Research On Machine Learning Based Acoustic Anomaly Detection Method For Power Transformers

Ensuring the normal operation of transformers is important for the uninterrupted and reliable operation of the grid.Acoustic online monitoring is a simple and reliable non-invasive monitoring method that does not interfere with the normal operation of the transformer and can reflect the working status and abnormalities of the transformer.The rapid development of communication,storage and computing technologies offers the possibility of acoustic online monitoring of transformers,but the large number of transformers,their complex structure and the variety of fault types make it difficult to reason directly about the health of a transformer based on sound.To this end,this study uses machine learning methods to detect acoustic anomalies in transformers and provides a corresponding solution to the problems of sparse anomalous transformer sounds,complex environmental noise,large data size and dimensionality,and high real-time requirements.The work in this study consists of the following.1)This study compares and analyses 21 traditional machine learning anomaly detection algorithms and four advanced deep learning detection algorithms,gives the principles and calculation methods of each type of algorithm,and compares and analyses the performance of each algorithm under different training modes based on actual measured transformer sound data,to provide reference for algorithm selection under different scenarios during practical application.2)The characteristics of sound data from different transformers and different operating conditions are analysed,and the use of linear frequency cepstrum coefficients for sound feature extraction is proposed in conjunction with the sound generation mechanism of transformers.3)To address the problem of transformer anomaly sound sparsity,this study proposes the use of unsupervised and semi-supervised anomaly monitoring methods.Unsupervised learning for anomaly detection does not require any labels,while semi-supervised learning only needs to ensure that the training set is all normal data,which fits the characteristics of redundant normal data and sparse anomalous data for transformer sounds.The experimental results validate the effectiveness of both methods.4)To address the problems of large data size and dimensionality and high real-time requirements when monitoring transformers online,this study proposes the use of a deep learning-based anomaly monitoring method.Experimental results show that the use of single-objective generative adversarial active learning and multiple-objective generative adversarial active learning methods have better detection effects and speed,and meet the requirements of high real-time performance under large-scale data.5)To address the problem of complex environmental noise,this study proposes an anomaly monitoring scheme for high noise environments.The initial model is obtained by combining unsupervised and semi-supervised learning and used for online monitoring,while the normal sound data set is continuously expanded for fine-tuning the training neural network,and the decision boundary is modified by the anomaly data until the model can correctly identify the operating conditions under various types of noise.Measured noise-containing data validate the effectiveness of the scheme.