Lifeng Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Lifeng Zhang has a total of 37 co-authors. Here are the top co-authors by number of shared publications:
- Zhihao Tang
Baoding Key Laboratory of State Detection and Optimization Regulation for Integrated Energy System, North China Electric Power University, Baoding, China
6 shared publications - Chuanbao Wu
Baoding Key Laboratory of State Detection and Optimization Regulation for Integrated Energy System, North China Electric Power University, Baoding, China
5 shared publications - Junzhi Liu
Baoding Key Laboratory of State Detection and Optimization Regulation for Integrated Energy System, North China Electric Power University, Baoding, China
5 shared publications - Bin Wang
Yanzhao Electric Power Laboratory of North China Electric Power University
3 shared publications - Guozhi Li
Baoding Key Laboratory of State Detection and Optimization Regulation for Integrated Energy System, North China Electric Power University, Baoding 071003, People’s Republic of China
3 shared publications - Hanlin Cheng
Yanzhao Electric Power Laboratory of North China Electric Power University
3 shared publications
Latest Publications
- 24 September 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (38) , 385405
Deep unfolding network-based temperature reconstruction for acoustic tomography-based measurement with multi-scale feature fusion
To improve the inverse reconstruction stage of acoustic tomography (AT)-based temperature field measurement, a novel neural network called CPP-Net is proposed. It is a temperature reconstruction model built on a multi-scale deep unfolding neural network inspired by the Chambolle–Pock algorithm. The model embeds trainable modules through a deep unfolding framework and introduces a dual path fusion block architecture to facilitate multi-scale feature extraction and feature fusion in the signal domain. This design improves the structural fidelity of the reconstructed temperature distribution. A series of representative temperature field models are constructed for simulation studies, and the reconstruction performance is further validated through experiments on an AT-based temperature measurement platform. Simulation results show that CPP-Net achieves improved reconstruction accuracy across various temperature distribution scenarios, with an average relative error of 0.13% and a root mean square error of 0.17%, outperforming various traditional algorithms and deep learning-based methods. In experimental evaluations, CPP-Net also yields the lowest temperature reconstruction error at thermocouple sampling points, indicating its potential for practical AT-based temperature measurement applications.
- 24 September 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (38) , 385405
Deep unfolding network-based temperature reconstruction for acoustic tomography-based measurement with multi-scale feature fusion
To improve the inverse reconstruction stage of acoustic tomography (AT)-based temperature field measurement, a novel neural network called CPP-Net is proposed. It is a temperature reconstruction model built on a multi-scale deep unfolding neural network inspired by the Chambolle–Pock algorithm. The model embeds trainable modules through a deep unfolding framework and introduces a dual path fusion block architecture to facilitate multi-scale feature extraction and feature fusion in the signal domain. This design improves the structural fidelity of the reconstructed temperature distribution. A series of representative temperature field models are constructed for simulation studies, and the reconstruction performance is further validated through experiments on an AT-based temperature measurement platform. Simulation results show that CPP-Net achieves improved reconstruction accuracy across various temperature distribution scenarios, with an average relative error of 0.13% and a root mean square error of 0.17%, outperforming various traditional algorithms and deep learning-based methods. In experimental evaluations, CPP-Net also yields the lowest temperature reconstruction error at thermocouple sampling points, indicating its potential for practical AT-based temperature measurement applications.
- 9 September 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (36) , 365301
Gas parameter measurement and flow pattern identification and analysis in gas–liquid two-phase flow using a multi-scale multi-task cross-feature adaptive network
Accurate measurement of gas – liquid two-phase flow parameters is of great significance for industrial instrumentation and process monitoring. In this paper, a multi-scale multi-task cross-feature adaptive network (MMCAN) is proposed based on signals acquired from an array conductivity sensor for simultaneous flow-pattern classification, gas volume fraction (GVF) measurement, and gas-flow-rate measurement. A multi-scale weighted incremental entropy (MWIE) representation is introduced to characterize the temporal complexity of conductivity signals across different scales and is integrated into the network as a physics-inspired statistical prior. Meanwhile, a lightweight gated cross-feature fusion mechanism adaptively balances MWIE features and deep temporal features. Under a stricter condition-wise split involving 28 unseen operating conditions, MMCAN achieves an average flow-pattern classification accuracy of 96.46%, with root mean square errors of 0.0572 and 0.0453 for gas flow rate and GVF, respectively. For within-distribution comparison under the window-level split, the corresponding values are 99.74%, 0.0024, and 0.0021. Ablation, entropy-feature comparison, and gating analyses further verify the effectiveness of MWIE, stochastic encoding, and the adaptive fusion strategy. These results demonstrate the accuracy of MMCAN and its condition-level generalization capability within the same experimental facility.
