Research On Adaptive Approach Of Sensor Fault Detection In Industrial Environment

With the continuous development of automatic technology, the problem of sensorfault detection in oil field system is no longer confined to situated AI inspection andmaintenance, so the intelligent method which is applied to the sensor fault has a highpractical value. With the expansion of the scale in oil field production system and thecomplexity of the constantly improving, the fault of sensor groups in the process ofproduction is increasing, and hard to be found out. As a result, higher requirements forsensor fault detection is required. The requirements can’t be satisfied with the traditionalapproach of limit alarm and manual check. It is necessary to study the advanced faultdiagnosis technology of oil field production system.Based on multivariate statistical analysis method, principal component analysis (PCA)is used in this paper due to many characteristics such as large complex structure, numeroussensors and other features in the oil field production system. The method of principalcomponent analysis (PCA) is used to diagnosis the sensor fault by analyzing the processdata. PCA model is established in the normal condition. The sensor fault is detected bycomparing the statistic and the control limit. After the failure is detected, this study focuson isolating faulty of the sensor by calculating the sensor fault index through the method offault image vector (FIV). An improved approach to sensor fault isolation based on particleswarm optimization-least squares support vector machine (PSO-LSSVM) is presented tosolve the problem that the output of sensors is volatilized The mathod ensure the accuracyof sensor fault isolation. The constant control limit of PCA is not suitable for the dynamicdata, because it can result in large false alarm, the method of recursive principalcomponent analysis (RPCA) is used to improve the detection method and singular valuedecomposition (SVD) method is used to simplify the covariance matrix of the recursiveprocess. Finally index weighted is used to recursive control limit, which can achieve theadaptive method of sensor fault detection. Completed work is summarized as follows: Firstly, the various methods of fault detection are researched in theory. The feasibilityand superiority of using PCA to detecting the fault in oilfield production system isconcluded by aiming at the characteristics of the sensor output data. The basic theory ofoilfield system fault diagnosis based on PCA is described. The simulation results show theeffectiveness of the method.Secondly, point at the problem of fault isolation in oil system, a new method for faultisolation based on the fault image vector is proposed. The method can effectively improvethe accuracy of fault isolation. PCA is used to deal with the sensor data and estimatewhether the sensor have fault through observing the information of sensor fault index.Thirdly, PSO-LSSVM is adopted to predict the output data to deal with the problemthat the output of sensors is volatilized which has an affect on the isolation result of FIV.The difference between prodicted and the measured sensor method is used to calculate thesensor fault index. This method can isolate the fault sensor effectively.Fourthly, SVD algorithm is used to improve RPCA, in order to achieving self-adapti-ve fault detection methods. The improved method can be well adapted to the time-varyingcharacteristic of the system which has a high accuracy in the detection and isolation fault.In the research, a large number of situlation experiments has been done for variousfault diagnosis method by using of the data from the field in this paper. Testing resultsshow that the method can detect the fault effectively and improve the accuracy of isolation.