The Optimized DBN Algorithm And It's Application In Fault Diagnosis Of Navigation System IMUs

With the continuous development and maturity of robot technology,robots have increasingly higher requirements for the accuracy and stability of their navigation systems.Inertial navigation units(IMUs)are one of the important components of a robot's navigation system,and their fault will affect the stability and operability of the robot.Therefore,fault detection and diagnosis of IMUs has become the focus of research on improving the stability and reliability of robot navigation systems.Based on this,we focus on three aspects for our research in this paper:the research and optimization of the deep belief networks(DBNs)for the fault diagnosis of IMUs,the creation of IMUs fault dataset for wheeled robots,and the evaluation of the optimized DBN's fault diagnosis performance.The content mainly covers the following aspects:Firstly,the verification of restricted boltzmann machine(RBM)feature extraction capability and the verification of DBN classification capability were completed based on the research on the working principle of DBN.According to the needs of the project,three RBMs are stacked to form a DBN,and a Softmax classifier is superimposed on the top of the DBN to complete the construction of the DBN fault diagnosis model.At the same time,the output signals of IMUs are affected by external interference signals and the cumulative error of the inertial sensor itself,resulting in complex and diverse faults and low correlation between data,which greatly reduces the real-time and accuracy of fault diagnosis of the DBN model.Therefore,the DBN is optimized from the aspects of weight optimization and the number of DBN's hidden-layer neurons.Inexact LSA-GA was introduced into the DBN's weight fine-tuning process,meanwhile combining Inexact LSA-GA with GA,the terminal optimization of the weights was accelerated and the accuracy of the weight optimization was improved.The bat algorithm(BA)is introduced to dynamically adjust the number of DBN's hidden neurons by changing the position of the bat colony.Secondly,through the combination of experiments and simulations,the wheeled robot equipped with MPU6050 chip is used as the research object to complete the creation of IMUs fault datasets,based on the research on the working characteristics of wheeled robot IMUs.During the experiment,local outlier factor(LOF)and normalization algorithm were used to pre-process the data to eliminate outliers caused by the inertial sensor's own error accumulation and external environment interference and to reduce errors between data samples in the same motion state.At the same time,the fault signals of common IMUs'faults were simulated through the simulation process,and the fusion of preprocessed data and fault signals was achieved through attitude solution.The IMUs fault dataset with the data structure of 39000x6 was created.Finally,a comprehensive evaluation of the performance of the optimized DBN fault diagnosis model is completed,based on the three aspects of fault diagnosis accuracy rate,fault classification performance and generalization ability.The results show that the optimized DBN fault diagnosis model has better fault classification ability and generalization ability on the given dataset,and the dynamically adjustable "DBN structure" facilitates the extraction of data associations between multiple types of fault categories.Therefore,the optimized DBN fault diagnosis model proposed in this paper provides a preferred reference model for general fault diagnosis problems based on data association.