Fault Diagnosis And Prediction Of Complex Dynamic Systems

With the rapid breakthrough of industrial technology,the development trend of the system is gradually becoming highly integrated and increasingly complicated.These complex systems will inevitably fail after long-term operation.They will be bound to lead to serious consequences if not discovered and processed in time.Therefore,it is essential to conduct fault diagnosis and fault prediction of the system timely and effectively.At the same time,advances in communication technology and the popularity of networks have led to the widespread use of networked control systems in the production process.However,when data is transmitted in the network environment,it will inevitably suffer from the inherent shortcomings of the network,such as delay,packet dropout and so on,which increases the difficulty of fault diagnosis of the networked control system.In this thesis,the problem of fault diagnosis for multi-sensor networked dynamic systems with multiple packet dropout rates and parameter perturbation is studied.Firstly,multiple independent Bernoulli sequences are used to model the packet dropout process of multisensors,and a networked unscented Kalman filter is designed.Then,a strong tracking idea is introduced based on the networked unscented Kalman filter and a networked strong tracking unscented Kalman filter is proposed.Next,through the state expansion,the estimated value of the fault amplitude is obtained while estimating the state.Finally,the above method is verified on the three water tanks.The results show that under the environment of multiple packet dropout rates and parameter perturbation,the designed filter can accurately track the system state and the magnitude of the fault,and timely and accurately fault.diagnosis.For some complex industrial systems,the specific mechanism is difficult to know.The data-driven methods can extract information directly from historical operational data without the need to understand the mechanism knowledge,and then realize fault diagnosis or prediction.However,the existing majority of data-driven based methods are offline,which leads to low model matching in the later prediction process.This thesis proposes an online prediction method for the remaining useful life of lithium batteries based on online extreme learning machine,which can not only update the prediction model online according to the latest incoming data,but also update the prediction error online to correct the prediction value,thereby improving the matching degree of the model and the accuracy of the prediction data.Finally,simulations of several lithium battery data verify the effectiveness of the proposed method.Finally,the work of the full thesis is summarized,and the future research contents are prospected.