基于SKInformer的列车部件加速度时序预测OA

The nonlinear and nonstationary characteristics of vibration acceleration time series from train components degrade the predictive accuracy of the Informer model.To address this issue,this study proposes a Selective Kernel Mechanism Informer(SKInformer)model that integrates the Selec-tive Kernel mechanism(SKmechanism)into the Informer architecture for predicting vibration accelera-tion time series of train components.First,sequence-dimension transformation converts one-dimensional time series into the two-dimensional feature maps required by the SKmechanism,en-abling two-dimensional convolution to capture spatial correlations within the time series and enhance feature interactions across time steps.Second,the SKmechanism performs split,fusion,and selection operations on convolutional branches,allowing the network to adjust its receptive-field size dynami-cally according to the input and thereby improve its adaptability to multiscale features.Third,the SK-mechanism is embedded in the Informer encoder.Through dynamic receptive-field selection,the model simultaneously captures short-term abrupt variations and long-term trends in the time series.Fi-nally,the SKmechanism is incorporated into the Informer decoder to adaptively adjust the range of at-tention over the historical sequence and extract more discriminative local temporal representations.Ab-lation experiments on a pantograph dataset and comparative experiments on the LRV4313 light-rail ve-hicle dataset are conducted to validate the proposed method.Compared with the baseline Informer,SKInformer reduces the test-set loss by 34.18%at a learning rate of 0.000 1 in the ablation experi-ments and achieves a maximum coefficient of determination(R²=0.99)in the comparative experi-ments.These results demonstrate that the proposed method provides greater accuracy and robustness in predicting acceleration time series for train components.