Layered multi-fork network structure efficient search method for rotating machine fault diagnosis

The invention discloses a layered multi-fork network structure efficient search method for rotating machinery fault diagnosis. A controller is created by using a long-short-term memory network to search and determine two types of meta-structures, and the meta-structures are stacked into a sub-model according to a definition. The meta-structure is composed of five nodes including two input nodes and three nodes to be searched, the nodes to be searched further comprise a left child node and a right child node, and the input and operation types of the child nodes are to be searched and determined. The sub-model is of a layered structure formed by stacking element structures, the search freedom degree can be controlled, the search efficiency is improved, and the migration performance of the sub-model is supported; and compared with a chain structure, the multi-fork structure of the nodes in the meta-structure enables the sub-model to extract deeper features from original data, and is of great significance for improving the diagnostic performance of the sub-model. When the method is used, a to-be-searched quantity of a meta-structure serves as input, a label determined by sub-model verification precision serves as output to train a controller, the controller designs two types of meta-structures, a complete sub-model is created according to stacking definition and rules, verification is carried out on a target data set, training of the controller and training of the sub-model are carried out alternately, and the target data set is obtained, and finally, the controller capable of designing a sub-model with good diagnosis performance and migration performance is obtained. By means of the characteristics, the method can automatically create a high-precision diagnosis model when facing different rotary machine diagnosis tasks, and other diagnosis tasks can be migrated and completed.