The mechanical vibration defect signals of gas insulated switchgear (GIS) equipment are highly complex. Addressing the prominent issue of limited feature extraction and low identification accuracy of mechanical defects based on vibration signal analysis due to dynamic changes in load current

this paper firstly uses the Markov transition field

spectral Markov transition field

and short-time Fourier transform method to transform the one-dimensional time series vibration signals into time domain

frequency domain

and time-frequency domain vibration graphs in turn

and the feature information of these three vibration graphs were fused using the multi-graph fusion method

which reduces the complexity of data processing effectively; then

the GoogLeNet is improved by using the adaptive batch normalization algorithm to construct the load current adaptive diagnosis model for GIS mechanical vibration defects with the fusion graphs as input. The results show that the improved model has a high diagnostic accuracy for typical mechanical vibration defects under variable load currents

with an average diagnostic accuracy of 91.21%

which is 28.14% higher than that of the unimproved GoogLeNet model

providing a valuable reference for the diagnosis of GIS equipment mechanical vibration defects under the dynamic change of load currents in the field.