelet transform (EWT) and the light gradient boosting machine (LightGBM) for ultrasonic solid phase fraction estimation. This synergistic approach overcomes traditional limitations by combining EWT’s adaptive feature extraction for complex signals with LightGBM’s efficient and robust prediction, achieving rapid and reliable online measurement. The proposed method adaptively extracts features from ultrasonic signals via energy partitioning using EWT. Then, a regression model is constructed using LightGBM to enable rapid prediction of the solid phase fraction, which provides an efficient, precise, and robust solution for real-time online measurement in liquid–solid two-phase system. To validate the proposed methodology, circulating flow experiments were conducted using polystyrene particles and water. Various feature extraction methods and regression models were tested and compared, using the combined McClements and Bouguer–Lambert–Beer–Law model (MCBL model) as a performance benchmark. Results demonstrate that EWT-based energy partitioning (EWT-EP) yields superior performance in feature extraction, while LightGBM achieves higher prediction accuracy and robustness. For instance, at a single ultrasonic frequency of 7.5 MHz, the EWT-EP combined with the LightGBM model achieved a coefficient of determination ( R ² ) of 0.975, representing a substantial 71.78% improvement over the MCBL model. In addition, the implementation of a multifrequency fusion strategy enhanced overall model performance by providing richer feature representations, result in improving both prediction accuracy and generalization capability, with R ² increasing to 0.978.

A self-attention powered graph convolution network (GCN) is proposed for electrical resistance tomography (ERT) and ultrasonic transmission tomography (UTT) dual-modality tomography. It’s promising to improve the imaging accuracy in cross-sectional media distribution detection. The proposed model uses polynomial feature generation to pre-process the two input layers of measurement data. The dual-branch measurement spatial information is fused by the feature level at the dense layer with the same dimension. The proposed model calculates attention scores to extract contrasts from complex non-Euclidean distance measurement data structures. The GCN with an attention mechanism (GCNAM) identifies expressive complementary features from the combined ERT/UTT measurement data, improving prediction accuracy. During training, the attention mechanism predicts the various modal features using learnable parameters; the model then adapts to the good reconstruction target. It promotes the effective fusion and image reconstruction of dual-modal information between different modal features. The optimizer RMSprop further fine-tunes the proposed model to enhance the generalization ability. The simulation and experiment tests are carried out to evaluate the proposed model’s performance. Furthermore, it conducts generalization and robustness tests on all state-of-the-art models to evaluate their performance under varying conditions. The results demonstrate that the proposed model maintains higher imaging accuracy and stability using structural similarity (SSIM) index and root mean square error (RMSE) metrics.

Electrical resistance tomography (ERT) is a visualization detection technology possessing unique advantages in industrial dynamic process monitoring. ERT image reconstruction is an ill-posed inverse problem that is nonlinear and susceptible to measurement noise. In recent years, deep learning models have been used to learn complex patterns and relationships in sparse data, leading to more accurate and robust image reconstruction. A stacked sparse autoencoder (SSAE) based deep learning approach is proposed to map complex nonlinear functions for ERT image reconstruction. In the proposed SSAE, global conditional gradient steps employ the back projection concept and non-convex local search with the predicted model for fine-tuning. The proposed method guarantees a global optimum for the neural network sparse constraint. The model improves feature learning and representation, while the optimization technique effectively optimizes weight changes during training. A sparse autoencoder imposes sparsity constraints on the hidden layers to minimize the sparsity effect. The proposed model uses adaptive strategies to dynamically shuffle and optimize encoder-decoder weights in the architecture and optimize sparse features by shuffle shepherd optimization (SSO), which can take advantage of the proposed model’s ability to capture complex hierarchical features and the SSO’s adaptability in optimizing autoencoder parameters. The fusion of SSAE and SSO algorithms obtains an average structural similarity (SSIM) index of 0.8371 under an optimal sparsity constraint of 0.047 for reconstruction. The proposed model significantly enhances reconstruction performance with a loss of convergence of 0.0124, providing faster training and robust image reconstruction capabilities for ERT. Simulation and experimental tests are conducted to evaluate the network’s generalization ability and practicality.

Ultrasonic transmission tomography (UTT) is a crucial imaging modality for non-destructive testing and quality assessment in industrial settings. The reconstruction of high-quality images from limited, irregular, and noisy measurement of ultrasound measurements poses a significant challenge due to the ill-posed nature of the inverse problem. Therefore, the proposed method involves formulating the UTT image reconstruction problem into a graph-based learning task. In this task, nodes symbolize locations on the inspected object, while edges illustrate the connections between those locations. Furthermore, it employs the graph Laplacian function to inject graph structure into the learning algorithm to extract the structural details of the acoustic parameter distribution from the ultrasound measurements. The spectral graph convolutional neural networks (SGCNNs) enforce smoothness across the graph. The unbiased risk estimation-based spectral approach efficiently shapes the irregularity graph structure. The proposed SGCNN is tailored to effectively learn and exploit the complex relationships between transmitted ultrasound measurements and the underlying substantial properties in an ultrasound spectrum. Simulation and experimental tests are carried out to validate the proposed algorithm’s performance. In addition, the SGCNN algorithm is compared to graph convolution neural network and traditional reconstruction methods, showcasing its superior ability to reconstruct detailed and accurate images by evaluating performance metrics.

Most industrial process data are highly dynamic and non-linear with time-varying characteristics, which poses great challenges to industrial fault diagnosis tasks. Additionally, in actual industries, limited training samples are often obtained due to safety requirements and high maintenance costs. Therefore, it is crucial to effectively extract representative features and identify faults in a scenario with limited data. To address these issues, this study proposes an elastic slow feature prototypical network (ESFPN) under the meta-learning framework for few-shot dynamic fault diagnosis. First, a Siamese embedding network is customized with a loss function for slow feature extraction, aiming at characterizing process states effectively. Second, an elastic mechanism is introduced to the metric-based prototypical network, which employs elastic distances to emphasize the classification impact of more important prototype vectors. ESFPN utilizes the extracted nonlinear slow features as the elastic metric network inputs to construct slow feature prototypes, which significantly improves the fault classification performance. The performance of ESFPN and other advanced comparative methods is evaluated using the Tennessee Eastman Process. The results demonstrate the significant superiority and effectiveness of the proposed method.

The monitoring of multiphase flow distribution in industrial processes in order to optimize production presents a challenge. Electrical resistance tomography (ERT) is a technique used to visualize the inner distribution of multiphase flow. Image reconstruction plays a vital role in ERT. However, the nonlinearity and ill-posedness of inverse problems make image reconstruction in ERT difficult. The development of advanced imaging algorithms has attracted much interest for this purpose. In this work, an improved U-shaped deep learning model is proposed, which combines the advantages of the multi-scale feature extraction of UNet++ and the residual feature fusion of Res2Net. The network is designed to post-process the pre-reconstruction results of traditional ERT image-reconstruction methods, combining the generalization ability of the model-based methods and the flexible feature-extraction advantage of deep learning methods. The post-processing includes super-resolution, image denoising and artifact removal. Simulations and experiments are designed to verify the generalization ability and effectiveness of the proposed post-processing model. Both simulation and experimental results show that the proposed U-shaped network approach outperforms other deep learning methods, and the proposed deep learning model is fit for post-processing in ERT, making it a robust solution for applications.