- 9 September 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (36) , 365301
Gas parameter measurement and flow pattern identification and analysis in gas–liquid two-phase flow using a multi-scale multi-task cross-feature adaptive network
Accurate measurement of gas – liquid two-phase flow parameters is of great significance for industrial instrumentation and process monitoring. In this paper, a multi-scale multi-task cross-feature adaptive network (MMCAN) is proposed based on signals acquired from an array conductivity sensor for simultaneous flow-pattern classification, gas volume fraction (GVF) measurement, and gas-flow-rate measurement. A multi-scale weighted incremental entropy (MWIE) representation is introduced to characterize the temporal complexity of conductivity signals across different scales and is integrated into the network as a physics-inspired statistical prior. Meanwhile, a lightweight gated cross-feature fusion mechanism adaptively balances MWIE features and deep temporal features. Under a stricter condition-wise split involving 28 unseen operating conditions, MMCAN achieves an average flow-pattern classification accuracy of 96.46%, with root mean square errors of 0.0572 and 0.0453 for gas flow rate and GVF, respectively. For within-distribution comparison under the window-level split, the corresponding values are 99.74%, 0.0024, and 0.0021. Ablation, entropy-feature comparison, and gating analyses further verify the effectiveness of MWIE, stochastic encoding, and the adaptive fusion strategy. These results demonstrate the accuracy of MMCAN and its condition-level generalization capability within the same experimental facility.
- 27 August 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (34) , 346108
PAAD-Net: a physics-aware asymmetric decoupling network for remaining useful life prediction of lithium-ion batteries
Accurate remaining useful life prediction of lithium-ion batteries is crucial for reliable operation of battery management systems. However, conventional data-driven models often blindly mix thermodynamic and kinetic properties, leading to prediction error during the non-linear capacity plunge at the end-of-life. To address this critical bottleneck, this paper proposes a physics-aware asymmetric decoupling network (PAAD-Net) with adaptive physical constraints. Specifically, the model utilizes causal convolutions and asymmetric cross-attention to decouple the thermodynamic-equivalent and kinetic-related health indicators, which are then processed by a specialized capacity decoder to ensure achieve prediction. Most importantly, this paper designs an adaptive physics-informed gradient constraint by leveraging the gating signal to dynamically modulate the weight of the physical loss. Extensive comparative experiments on CALCE and Oxford datasets demonstrate that PAAD-Net achieves low mean absolute error of 0.0070 Ah and 0.0011Ah, respectively. Experimental results verify that PAAD-Net achieves high-precision prediction throughout the full lifecycle, with comparative and ablation experiments further validating its accuracy and generalization.
- 27 August 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (34) , 346108
PAAD-Net: a physics-aware asymmetric decoupling network for remaining useful life prediction of lithium-ion batteries
Accurate remaining useful life prediction of lithium-ion batteries is crucial for reliable operation of battery management systems. However, conventional data-driven models often blindly mix thermodynamic and kinetic properties, leading to prediction error during the non-linear capacity plunge at the end-of-life. To address this critical bottleneck, this paper proposes a physics-aware asymmetric decoupling network (PAAD-Net) with adaptive physical constraints. Specifically, the model utilizes causal convolutions and asymmetric cross-attention to decouple the thermodynamic-equivalent and kinetic-related health indicators, which are then processed by a specialized capacity decoder to ensure achieve prediction. Most importantly, this paper designs an adaptive physics-informed gradient constraint by leveraging the gating signal to dynamically modulate the weight of the physical loss. Extensive comparative experiments on CALCE and Oxford datasets demonstrate that PAAD-Net achieves low mean absolute error of 0.0070 Ah and 0.0011Ah, respectively. Experimental results verify that PAAD-Net achieves high-precision prediction throughout the full lifecycle, with comparative and ablation experiments further validating its accuracy and generalization.
- 11 August 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (32) , 325403
A planar electrical capacitance tomography method based on the Bayesian linear regression with Cauchy prior for defect detection
Planar electrical capacitance tomography is a technology for visualizing permittivity distribution and has broad prospects in nondestructive testing. Image reconstruction is the core of the technology, but it is often an underdetermined problem and is sensitive to noise, which are essential factors that limit the improvement of image reconstruction quality. Therefore, an image reconstruction algorithm based on Bayesian linear regression with Cauchy prior is proposed. Firstly, the prior permittivity distribution is constructed based on the Cauchy distribution. Secondly, the Laplace approximation is used to obtain the local Gaussian approximation of the posterior distribution of permittivity. Finally, the approximate expectation maximization algorithm is used to automatically update the hyperparameters of the posterior distribution to achieve image reconstruction. Experimental results show that compared with the Cauchy–Gaussian method, sparse Bayesian learning method, Tikhonov algorithm, and Landweber algorithm, this method has the best correlation coefficient and relative error, resulting in the most realistic images reconstructed, with the fewest artifacts and the highest defect detection accuracy.
