Research On Zero Sample Fault Diagnosis Method Based On Surrogate-Assisted Many-objective Structure Optimization

With the rapid development of the Internet of Things and artificial intelligence,fault diagnosis technology has stepped into the data-driven era catalyzed by the massive data generated by mechanical equipment.Based on these data,mining,analysis,prediction and diagnosis of emerging technologies have gradually become the cornerstone to ensure the lasting and safe operation of machinery.As the core parts of today’s mechanical equipment,rolling bearing fault diagnosis research can greatly guarantee the healthy operation of mechanical equipment,and has practical engineering significance and theoretical research value.At present,the data-driven fault diagnosis method has a broad application prospect and development prospect,but it has not formed a complete theoretical framework and system,and there are many theoretical and technical problems to be solved.The problems are mainly manifested in: 1)It is difficult to obtain massive data of rare categories in practical engineering diagnosis,and it is difficult to apply to high-performance supervised fault diagnosis methods;2)There are limitations in single-source sensing information mining,low information richness in timing signal fault discrimination,and insufficient visual feature extraction;3)Bearing fault feature extraction model has large structure and parameter redundancy,strong dependence on designers’ professional domain knowledge,weak generalization performance,and time-consuming solution of multiple evaluation indexes in the process of parameter optimization.In view of the above problems and difficulties,the main research carried out in this paper is as follows:(1)Taking bearing fault diagnosis under the condition of few samples and zero samples as the research object,the coupling relationship between labeled fault categories and new unseen faults is intended to be established through embedding space at visual and semantic levels,and semantic vectors of known and unseen classes are extracted to build semantic space and learn the shared subspace between visual features and semantic attributes.The mapping function is used to obtain the visual feature prototype of unknown faults,and the zero small sample diagnosis model for complex/unseen faults is constructed finally,in order to realize the prediction and recognition of different types of unknown faults.The experimental results prove the necessity of using zero-sample diagnosis model in the diagnosis process and the effectiveness of the model compared with the classical fault diagnosis algorithm under the background of no training samples.(2)Study the transformation of noise reduction data sequences in different spatial domains under different sensing environments,establish a multi-channel cascade mechanism to realize multi-spatial domain information fusion,combine the advantages of each domain to highlight the channels sensitive to fault,and transform the detection clustering problem into a multi-domain two-dimensional data model processing problem;The signals collected by different measuring points were constructed into a signal set with multiple domain spatial attributes.The energy information distribution characteristics of each measuring point were observed.The multi-sensor information fusion was realized by setting weight reward and penalty strategy fusion without obviously increasing the complexity of the model.Finally,a multi-spatial reconstruction model of fault characteristics based on heterogeneous sensing is presented.The effectiveness of the proposed method is verified by comparison experiments on two bearing fault data sets.(3)Analyze the search space of the architecture components of the model to be optimized,determine the parameters to be optimized in the model and take them as decision variables,set the constraints and search range existing in each variable,and formulate evaluation indicators that can reflect the evaluation speed,fault sample evaluation accuracy,model computational complexity and other demand items.The problem of fault diagnosis model construction is transformed into a multi-objective optimization problem.At the same time,considering that it is necessary to solve the time-consuming problem of model evaluation index for several times in the case of limited computing resources,it is intended to calculate the cheap agent auxiliary model instead of calculating the complex evaluation index solution,and the multi-objective optimization algorithm with efficient optimization performance is used as the search strategy to search the optimal solution set of the agent model.In this way,the non-dominated solution set of balanced multiple evaluation indexes is provided.According to the actual engineering needs,the reliability component of the diagnostic model is optimized by selecting the appropriate Pareto optimal solution,and the requirements of the accuracy of the reconstructed model are finally balanced.Experimental results demonstrate the superiority of the proposed optimization strategy over other classical evolutionary algorithms and its effectiveness in dealing with practical problems such as convolutional neural network structure optimization.Based on the above research content,the multi-type fault intelligent recognition system of rolling bearing is designed and implemented,so as to realize the fault diagnosis of rolling bearing objectively,intelligently and efficiently.The research results of this paper can be widely used in the country’s modern industrialization construction,to meet the needs of national development,maintain social security and stability,intelligent manufacturing,intelligent detection and monitoring and other fields,with huge market prospects.