Nianfeng Shi

Deep-unrolled inverse-scattering networks combine physics-based iterations with learned corrections, yet stability of the trained composite is not implied by solver convergence. We analyse the layers of a Born/contrast-source formulation and turn that gap into a computable test. The layer Jacobian contains an explicit contrast-to-field coupling block, giving a state-dependent sufficient condition for non-contraction: above the threshold α_c no correction confined to the contrast channel makes the evaluated layer contractive. Under β₀< 1 the field subsystem contracts, and a nonexpansive contrast projection with a positive damping floor gives the rollout an invariant bounded set; a finite-depth perturbation bound follows without assuming contraction of the composite. The weak-scattering constant β₀ follows from the acquisition operator and the contrast bound, whereas the threshold α_c depends on the field state at which the layer is evaluated; both are computable before training for a reference configuration. On synthetic data β₀=0.8303< 1 while the measured spectral norm of the full layer exceeds one; experiments test the bound, the budget constraints, depth and noise. On the published Institut Fresnel geometry the test returns β₀>1, so no a-priori certificate is claimed there, and the measured reconstruction is a finite-depth evaluation with untransferred weights.

Methane is not only the primary cause of gas explosions in coal mines but also a potent greenhouse gas. Accurate monitoring of its concentration is of great significance for ensuring production safety and maintaining ecological stability. In this paper, based on the Vernier effect, an ultra-sensitive photonic crystal fiber (PCF) methane sensor utilizing the two-loop parallel configuration is proposed. Two Sagnac loops function as the sensing and reference arms, with each containing a different length of methane-sensitive film-coated PCF. A combined design of microholes and semi-elliptical grooves is employed in the PCF, which not only achieves a high birefringence effect but also overcomes the difficulty of depositing the methane-sensitive film in the air holes. A theoretical analysis was carried out on the sensing performance of the proposed PCF methane sensor using the finite element method. Simulation results showed that the sensor achieved a maximum sensitivity of up to 952 nm/% within the methane concentration range of 0.00%–1.50%, corresponding to an 11.33-fold enhancement over the single-loop configuration. In addition, the PCF had a fabrication tolerance of ±1° for the semi-elliptical grooves. As far as we know, the proposed PCF methane sensor exhibits the highest sensitivity compared with other existing methane sensors. Owing to its merits of ultra-high sensitivity, simple structure, and high fabrication tolerance, the proposed methane sensor exhibits considerable potential for application in coal mine gas monitoring.

For practical industrial production scenarios where collected vibration signals are easily interfered by environmental noise and bearing operating conditions are complex and variable, this paper proposes a Double Branch Lightweight Convolutional Neural Network (DBLCNN). The model adopts a dual-branch architecture: the one-dimensional branch enhances feature extraction capability under low signal-to-noise ratio conditions, while the two-dimensional branch improves feature representation while significantly reducing the number of parameters. The complementary fault features extracted by the dual branches effectively enhance the accuracy of fault diagnosis. Under varying operating conditions, the model achieves an average accuracy of 95.58%; with the addition of 0 dB Gaussian white noise, its average accuracy under varying conditions remains at 90.17%. This study demonstrates that, even based on raw vibration signals without cumbersome preprocessing, the model can achieve excellent diagnostic performance in noisy environments and under variable operating conditions.

Refractory bricks, with high-temperature resistance, are the core lining materials in all construction fields. Surface defect detection plays a crucial role in ensuring the quality of refractory bricks. By accurately identifying defects such as cracks and pits, it effectively prevents equipment damage or production accidents caused by material failure in high-temperature environments. However, manual inspection is still relied upon to identify surface defects in firebricks. It suffers from low detection accuracy and high costs, while also being heavily influenced by human experience and subjective factors, failing to meet modern demands. This paper proposes a surface defect detection model for refractory bricks named Crack-Net, enabling efficient and precise identification of defects in refractory bricks. First, a specialized dataset for training and verification is constructed through field photography and image preprocessing. Second, we propose the CFNeXt Block for feature extraction and a defect feature amplification module for feature refinement, enabling accurate and efficient classification of surface defects in refractory materials. Finally, the model is verified on the self-built dataset, and the adverse effects of different defects on the industrial recognition algorithm were analyzed. The experimental results show that the mAP of the four defects reaches 81.61%, and the Latency is 2.52ms, which is better than the current mainstream detection models. Crack-Net introduces a novel intelligent inspection method for building material quality assessment, driving a paradigm shift in inspection standards from traditional empirical judgment to data-driven approaches. This innovation provides a quantifiable technological foundation for the high-quality development of the materials industry.

