Fault Diagnosis Method Based On Improved ICA

Independent component analysis (ICA) is developed in the late20th century. As amulti-channel signal decomposition method, ICA becomes a new approach in the signalfield and one of the hotspots in the development of signal processing techniques. Inmost cases, signals observed by sensors are generally signals mixed by someindependent components. Independent signals are separated by ICA and thecharacteristic signals are extracted. And the application of ICA in fault monitoring isalso very successful, and ICA has also become a novel method of fault diagnosis.There are5chapters in this thesis. In first part, the basic principles of ICA wereexhaustively analyzed, including the description of ICA model, basic assumptions anduncertainty analysis. Two kind of algorithms, batch and adaptive, were elaborated froman online perspective. In next part, the statistical methods based on PCA and ICA werestudied, we took advantage of the I2and SPE statistic variables to monitor the fault andcompared with T2statistics, the signals extracted by ICA were classified by use ofself-organizing map (SOM). Then, we conducted some simulation, which include aECG feature extraction experiment, a fault monitoring and fault classificationsimulation of batch processes and vibration process. Each algorithm was compared withPCA. Finally, this thesis proposed a idea which has merged adaptive algorithm andbatch algorithm. To improve the traditional adaptive methods, a variable movingwindow adaptive ICA was proposed, and provided a form of change criteria at the sametime. The simulation on adaptive ICA compared with the simulation on variable movingwindow ICA in convergence and separation.