- 11 August 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (32) , 325403
A planar electrical capacitance tomography method based on the Bayesian linear regression with Cauchy prior for defect detection
Planar electrical capacitance tomography is a technology for visualizing permittivity distribution and has broad prospects in nondestructive testing. Image reconstruction is the core of the technology, but it is often an underdetermined problem and is sensitive to noise, which are essential factors that limit the improvement of image reconstruction quality. Therefore, an image reconstruction algorithm based on Bayesian linear regression with Cauchy prior is proposed. Firstly, the prior permittivity distribution is constructed based on the Cauchy distribution. Secondly, the Laplace approximation is used to obtain the local Gaussian approximation of the posterior distribution of permittivity. Finally, the approximate expectation maximization algorithm is used to automatically update the hyperparameters of the posterior distribution to achieve image reconstruction. Experimental results show that compared with the Cauchy–Gaussian method, sparse Bayesian learning method, Tikhonov algorithm, and Landweber algorithm, this method has the best correlation coefficient and relative error, resulting in the most realistic images reconstructed, with the fewest artifacts and the highest defect detection accuracy.
- 1 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. PP (1530437X) , 1
Sparse-Prior-Assisted Dual-Encoder Multi-Scale Fusion Network for Acoustic Tomography Temperature Reconstruction
Acoustic tomography (AT) is a non-contact, radiation-free method well suited for measurement of temperature fields in complex industrial environments. To address the ill-posed inverse problem encountered by AT, a temperature reconstruction method that integrates compressed sensing and deep learning is presented to improve reconstruction accuracy and stability. First, the Smoothed lp-over-lq (SPOQ) regularization term is incorporated into the sparse reconstruction model for AT temperature measurement, leading to a compressed sensing–based initial reconstruction approach. On this basis, a Dual-Encoder Multi-Scale Fusion Network (DEMF-Net) is developed. The CNN encoder is used to extract multi-scale features from the TOF measurements. The Transformer encoder captures global dependencies in the initial temperature distribution. The extracted features are then fused for temperature reconstruction. Simulation results demonstrate that the proposed method achieves an average relative error of 0.1% and a root mean square error of 0.15% for single-peak, double-peak, and four-peak temperature distributions. The results demonstrate that the proposed method achieves improved reconstruction accuracy and anti-noise performance. The static experiments further verify the practical potential of the proposed method.
- 1 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. PP (1530437X) , 1
Sparse-Prior-Assisted Dual-Encoder Multi-Scale Fusion Network for Acoustic Tomography Temperature Reconstruction
Acoustic tomography (AT) is a non-contact, radiation-free method well suited for measurement of temperature fields in complex industrial environments. To address the ill-posed inverse problem encountered by AT, a temperature reconstruction method that integrates compressed sensing and deep learning is presented to improve reconstruction accuracy and stability. First, the Smoothed lp-over-lq (SPOQ) regularization term is incorporated into the sparse reconstruction model for AT temperature measurement, leading to a compressed sensing–based initial reconstruction approach. On this basis, a Dual-Encoder Multi-Scale Fusion Network (DEMF-Net) is developed. The CNN encoder is used to extract multi-scale features from the TOF measurements. The Transformer encoder captures global dependencies in the initial temperature distribution. The extracted features are then fused for temperature reconstruction. Simulation results demonstrate that the proposed method achieves an average relative error of 0.1% and a root mean square error of 0.15% for single-peak, double-peak, and four-peak temperature distributions. The results demonstrate that the proposed method achieves improved reconstruction accuracy and anti-noise performance. The static experiments further verify the practical potential of the proposed method.
- 20 March 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (11) , 116108
CEO-GPNFormer: a multi-source feature fusion framework for SOH estimation of lithium-ion batteries
Accurate state-of-health (SOH) estimation is crucial for ensuring the safe and reliable operation of lithium-ion batteries (LIBs). However, conventional SOH estimation methods mainly rely on a small number of highly correlated features or isolated model architectures, which often fail to capture the multi-dimensional degradation characteristics and complex nonlinear behaviors of LIBs. Therefore, this paper proposes a multi-source feature fusion-based chaotic evolution-optimized gated parallel echo state network-Transformer (CEO-GPNFormer) for battery SOH estimation. Multi-dimensional indirect health indicators are first selected using dual correlation criteria. CEO-optimized variational mode decomposition is then applied to extract capacity-mode features, which are effectively fused with the selected indicators to construct high-quality inputs. These fused features are subsequently fed into the CEO-GPNFormer model to complete SOH estimation. Validation on two datasets and six benchmark models shows that the proposed approach achieves marked improvements in both within-dataset cross-validation and single-battery prediction tasks, yielding mean absolute error values below 0.007 and average root mean squared error values below 0.01 across both datasets. These results verify the superior accuracy, robustness, and generalization capability of the proposed method.