In order to realize the efficient exploitation of coalbed methane (CBM) in the southern Qinshui basin in China, raw coal specimens with original fractures of the No.3 coal reservoir in the Sihe mine were selected to conduct compression tests with different confining pressures. During this loading process, the synchronous and staged real-time CT scanning tests were carried out by a high-precision X-ray scanner. The mechanical properties, the failure mode, and the evolution of fractures of coal samples due to the increased effective stress were analyzed from the microscopic point of view. The result showed that the original fractures in the coal specimens of the Sihe coal mine are relatively developed. Four mechanical stagescompaction, elastic, yield, and plastic failurewere consistently identified; image-based segmentation allowed real-time, precise quantification of crack initiation and growth. Within the tested confining pressure range, further confirmation through real experiments that compressive strength, peak strain, and elastic modulus rose with confinement, while residual strength was preserved by frictional resistance along fracture surfaces. New fractures tend to start cracking near the original fractures. Crack networks were more intricate under lower confinement (uniaxial tensile failure) and simplified yet more distributed under higher confinement (triaxial shear failure), indicating a transition from brittle to semibrittle/semiductile behavior. Quantitative CT metricsgray value, fracture volume, and fracture porosityproved reliable indicators of damage progression. Porosity is below 1% at the compaction and elasticity stage, twice the size of the original fractures at the initial damage stage. The integrated experimental–visualization workflow elucidates fracture propagation mechanisms in anthracite and offers theoretical guidance for CBM stimulation and ground-control strategies in the Qinshui Basin.

Intrusion detection aims to identify the unauthorized activities within computer networks or systems by classifying events into normal or abnormal categories. As modern scenarios often involve multi-source data, multi-view fusion deep learning methods are employed to leverage diverse viewpoints for enhancing security threat detection. This paper introduces a novel intrusion detection approach using multi-view fusion within a federated learning framework, proposing an integrated AE Neural SVM (AE-NSVM) model that combines auto-encoder (AE) multi-view feature extraction and Support Vector Machine (SVM) classification. This approach simultaneously learns representative features from multiple views and classifies network samples into normal or seven attack categories while employing federated learning across clients to ensure adaptability and robustness in diverse network environments. The experimental results obtained from two benchmark datasets validate its superiority: on TON_IoT, the CAE-NSVM model achieves a highest F1-measure of 0.792 (1.4% higher than traditional pipeline systems); on UNSW-NB15, it delivers an F1-score of 0.829 with a 73% reduced training time and an 89% faster inference compared to baseline models. These results demonstrate the advantages of multi-view fusion in federated learning for balancing accuracy and efficiency in distributed intrusion detection systems.

To address the challenge of reduced prediction accuracy caused by capacity regeneration during the use of lithium-ion batteries, this study proposes an RUL (remaining useful life) prediction method based on mode decomposition and an enhanced Informer-LSTM hybrid model. The capacity is selected as the health indicator, and the CEEMDAN (complete ensemble empirical mode decomposition with adaptive noise) algorithm is employed to decompose the capacity sequence into high-frequency and low-frequency components. The high-frequency components are further decomposed and predicted using the Informer model, while the low-frequency components are predicted with an LSTM (long short-term memory) network. Pearson correlation coefficients between each component and the original sequence are calculated to determine fusion weights. The final RUL prediction is obtained through weighted integration of the individual predictions. Experimental validation on publicly available NASA and CALCE (Center for Advanced Life Cycle Engineering) battery datasets demonstrates that the proposed method achieves an average fitting accuracy of approximately 99%, with MAE (mean absolute error) below 0.02. Additionally, both MAPE (mean absolute percentage error) and RMSE (root-mean-square error) remain at low levels, indicating improvements in prediction precision.

To address the limitations of the traditional methods that are used to extract features from non-stationary signals and capture temporal dependency relationships, a rolling bearing fault diagnosis method combining variational mode decomposition (VMD) and deep learning is proposed. A hybrid VMD-CNN-Transformer model is constructed, where VMD is used to adaptively decompose bearing vibration signals into multiple intrinsic mode functions (IMFs). The convolutional neural network (CNN) captures the local features of each modal time series, while the multi-head self-attention mechanism of the Transformer captures the global dependencies of each mode, enabling the global analysis and fusion of features from each mode. Finally, a fully connected layer is used to classify the 10 fault types. The experimental results on the Case Western Reserve University bearing dataset demonstrate that the model achieves a fault diagnosis accuracy of 99.48%, which is significantly higher than that of single or traditional combined methods, providing a new technical path for the intelligent diagnosis of rolling bearing faults.