Fault Identification And State Prediction For Process Equipment Data

With the development of detective equipment and instruments,data can be accessed more easily which results in exponential increase for the magnitude of the data.Consequently,how to extract useful information from massive data to guide the practical industrial production and the current life is becoming one of hot spots.Aiming at big data generated from process equipment during production process,some appropriate machine learning algorithms are chosen to combine or improve for fitting these data’s characteristics.And then,several effective techniques based on these data-driven models are developed to deal with some key challenges such as fault diagnosis and process craft optimization.The main contributions of this thesis is generalized as follows.(1)Aiming at the characteristics of industrial water pumps in actual operating environment,such as large amount of working data,long operating time,and multiple feature types,a combined fault diagnosis classification based on feature weighted Gaussian weighted K nearest neighbor-support vector machine(GWKNN-SVM)is proposed.And then,the effectiveness of the designed fault diagnosis model is shown through comparing with other three algorithms under the same data set generated by three pumps in the same actual operating environment.(2)In view of the fact that the current octane loss value is difficult to obtain in the gasoline catalytic cracking process,the process flow is complex,and the characteristic data are coupled with each other,a random forest algorithm based on information gain and particle swarm optimization(IG-PSO-RF)is proposed to obtain octane loss prediction model.And then,the effectiveness of the proposed predication model is verified through comparisons with others classic models when all models are trained and built by the same data set generated from different operating periods.(3)Aiming at the characteristics of penicillin fermentation process in actual production,such as long cycle,multiple characteristic stages,and time-consuming difficulty in data collection,a fuzzy C-means clustering and Drosophila optimized least square support vector(FCM-PSO-LSSVM)is proposed.Moreover,the proposed prediction model of penicillin concentration is compared with other algorithms trained with the same data generated by the well-known Pensim simulation platform and the effectiveness of the prediction model is also verified.