Faye Zhang

Open-set cross-condition fault diagnosis requires learning models to mitigate operating-condition shifts while rejecting previously unseen faults. However, conventional domain-adaptation methods may force unknown target samples into predefined classes and align them with known distributions, resulting in negative transfer. To address these limitations, this paper proposes a Physics-Constrained Data-Driven Dual-Gating Fusion Network (PCDGFCN) for power-transformer fault diagnosis. PCDGFCN couples physics-informed representation learning, unknown-fault estimation, and confidence-aware domain adaptation within a unified optimization framework. Vibration transmissibility, modal, harmonic, and energy-distribution characteristics are integrated to construct a fault-sensitive and physically interpretable representation space. Relatively stable harmonic and energy characteristics are further employed as soft physical anchors to preserve known-class semantics across voltage and load variations. Within the learned space, a cluster-conditioned extreme value theory gate quantifies local tail abnormality, while a prototype gate evaluates incompatibility with known fault structures. Rather than combining the two criteria using fixed weights, their sample-wise contributions are adaptively coordinated according to optimization feedback and path reliability. The resulting unknown-fault score regulates target-domain alignment and physical-consistency regularization, establishing closed-loop interaction between unknown-fault rejection and representation adaptation. Experiments on a full-scale 110-kV oil-immersed transformer platform and a public dry-type transformer dataset demonstrate that PCDGFCN achieves an average macro-F1 score of 97.77% in cross-voltage tasks under loaded conditions and consistently outperforms the baseline methods.

Scanning laser Doppler vibrometers (SLDVs) are indispensable in non-contact vibration measurement. However, the inherent lack of 3-D geometric perception in native SLDV systems severely constrains multi-head co-alignment and the spatial co-registration of 3-D operational deflection shapes (ODS). Existing 3-D surface measurement schemes often rely on structured light projection or line laser sources, which are inherently incompatible with native SLDV hardware architectures and vibration measurement workflows. To address this limitation, a 3-D geometrically aware scanning framework is proposed for conventional SLDVs. Under the constraint of zero-hardware-intrusion, a unified analytical geometric model of the galvanometer--camera system is established by utilizing the built-in dual-axis galvanometer, the single-point laser beam, and a built-in 2-D camera. Consequently, full-field absolute 3-D coordinate mapping is achieved via a single-plane 2-D calibration method. Furthermore, a decoupled simulation-based error propagation analysis is conducted to isolate and quantify the respective impacts of optomechanical assembly tolerances and visual calibration parameter deviations on the reconstruction performance. Experimental validations demonstrate that the proposed framework robustly achieves sub-millimeter surface reconstruction accuracy, effectively extending the measurement capabilities of native SLDVs into the real 3-D physical space. Thus, a highly practical technical path is provided for the efficient multi-head co-alignment of vibrometer probes and the post-processing-free, high-fidelity spatial co-registration of ODS on complex curved surfaces.

Reliable rolling bearing fault diagnosis under unknown degradation remains challenging because fault-related vibration signatures can be weakened by background disturbances, impulsive interference, colored noise, periodic interference, and operating-condition variation. Many noise-robust models rely on fixed feature fusion or assume that signal-quality information, such as the true signal-to-noise ratio (SNR), is available during inference. This paper proposes blind reliability-guided physics-aware time–frequency network (BRP-TFNet), a blind-at-inference reliability-guided physics-aware time–frequency network for noise-robust rolling bearing fault diagnosis. BRP-TFNet learns degradation-related information directly from the observed vibration signal without using the true SNR as an inference input. It extracts waveform-based temporal evidence and physics-guided spectral evidence around characteristic fault-frequency harmonics, including ball-pass frequency of the inner race, ball-pass frequency of the outer race, and ball spin frequency, and combines the two branch predictions through sample-dependent reliability fusion. A unified multi-label target [ IF , OF , BF ] is adopted to represent healthy, single-fault, and compound-fault conditions. Experiments on the Shandong University (SDU) and Case Western Reserve University (CWRU) bearing datasets under Gaussian, impulsive, colored, periodic, and mixed degradation demonstrate competitive robustness under moderate noise and several structured degradation conditions. Under the raw-file-level SDU split, BRP-TFNet obtains the highest mean Report- F 1 under clean, 0 dB additive white Gaussian noise (AWGN), and − 5 dB AWGN conditions among the reproduced baselines. Combined-source supervised training improves robustness across all reported SDU conditions, including an increase in Report- F 1 from 0.540 to 0.646 under − 5 dB AWGN. Its effect on CWRU is condition dependent, with improvements under several structured degradation conditions but reduced performance under severe AWGN.

