Maofa Wang

Petrographic thin-section image classification is important for geological education, petrographic interpretation, and digital geoscience workflows, but robust automation remains difficult when labelled examples are scarce, inter-class differences are subtle, and microscope acquisition conditions vary. Standard episodic meta-learning usually samples tasks uniformly and gives all sampled episodes equal influence during meta-optimization. This design is poorly matched to petrographic imagery, where the scientific value of an episode depends not only on classification loss but also on predictive ambiguity, PPL/XPL or acquisition-domain discrepancy, and geologically plausible confusion between lithologies. We propose Domain-Aware Top-k Weighted Meta-Learning (DA-TWML), a reproducible computational workflow that estimates episode value from query loss, predictive uncertainty, domain discrepancy, and prototype confusion.DA-TWML retains high-value episodes through a Top-𝑘 scheduler and aggregates them with a softmax-weighted outerloop update on a Prototypical Network backbone. The completed audit covers mixed-polarization 5-way 1-shot and 5-way 5-shot settings, multi-seed evaluation, and score-factor ablation on an audited 108-category thin-section dataset. In the 1-shot ablation, the strongest DA-TWML-style score variant, 𝐿 + 𝑈 + 𝐶, reaches 63.01% mean test accuracy and improves over loss-only scoring, indicating that uncertainty and prototype confusion provide useful information beyond query loss. In the 5-shot ablation, the 𝐿 + 𝐶 variant obtains the highest current mean, suggesting that prototype-level geological confusion remains informative when support prototypes are estimated from more examples. These results position DA-TWML as an interpretable and auditable task-valuation framework for petrographic few-shot learning.Beyond accuracy reporting, the workflow formalizes data curation, split control, episode logging,software packaging, and code-release requirements for reproducible geoscientific image analysis.

The safe and stable operation of coal-fired generating units is of critical strategic importance for ensuring the reliable supply of power systems. However, the fault evolution of industrial thermal systems exhibits the characteristics of strong nonlinearity and a long incubation period, coupled with the extreme scarcity of key fault samples (Few-shot) in actual production, which severely limits the engineering application of traditional data-driven diagnostic methods. Existing deep learning models, which are highly dependent on massive and balanced labeled data, not only struggle to overcome the overfitting bottleneck in scenarios with scarce fault samples, but also frequently introduce severe label noise (Label Noise) by ignoring the physical incubation period of faults, resulting in the degradation of the model’s decision boundary. To address the above challenges, this paper proposes a novel fault diagnosis framework integrating multi-scale spatiotemporal feature engineering and the Tabular Prior-Data Fitted Network (TabPFN). Starting from the physical mechanism of the system, this paper develops a dynamic label cleaning strategy based on multivariate statistical deviation, which accurately defines the fault divergence point to eliminate the noise in the incubation period. The constructed multi-scale spatiotemporal feature engineering integrating first-order difference and sliding window statistics can effectively map the transient mutation and steady-state evolution trend of the system. The introduced pre-trained TabPFN model based on the Transformer architecture, relying on its Bayesian inference capability and in-context learning (In-Context Learning) mechanism, can realize parameter-tuning-free and efficient classification for scarce samples. Experiments based on high-fidelity dynamic simulation data from GE Steam Power show that under the strict setting of limiting the training set to only 2000 samples, the proposed method achieves a comprehensive diagnostic accuracy of up to 99.29% and an F1-score of 0.9929 for seven typical operating conditions. Multi-dimensional comparative experiments and ablation studies confirm that the proposed framework comprehensively outperforms six mainstream baseline models, including XGBoost and SVM, in terms of precision, recall, and anti-interference robustness, and also delivers outstanding performance when benchmarked against deep learning models. This provides a brand-new theoretical perspective and technical paradigm for equipment health management in the context of industrial big data.

In recent years, UAV-based visible and thermal (RGB-T) multimodal object detection has demonstrated significant potential for monitoring forest fires in complex environments. However, restricted by the intrinsic discrepancies between modalities, existing RGB-T models still struggle to achieve optimal detection accuracy. To address the aforementioned issues, this paper proposes FireRGBTNet, a lightweight and efficient RGB-T fusion model for UAV-based forest fire detection. First, a heterogeneous dual-stream backbone is designed to precisely capture modality-specific information via customized feature extraction modules, utilizing Target-Enhanced Downsampling to adaptively preserve the fragile features of small targets. Second, a multi-scale semantic alignment enhancement branch is proposed to explicitly bridge the semantic gap through the dual constraints of statistical distribution and semantic direction. Finally, a multimodal spatial gated fusion module is constructed, which utilizes spatial gated interactions to effectively suppress multimodal noise and incorporates Mish Gated Linear Units to compensate for global semantics, thereby achieving high-quality heterogeneous feature fusion. Extensive experiments on the RGBT-3M dataset demonstrate that with a mere 4.13M parameters, FireRGBTNet achieves accuracies of 95.9% and 62.8% in terms of mAP@0.5 and mAP@0.5:0.95, respectively. The proposed model achieves an optimal trade-off between detection accuracy and computational efficiency, providing a highly effective solution for UAV-based forest fire monitoring in complex environments.

Reliable sensor fault detection is critical for the safe and efficient operation of complex industrial systems, such as thermal power plants. However, traditional data-driven methods and standard deep learning models often struggle to detect incipient gradual drift faults under severe environmental noise, primarily because they ignore the inherent physical correlations among multivariate sensor signals. To address this challenge, this paper proposes a novel Physics-Coupled Deep Long Short-Term Memory Autoencoder (PC-Deep-LSTM-AE). Specifically, we integrate a deep LSTM architecture with an explicit non-linear information compression bottleneck and layer normalization to enhance robust feature extraction in high-noise environments. Furthermore, we innovatively introduce a Physics-Coupling Loss (PCC Loss) that jointly optimizes the mean squared reconstruction error and the Pearson correlation coefficient, forcing the model to strictly preserve the dynamic physical relationships among multivariable signals. Extensive experiments were conducted on a real-world thermal power plant dataset with severe noise injection. The results demonstrate that the proposed PC-Deep-LSTM-AE achieves an outstanding F1-score of over 0.98, significantly outperforming mainstream baseline models, including Vanilla LSTM-AE, GRU-AE, Bi-LSTM-AE, and CNN-AE. The proposed method exhibits exceptional robustness and high interpretability for root-cause analysis, highlighting its immense potential for real-world industrial deployment.