Research On Industry Process Data Reconciliation

Measurements are the base and starting point of the engineering technique work such as process design, simulation, optimization and control. However, process measurements from industrial plants are inherently containing random errors and possibly contaminated by gross errors, which will lead these measurements are not consistent with corresponding conservation laws and other process constraints. The goal of data reconciliation is to smooth the measured data then enhance their accuracy and reliability and to estimate unmeasured process variables. Data rectification contains three aspects:data reconciliation, gross error detection, redundancy analysis. My work is concerned with steady state process mainly studying in several aspects as follows.First, I have embedded analyzed steady model and several solved problems to date reconciliation, known nonlinear and dynamic condition problems at the same time. Because of the existence of non-reconcilable variables and non-estimable variables, this can cause the stoppage of compute. To solver this problem, we have studied date class methods, for example, Crowe matrix projection method, QR method, and improved on two-step matrix projection method. Through theory prove and example analysis, improved two-step matrix projection can make date class correctly, not only can avoid the lose of reconcilable variable, but also can reduce matrix dimension and predigest computation, so this can solver the problem of stoppage of computation.Second, I have studied the principle of gross error detection and the methods of gross error detection, for example:Node Test (NT), General Test (GT), Measure Test (MT), PCA, analysed strongpoint and shortcoming of MT and NT, studied gross error detection method MT-NT; studied on variance estimate, contain direct method and indirect method.Last, I have studied detecting gross errors using ANN. The practical application shows that this method detects gross errors rapidly and efficiently after ANN being trained properly.Artificial Neural Networks is a useful simulative tool and has a potential to identify rapidly and easily on nonlinear system on the basic of historical data. The specified process model and definitive knowledge of the measurement error statistics are not necessary.