A small-sample supervised learning diagnosis method for scroll compressors based on uncertainty-calibrated dynamic memory-aware graph convolutional network

Existing fault diagnosis methods typically rely on static-weight attention models, which often produce overconfident predictions. This issue is particularly pronounced in complex machinery, such as scroll compressors, where small-sample conditions exacerbate diagnostic biases caused by such overconfidence. Consequently, both the reliability and accuracy of diagnosis are diminished. To address this problem, this paper proposes a dynamic memory-aware mechanism. Specifically, the method integrates bidirectional Long Short-Term Memory (LSTM) with learnable positional encoding, enhancing the model’s ability to capture temporal sequences effectively. Additionally, an adaptive uncertainty calibration loss function (AdaUCal-Loss) is introduced, which optimises the relationship between predicted confidence and classification accuracy, thereby improving diagnostic robustness. Furthermore, a dynamic memory-aware graph convolutional network model, grounded in uncertainty calibration, is presented. This model incorporates spatial topology, temporal evolution, and probabilistic calibration, collectively enhancing classification accuracy-particularly for boundary-ambiguous samples under small-sample conditions. Experimental results demonstrate that the model achieves an impressive accuracy of 99.19% on a small-sample scroll compressor dataset, outperforming existing methods. Moreover, it shows strong generalisation and diagnostic performance on two publicly available datasets from Xi’an Jiaotong University (XJTU Spurgear and XJTU Gearbox), validating the effectiveness of the proposed approach.