High-frequency ultrasonic testing typically requires high sampling rates and dense spatial scanning, resulting in massive volumes of data and imposing significant burdens on data storage and transmission. Compressive Sensing (CS) provides an effective framework for reducing sampling data while enabling high-quality signal reconstruction. However, conventional CS reconstruction methods still exhibit limitations in modeling complex high-frequency ultrasonic echo structures and in reconstruction efficiency. To address these challenges, this paper proposes a high-frequency ultrasonic compressive sensing reconstruction method for Scanning Acoustic Microscopy (SAM), termed SAM-TCS. Specifically, a learnable compression measurement layer is first employed to achieve signal acquisition in the compressed domain, followed by a linear reconstruction module to obtain an initial signal estimate. On this basis, a hybrid CNN-Transformer network is introduced for refined reconstruction. The convolutional neural network (CNN) is utilized to extract local temporal structural features of ultrasonic signals, while the Transformer module captures long-range dependencies through a self-attention mechanism, thereby enhancing the recovery of complex echo structures. Furthermore, a joint loss function combining time-domain reconstruction error and frequency-domain structural constraints is designed to preserve the narrowband spectral characteristics of high-frequency ultrasonic signals. Experimental results demonstrate that, compared with conventional CS methods and representative deep learning approaches, the proposed method achieves superior performance in both signal reconstruction accuracy and imaging quality. At an 8× compression ratio (sampling ratio of 12.5%), the proposed method can effectively recover high-frequency ultrasonic signal structures, yielding clearer representations in A-scan, B-scan, and C-scan imaging results.

Scanning Acoustic Microscopy (SAM) has been widely applied in semiconductor packaging and advanced electronic manufacturing. However, small scanning steps can obtain high-resolution images but significantly reduce detection efficiency; conversely, large scanning steps improve scanning speed but may lead to the loss of image details and degradation of spatial resolution. To achieve high efficiency and precision simultaneously in SAM inspection, this paper proposes a Depth-Aware Super-Resolution Reconstruction Network (DASR-Net) for multi-depth slice imaging in SAM. First, construct a paired dataset of high- and low-resolution multi-depth SAM slices acquired with scanning steps of 20μm and 50μm. Subsequently, a network architecture is designed, integrating Global Encoder, Depth-Aware Encoder, Cross Depth Attention, Feature Fusion, and Reconstruct Decoder, to jointly extract global spatial features and cross-depth structural information. Meanwhile, a joint optimization strategy incorporating reconstruction loss, gradient loss, depth consistency loss, and Laplacian edge loss is introduced to enhance the recovery capability of image texture details and structural information. Experimental results demonstrate that DASR-Net achieves a PSNR of 30.49dB and an SSIM of 0.9325 on the test set, improving by 2.71dB and 8.42%, respectively, compared with the best-performing comparison method. RMSE, MAE, and LPIPS are reduced by 25.11%, 24.73%, and 36.51%, respectively. In experiments involving real plastic-encapsulated chips, the average PSNR and SSIM reach 29.44dB and 0.9163, respectively, and LPIPS is reduced by 35.14%, demonstrating the model's ability to recover detailed information of complex packaging structures. Furthermore, the proposed model contains only 4.79M parameters, reducing the parameter size by 35.99% compared with the best-performing comparison method. The forward inference time is only 7.15ms, achieving a 30.24% increase in speed. Ablation experiments further verify the effectiveness of each network module and the joint loss function. The proposed method can significantly improve SAM image quality while maintaining high detection efficiency, providing an efficient and lightweight super-resolution reconstruction solution for scanning acoustic microscopy in industrial inspection.

Rolling bearing fault classification degrades under severe noise because fault-related transients are weakened and class separability is reduced. To address this problem, this paper proposes an signal-to-noise ratio (SNR)-aware time-frequency dual-branch framework for controlled noise-robust rolling bearing fault classification. The model combines a temporal branch that learns from raw vibration waveforms with a spectral branch that learns complementary information from the fast Fourier transform log-magnitude spectrum. To adapt the temporal representation to different degradation levels, feature-wise linear modulation (FiLM) is introduced and conditioned on an SNR-related input. In addition, a symmetric Kullback–Leibler prediction-consistency term is used during training to encourage agreement between the two branches, while the final decision is obtained through lightweight logit-level fusion. Experiments on the Case Western Reserve University and Shandong University bearing datasets under additive white Gaussian noise, impulsive noise, and supplementary 1/f, harmonic, and non-stationary disturbance models show that the proposed framework maintains strong in-dataset classification performance from clean to severely degraded conditions within the adopted controlled evaluation protocol. Ablation results indicate that FiLM-based SNR conditioning provides the clearest benefit under severe noise, while the dual-branch design improves robustness relative to relying only on the temporal branch. These results support the proposed framework as an effective approach for controlled in-dataset noise-robust bearing fault classification, while broader deployment settings such as estimated-SNR conditioning and cross-dataset transfer remain for future study.