Study On The Key Technologies Of Prognostics And Health Management For The Suspension System Of Middle-low Speed Maglev Trains

As a new type of urban rail transportation,more and more public recognition and attention have been paid to the middle-low speed maglev trains.Beijing and Changsha have established demonstration lines and opened operations,while Guangdong Qingyuan and Hunan Phoenix are also constructing middle-low speed maglev tourist lines.As one of the key systems of the middle-low speed maglev train,the safety and reliability of the suspension system play an important role in the operation of the middle-low speed maglev train.Therefore,prognostic and health management(PHM)technology,as an effective means to improve system operation safety,has received extensive attention and did research from the academia and industry field.So,it is of great significance to study the PHM technology of the suspension system for improving safety and reliability.This thesis has paid attention to the key technologies of PHM,both the onboard PHM system and the ground PHM system.At the onboard PHM system level,firstly,two anomaly detection models are proposed,one is a single-dimensional time series anomaly detection model inspired by the Gaussian distribution of hyperspheres,the other is a multi-dimensional time series anomaly detection model using improved canonical correlation analysis(CCA);Then,it comes to fault detection model where autocorrelation length and stable kernel representation(SKR)has been successfully applied to.At the ground PHM system level,two methods are proposed.One is a health assessment method hinging on standard core status and information fusion to regard unobvious sequence trends caused by multiple operating conditions(such as changes in load,speed,orbit,etc.);the other is the remaining life prediction method combining dynamic time warping(DTW)and kernel density estimator(KDE),aiming at tackling shortcomings in the similarity function and weight function of the traditional similarity-basedThe main objective and innovations of the thesis are listed into five parts as follows:(1)The first focus is applying the monitoring data of the suspension gap to deal with data imbalance in anomaly detection,as a result,a single-dimensional time series anomaly detection method is proposed,where the Gaussian distribution of hyperspheres is used.After extracting the features of the suspension gap via Fast Walsh-Hadamard Transform(FWHT)under normal conditions;Standardized processing and principal component analysis(PCA)are used to minimize the difference,correlation,and high dimension so that the distribution space of features is transformed into an approximate hypersphere;finally,the anomaly threshold is determined according to the distribution characteristic of Euclidean distance,and an anomaly detection model is established.Validation of operating line data shows that the proposed method has a better detection rate comparing with the empirical threshold methods,PCA methods,and support vector data description(SVDD).(2)The second attempt is a multi-dimensional time series anomaly detection method hinging on improved CCA,and the model is inspired by the single-dimensional time series anomaly detection method but it uses the multi-type monitoring data of the suspension system.This method uses CCA to process multi-dimensional data,including gap,current,acceleration,voltage,and velocity of the suspension system,to obtain secondary statistics(non-Gaussian distribution);Followed by using Box-Cox transformation to convert the secondary statistics into Gaussian distribution variables;After that,the characteristics of Gaussian distribution are used to determine the abnormal threshold.The operating line data verified that this method can obtain a better threshold compared with the method based on K-medoids and the method based on SVDD,and it can obtain a higher detection rate.(3)The third aim is to achieve self-diagnosis of the electromagnet and other components in the suspension system that has not to be realized,and a fault detection method using autocorrelation length and SKR is studied.First,the length of sampling the data used for SKR is determined by the autocorrelation length;then,SKR is used to obtain the residual of the sampled data and set the residual threshold.The experimental results show that the research method can detect the failure of the suspension control box in real-time effectively.The simulation results show that the research method can detect the failure of the electromagnet in real-time effectively.(4)Aiming at achieving health assessment under multiple operating conditions(such as changes in load,speed,track type,etc.)and circumventing unobvious sequence trends,a method based on standard core status and information fusion is proposed.First,K-means is used to cluster the operating conditions and obtain the core state under the concept of core state;then,after analyzing the impact of different core states on the estimated health index(HI),the sequence under different core states is converted into the standard core by standardization processing.Then,according to the monotonicity of the sequence,the sequences with obvious degeneration trends are selected;Finally,according to the contribution rate of the sequences,the linear regression model is used to obtain the mapping relationship between the sequence and the HI,to obtain the estimate HI.The simulation results and experimental results show that the HI obtained by this method is very close to the real HI,and it has practical value.(5)To handle the two main drawbacks of the traditional similarity remaining life prediction: one is ignoring the sequence of relative points when a Euclidean distance similarity function is involved and the other is causing a large error when the results are obtained via the weighted average using a few available training samples,is of importance for higher accuracy of predicting.So,the remaining life prediction method taking DTW and KDE into consideration to get rid of the drawback is presented.The exponential similarity function inspired by the DTW distance is used to calculate the similarity between the HI of all training samples and the HI of the test samples;further,after obtaining the remaining life of the first p training samples with high similarity,KDE is utilized to obtain the remaining life probability distribution of the test sample and estimate its remaining life.Both the simulation results and experimental results show that this method can not only gain better prediction results when fewer training samples are involved but also helps maintainers to perform predictive maintenance on the suspension system.To meet the actual engineering needs of middle-low speed maglev operating lines,this thesis conducts in-depth research on four aspects: abnormality detection,fault detection,health assessment,and remaining life prediction in the suspension system PHM.The research results lay a foundation for the intelligent operation and maintenance of middle-low speed maglev operating lines and can be promoted and applied to the intelligent operation and maintenance of high-speed maglev trains,which have important reference values.