- 20 March 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (11) , 116108
CEO-GPNFormer: a multi-source feature fusion framework for SOH estimation of lithium-ion batteries
Accurate state-of-health (SOH) estimation is crucial for ensuring the safe and reliable operation of lithium-ion batteries (LIBs). However, conventional SOH estimation methods mainly rely on a small number of highly correlated features or isolated model architectures, which often fail to capture the multi-dimensional degradation characteristics and complex nonlinear behaviors of LIBs. Therefore, this paper proposes a multi-source feature fusion-based chaotic evolution-optimized gated parallel echo state network-Transformer (CEO-GPNFormer) for battery SOH estimation. Multi-dimensional indirect health indicators are first selected using dual correlation criteria. CEO-optimized variational mode decomposition is then applied to extract capacity-mode features, which are effectively fused with the selected indicators to construct high-quality inputs. These fused features are subsequently fed into the CEO-GPNFormer model to complete SOH estimation. Validation on two datasets and six benchmark models shows that the proposed approach achieves marked improvements in both within-dataset cross-validation and single-battery prediction tasks, yielding mean absolute error values below 0.007 and average root mean squared error values below 0.01 across both datasets. These results verify the superior accuracy, robustness, and generalization capability of the proposed method.
- 8 January 2026
- Published by Elsevier in Measurement
- Vol. 265
A physics-consistent image reconstruction framework for ECT via sparse priors and physics-guided deep learning
- 8 January 2026
- Published by Elsevier in Measurement
- Vol. 265
A physics-consistent image reconstruction framework for ECT via sparse priors and physics-guided deep learning
- 2 September 2025
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 36 (9) , 095301
Gas volume fraction measurement and flow pattern identification and analysis in gas–liquid two-phase flow using a multi-source cross-attentional task network
To address gas–liquid two-phase flow modeling, this study proposes a multi-source cross-attentional task network (MCATN) for simultaneous gas volume fraction (GVF) measurement and flow pattern identification. The model features a parallel convolutional neural network-Transformer structure for joint extraction of local and global features, improving GVF prediction accuracy. A cross-attention mechanism is employed to fuse multi-source features with adaptive weighting, enhancing sensitivity to critical patterns. Furthermore, multivariate multiscale symbolic entropy (mvMSE) is introduced to extract dynamic complexity features, aiding in the discrimination of subtle flow pattern differences. Experimental results show that MCATN achieves superior performance, with a GVF measurement mean absolute error of 0.0015 and R 2 of 0.9999, and a flow pattern identification accuracy of 97.65%. Beyond classification and regression tasks, mvMSE also enables flow characteristics analysis by capturing temporal structural variations across multiple time scales, offering deeper insight into the complexity evolution of gas–liquid two-phase flow under varying operating conditions. Overall, MCATN demonstrates strong generalization and robustness in structure and multi-task optimization, providing a comprehensive and interpretable solution for complex flow behavior modeling in marine pipeline systems.
- 2 September 2025
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 36 (9) , 095301
Gas volume fraction measurement and flow pattern identification and analysis in gas–liquid two-phase flow using a multi-source cross-attentional task network
To address gas–liquid two-phase flow modeling, this study proposes a multi-source cross-attentional task network (MCATN) for simultaneous gas volume fraction (GVF) measurement and flow pattern identification. The model features a parallel convolutional neural network-Transformer structure for joint extraction of local and global features, improving GVF prediction accuracy. A cross-attention mechanism is employed to fuse multi-source features with adaptive weighting, enhancing sensitivity to critical patterns. Furthermore, multivariate multiscale symbolic entropy (mvMSE) is introduced to extract dynamic complexity features, aiding in the discrimination of subtle flow pattern differences. Experimental results show that MCATN achieves superior performance, with a GVF measurement mean absolute error of 0.0015 and R 2 of 0.9999, and a flow pattern identification accuracy of 97.65%. Beyond classification and regression tasks, mvMSE also enables flow characteristics analysis by capturing temporal structural variations across multiple time scales, offering deeper insight into the complexity evolution of gas–liquid two-phase flow under varying operating conditions. Overall, MCATN demonstrates strong generalization and robustness in structure and multi-task optimization, providing a comprehensive and interpretable solution for complex flow behavior modeling in marine pipeline systems.