基于频域变换和Transformer的动检数据异常分类算法研究

Railway dynamic inspection data is the multi-dimensional time-series data collec-\\ted periodically by high-speed railway comprehensive inspection vehicle to monitor the track-\\status. Due to the influence of collection equipment, climate and other complex factors, there are various data anomalies in the collected dynamic inspection data, and even the manual de-\\termination of anomaly categories is laborious and erroneous. The data anomalies affects the subsequent data mining and utilisation. Starting from the time-frequency correlation between the attributes of multi-dimensional motion detection data, this paper proposes an anomaly cl-\\assification model AFFT-Dformer that combines the adaptive frequency domain transform m-\\odule and the double Transformer encoder. The Adaptive Fast Fourier Transform (AFFT) m-\\odule removes the high-frequency noise and enhance the data representation. The dual Trans-\\former encoders module captures data channel and timing dimension features. Experiment on the field motion detection dataset show that the AFFT-Dformer can classify five types of abn-\\ormal data with a classification accuracy of 92\%. Meanwhile, we also conducted experiments on the four public datasets to validate the generalisability of the proposed AFFT-Dformer for time-series data classification.