Aiming at the problems that the fault diagnosis models suffer from poor cross-domain diagnostic capability due to the scarcity of fault annotation data and significant feature discrepancies across domains under varying loads and control modes in thermal power plants

and that traditional domain adapttation methods are mostly confined to single target domain migration

this paper proposes a cross domain fault diagnosis method for boiler steam-water system based on mixture of experts (MoE) and multi-target domain adversarial migration network. First

this paper establishes a feature extractor to extract the potential features of fault data using bidirectional gated recurrent unit (BiGRU) network with soft attention mechanism which helps BiGRU to focus on key features. Then

this paper builds a multi-expert network to reprocess the output features of the feature extractor

which are dynamically weighted by the gated network and then output to the label classifier and domain discriminator. Finally

a separate domain discriminator is set for each target domain

and the domain invariant features of the source domain and multiple target domains are jointly extracted by introducing the confrontation training mechanism through the gradient inversion layer and cooperating with the maximum mean discrepency (MMD). At the same time

the feature extractor

MoE and label classifier are continuously trained by using the source domain annotation data. The fault diagnosis experiment of boiler steam water system using a 600 MW thermal power plant simulator shows that this method can effectively migrate multiple target domains at the same time

and the average diagnostic accuracy of domains in each migration task is more than 98%.