Research On Fault Diagnosis And Parameter Identification Method Based On Multi-dimensional Taylor Network

Neural networks have been widely used in various fields,but existing neural networks not only have high computational complexity but also make it difficult to explain the results.When the neural network node is an exponential non-linear function,it can also be reduced to a multi-dimensional Taylor high-order expansion polynomial for approximation,and it is expected to reduce the computational complexity and enhance the interpretability of the results.This kind of simulation is based on multi-dimensional Taylor network and develops fault diagnosis.The main work of the thesis is as follows.1)System variable importance evaluation and fault diagnosis based on single-layer multi-dimensional Taylor network.Firstly,the evaluation function of the importance of each variable is established based on a single-layer multidimensional Taylor network,and then based on the analysis of the average impact of each variable,or the partial derivative results of the variables,the weight factors that reflect the importance of each variable are obtained;The relative changes of the weighting factors are used to establish a fault diagnosis method driven by k-nearest neighbors.Finally,experimental simulations are used to verify.2)Kalman filter-based adaptive identification method for multi-dimensional Taylor network parameters.First,the input-output equations of the multi-dimensional Taylor network are regarded as the measurement equations of the parameters to be identified;second,the parameters are described as a dynamic process based on random walks;and the linear Kalman filter is used to establish an adaptive identification of network parameters with increasing samples.Algorithm;Finally,the validity of the method to be identified is verified numerically.3)Establish a multi-dimensional multi-dimensional Taylor network self-encoding and performance analysis method.First,based on different orders,build a multi-layered Taylor network self-encoding network framework that gradually increases with the order.Second,use the continuous learning method to layer by layer.Train network parameters;finally,test and evaluate the representation capabilities of the network and the existing stack-type self-encoding network through the same input and output tag data.