Transformer fault early warning based on multi-dimensional representation and multi-view structure prediction model

Aiming at the existing transformer early warning method based on dissolved gas analysis in oil, which ignores the lack of relationship structure between multi-gas parameters, the lack of representation of local changes, and the lack of high computational complexity, a multi-dimensional representation and multi-view structure are proposed. Prediction model early warning method. First, the importance of each fault type is represented by gray correlation analysis and measurement of multiple gases, and the weight of each gas parameter under each fault type is determined; secondly, a multi-view structure prediction model early warning method is established, which includes: inputting multiple The gas sequence sample constructs a parameter time graph by calculating the correlation between all gas parameters at different time steps; generates a parameter dimension graph through differential operation to extract effective information on the relative fluctuations of each parameter; and divides the multi-element gas sequence into local sequence and calculate the correlation coefficient that dynamically changes with time between neighboring segments to construct a cross graph; use a Parallel attention heads mechanism to convert and connect multi-view features to obtain the final prediction result. The minimum sum of squared weighted errors is the criterion to determine the category to which the sample belongs. Compared with other methods, this method can reduce the error rate by 17.41%. In addition, Through optimization, the memory consumption was reduced by 63.11%, and the execution time was also reduced by 77.12%.