Study On Feature Extraction Methods For Fault Detection Of Industiral Processes

Nowadays, the production rate and safety gain a great improvement with the introduction of machines, but their faulty conditions may lead to some disastrous events for both people and facilities. With the development of plant scales, it is no longer feasible for engineers to keep an eye on all varibles one by one. Simutenously, a large amount of data which are responsible for various operation states of industrial processes are stored owing to the introduction of distributed control system. Therefore, in order to improve the monitoring efficiency, data driven process monitoring methods are raised up.In fact, there are several characteristics of real industrial processes such as multimode, time-varying and complex data distribution, etc. In the early period, people would like to make some assumptions about the processes or collected data to simplify the research problems. But with the intensive study of process properties and the continuous emergence of powerful algorithms, it is time to get rid of those assumptions which get in the way of the development of process monitoring. The present thesis intends to utilize the local and global data structure information to cope with the multimode, time-varying and complex data distribution behaviors of industrial processes and build more accurate monitoring models. Detailed contributions are listed as follows:(1) In order to deal with the complex data distribution problem, a novel process monitoring method is proposed based on locally linear embedding (LLE) and support vector data description (SVDD). The algorithm takes the advantage of LLE and SVDD where no assumptions of the data distribution are made and can achieve an accurate monitoring model. Moreover, a partial least approach is introduced to calculate the mapping matrix from the data space to the feature space and thus online samples can be directly conducted.(2) Since conventional methods may destroy the local or global data structure during feature extraction, a process monitoring method called local and nonlocal embedding (LNLE) is put forward to tackle this problem. Aligned with the objective function of LLE which means to preserve the local data structure, a new objective function is developed to preserve the relationship between a sample and others which lie in its nonlocal area. Then, a unified optimization is constructed by minimizing the distances among neighborhood samples and maximizing the distances among nonlocal samples with an orthogonal constraint of the mapping matrix for extracting a compact representation of the original data space. Therefore, both local and global data structure information in the training data can be preserved in the feature space.(3) A novel scheme based on aligned mixture factor analysis (AMFA) is proposed for multimode process monitoring. On the basis of the traditional approaches by constructing multiple local monitoring models where statistical fingerprints of the training data can be ahieved, AMFA further aligns the separated local models together into a global model by constraining that all samples should have a unique expression in the final coordination. Through this approach of dividing and integrating, both within-mode and cross-mode correlations are believed to be greatly preserved in the global model. In addition, since it does not need to calculate the posterior probabilities of new samples in the online monitoring period, the monitoring efficiency can be guaranteed.(4) For model alignment and multimode process monitoring, a new framework named neighborhood based global coordination (NBGC) is proposed. In order to identify the different patterns in the training database, automatically calculate the mode number and avoid the local extremum problem, a new clustering method is derived by utilizing the serial correlations between adjacent samples. With local outlier probability (LoOP) and an arrangement approach, the fracture parts between multiple modes can be located and similar segments can be pieced together. Furthermore, during the approach of aligning all local models into a global one, NBGC seeks to preserve both local and global data structure in the coordinated space and a weighted factor is introduced to balance the tradeoff between the above two parts.(5) Two numerically efficient algorithms called moving window local outlier factor (MWLOF) and moving window local outlier probability (MWLoOP) are proposed. The key features of the proposed algorithms can be identified as its underlying capability to handle complex data distributions and incursive operating condition changes including both slow dynamic variations and instant mode shifts. Based on the advantages of the two algorithms, the model accuracy can be guaranteed under the effect of complex data distributions. With some updating rules developed for accelerating the computation speed, two-step adaption approachs are introduced to keep the monitoring models up-to-date. For operating mode changes, two semi-supervised switch strategies are designed and corresponding update termination rules are incorporated to prevent the models from accepting faulty samples or disturbances.Except for some theoretical analysis of the proposed methods, their feasibility, utility and superiority over the compared methods are demonstrated through numerical examples, the non-isothermal continuous stirred tank reactor (CSTR) process and the Tennessee Eastman (TE) benchmark process.