Research And Implementation On Data-driven Fault Diagnosis Methods Of Service Robot's Dead-reckoning Subsystem

With the rapid development of artificial intelligence technology and cloud computing,the research and industrialization process of service robots have been greatly accelerated in recent years.Working in a dynamic,complex and semi-structured home environment,service robots possibly offer services to the group with weak self-protection ability,such as the elderly,the young,the disabled and etc.Therefore,instead of pursuing high speed,efficiency and precision of industry robots,people put more emphasis on the reliability,security and intelligence level of home service robots.At present,in addition to the deficiency of intelligence,the low reliability and security have become one of the constraint factors to the popularization of service robots.The research of fault diagnosis method is one of the effective ways to improve the security and reliability of service robots.However,the existing researches on service robots mainly concentrate on navigation,control,external environment perception and understanding(e.g.,human behavior understanding and home scenario understanding),and etc.But the perception of the robot's internal environment(i.e.,health monitoring and fault diagnosis)is less studied.In order to improve the security and reliability of service robots,a software simulation platform and a hardware platform for robot fault diagnosis are designed as the experimental technique.Taking common faults of dead reckoning subsystem as examples,using the data-driven intelligent fault diagnosis method,the article focuses on the key issues of robot fault diagnosis.The main researches are as follows.(1)A full database of fault samples is the basis of studying data-driven fault diagnosis method.In order to solve the problem of lack of robot fault sample database and standard experimental platform for robot fault diagnosis,the paper presents a design scheme of hardware platform which provides real data for the research of data-driven fault diagnosis method.In addition,a high-precision visual localization approach based on probabilistic confidence model is proposed to reduce the impact of inaccurate pose information on diagnosis results.Besides,a service robot fault software simulation platform which can realize multiple functions(e.g.,fault injection,data collection and analysis,3D visualization,algorithm verification)is designed.(2)A fault diagnosis method based on multi-model perception and decision fusion is proposed.A service robot is a complicated mechatronic intelligent device with various heterogeneous sensors.Additionally,it works in a dynamic environment and carries out multiple service tasks.All the factors mentioned above jointly determine the complexity of robot fault diagnosis.Due to the incomplete perception of operational information about a robot based on single PC A model,an improved fault diagnosis method for the dead-reckoning subsystem based on multiple PCA models and decision fusion is proposed.Firstly,multiple PCA models are established.Fault detection and feature extraction are accomplished with these models.Then parameter optimization of the multi-class probabilistic SVM using an improved PSO algorithm is carried out,and the preparatory separation of fault is realized.At last,a new BPA definition method based on confusion matrix is proposed,and the fusion of primary fault isolation results is done using DS evidence theory.With the establishment of multiple PCA models,parameter optimization based on improved PSO algorithm and decision fusion,the proposed fault diagnosis method effectively improves the diagnosis precision of the dead-reckoning subsystem of a service robot.(3)We propose a new generalized Gaussian kernel.The kernel is proved to be positive and symmetrical which satisfies the Mercer condition.The classification performance of the proposed kernel function is compared with the existing kernel functions based on UCI benchmark test set and fault samples.Experimental results show that the proposed kernel function is a generalized form of the classical Gaussian kernel and has a better classification performance.By introducing the new generalized Gaussian kernel and optimizing fault diagnosis architecture mentioned above,the improved fault diagnosis method further raises the diagnosis precision of the dead-reckoning subsystem of a service robot.(4)In order to solve the problem of complexity of SVM training and the inability to efficiently process large-scale data,a fast SVM training method based on GPU and support vector selection algorithm is proposed.This method can accelerate the training process of fault diagnosis model and improve the real-time performance of the algorithm.The advantages of the method include the following aspects.On one hand,the support vector selection algorithm can effectively reduce the training samples while retaining support vectors as much as possible.On the other hand,the method can avoid numerous repeated calculations in the process of parameter optimization,cross validation and SMO by precalculation of kernel matrix utilizing the parallel computing ability of GPU.The effectiveness of the proposed method is proved by multi-group artificial data sets and robot fault diagnosis experiments.