In order to realize the fast and accurate detection of abnormal sound of transformer and reduce the inconvenience of manual acquisition and labeling of abnormal sound signals,an aligned auto-encoder model was proposed.Firstly,the different framing methods were applied to the training set and the test set.The training set used overlapping framing to facilitate the extraction of complete features,and the test set used non-overlapping framing to improve the detection speed;secondly,the mixed activation function was proposed to replace the single activation function,and Huber loss was used to replace the common mean square error,which reduced the number of iterations and increased the robustness of the model;finally,a sequence alignment method was proposed to construct an aligned auto-coder.The three-level clipping and correlation calculation were used to estimate the delay of the prediction sequence relative to the test sequence.The prediction sequence was aligned by the reverse delay compensation to improve the prediction accuracy.The test results show that the aligned auto-encoder improves the accuracy and speed of abnormal sound detection,and the detection of each scoring index reaches more than 95％,which can provide a good reference for engineering applications.
维普
万方数据