[Objective] With the increasing integration of intermittent renewable energy sources into the power grid, large-capacity thermal power units are frequently required to participate in deep peak -shaving operations. This operational transition leads to frequent and drastic c hanges in boiler load, which significantly intensifies the thermal stress on heating surfaces and increases the risk of boiler "four-tube" leakages. However, early-stage weak leakage diagnosis faces two critical challenges: feature distribution drift caused by variable-load conditions and an extreme scarcity of fault samples in industrial sites. Traditional deep learning-based diagnostic models, which rely on fixed classification heads, often suffer from feature manifold collapse and severe overfitting when transferred from base -load to deep peak -shaving scenarios. To address these issues, this paper proposes a cross -condition few-shot fault diagnosis framework named Proto-MixMamba, which integrates time-series multi-scale decoupling with prototypical metric learning. 

[Methods] The research utilizes high-precision simulation data from a 600 MW supercritical unit to construct comprehensive datasets covering various load conditions and leakage severities. To move beyond traditional single -scale sequence modelin g and parameter -solidified classification paradigms, a dual-branch backbone network comprising a One -Dimensional Position Attention Module (1D -PAM) and a Mamba State Space Model is established. Specifically, the 1D -PAM branch is designed to capture local w eak transient shocks and high -frequency fluctuations in thermal parameters, while the Mamba branch leverages its selective scanning mechanism to model global low -frequency evolution trends and long -term dependencies with linear computational complexity. Th ese multi -scale features are adaptively integrated through a gating fusion mechanism that balances the importance of local and global information. To resolve the feature confusion caused by amplitude baseline shifts during load variations, the framework in troduces an N -way K -shot episodic meta-learning strategy. A prototypical metric layer is employed to replace the conventional fully connected classifier. By calculating the relative Euclidean distance between query samples and class prototypes in a high-dimensional embedding space, the model constructs a robust decision boundary that is invariant to absolute parameter magnitudes. 

[Results] Extensive experimental evaluations demonstrate the superiority of the Proto-MixMamba model. The results indicate that t he proposed method effectively overcomes the long -standing industrial challenge of misjudging physically symmetrical leakage locations, such as the left and right economizers, which exhibit highly similar thermal responses under variable -load interference. In the challenging cross-condition diagnosis task transitioning from 600 MW to 480 MW (deep peak -shaving), and under the extreme constraint of the 1 -shot setting (only one reference sample per category), the proposed model achieves a diagnostic accuracy o f 88.4%. This performance significantly outperforms traditional baseline models such as Long Short-Term Memory (LSTM) and ResNet, whose accuracies drop below 50% due to severe overfitting and domain shift. Visualization via t -SNE confirms that the prototyp ical metric mechanism successfully maintains high intra -class compactness and inter -class separability in the target domain.

[Conclusion] The proposed Proto-MixMamba framework demonstrates exceptional robustness and generalization c apabilities in cross-condition few-shot scenarios. By successfully extracting invariant fault features and establishing a reliable metric space, this research provides a technically feasible and highly effective route for the accurate early warning and diagnosis of boiler "four-tube" leakages in label-scarce and operationally complex industrial environments.