To integrate multi-sensor and multimodal partial discharge data and improve the diagnostic accuracy of small occasional defects

this paper proposes a multi-perception dual-order fusion neural network (MPDOFNN) for gas insulated switchgear (GIS) insulation defect diagnosis. MPDOFNN is built based on multimodal neural network (MMNN) and embedded in a dual-order fusion framework. The first-order fusion is used to fuse the isomorphic information from various sensors. Through a multi attention fusion module

the information contribution of each channel is adaptively calculated

and the isomorphic information of each channel is transformed into a lossless variable-weight fusion based on the information weight. On the basis of the first-order fusion

the second-order fusion utilizes Maxout concatenated layers to perform heterogeneous information fusion of phase resolved partial discharge (PRPD)

phase resolved pulse sequence (PRPS)

and empirical features. This concatenated layer only retains the advantageous features of data from each modality through information competition

thereby filtering out redundant information from the same source and heterogeneous data of partial discharge

and reducing the scale and difficulty of model training. The experimental results show that MPDOFNN can more effectively integrate multi-sensor and multimodal data of partial discharge

and its diagnostic performance is superior to other diagnostic methods with single sample counts of 50

100

500

and 1 000 pulses. It is more suitable for on-site detection conditions with fewer partial discharge pulses and stronger noise interference.