摘要:
Systems and methods (300) for offline/online performance monitoring of batch processes (BPs) involving obtaining archived data (AD) obtained during runs of BP and including information defining a batch quality attribute for each run. The method also involves forming clusters by classifying AD for the runs into classes based on the batch quality attribute(s) and building a first multivariate statistical model (MSM) using AD. The method can further involve building a wavelet analysis based feature matrix (FM) using AD, forming a first projection (1200) by projecting FM onto a first MSM, building a second MSM (1300) using information obtained from the first projection, and computing centroids (C902, . . . , C918) and boundary profiles for the clusters (902, . . . , 918). The method can involve performing an online/offline performance monitoring (700/800) using an integrated version of the first and second MSM, a classification algorithm, centroids, and boundary profiles.
摘要:
A method (300, 400, 500, 1200) for offline/online monitoring of batch processes. The method involves (312) decomposing a time domain of a batch process run (BPR) into several blocks and (334) building multivariate statistical models (MSMs) for each of them using archived data for a batch process (ABPD). ABPD comprises stored data obtained during BPRs. The method also involves (506, 1204) retrieving recently stored data (RSD) for a recent fully performed BPR run (FPRNEW) or current BPR run. The method further involves (520, 1210) building a feature vector matrix (FVM) using RSD. FVM contains feature vectors representing statistical measures of wavelet coefficients determined for variables (v0, . . . , vJ). A projection (1100, 1150, 1190) is formed by projecting feature vectors onto at least one MSM or a combined multivariate statistical model (CMSM). CMSM is a weighted average of at least two MSMs. Subsequently, at least one estimate is generated for FPRNEW or current BPR run using information provided by the projection (528, 1220).
摘要:
Systems and methods (300) for offline/online performance monitoring of batch processes (BPs) involving obtaining archived data (AD) obtained during runs of BP and including information defining a batch quality attribute for each run. The method also involves forming clusters by classifying AD for the runs into classes based on the batch quality attribute(s) and building a first multivariate statistical model (MSM) using AD. The method can further involve building a wavelet analysis based feature matrix (FM) using AD, forming a first projection (1200) by projecting FM onto a first MSM, building a second MSM (1300) using information obtained from the first projection, and computing centroids (C902, . . . , C918) and boundary profiles for the clusters (902, . . . , 918). The method can involve performing an online/offline performance monitoring (700/800) using an integrated version of the first and second MSM, a classification algorithm, centroids, and boundary profiles.
摘要:
A method (300, 400, 500, 1200) for offline/online monitoring of batch processes. The method involves (312) decomposing a time domain of a batch process run (BPR) into several blocks and (334) building multivariate statistical models (MSMs) for each of them using archived data for a batch process (ABPD). ABPD comprises stored data obtained during BPRs. The method also involves (506, 1204) retrieving recently stored data (RSD) for a recent fully performed BPR run (FPRNEW) or current BPR run. The method further involves (520, 1210) building a feature vector matrix (FVM) using RSD. FVM contains feature vectors representing statistical measures of wavelet coefficients determined for variables (v0, . . . , vJ). A projection (1100, 1150, 1190) is formed by projecting feature vectors onto at least one MSM or a combined multivariate statistical model (CMSM). CMSM is a weighted average of at least two MSMs. Subsequently, at least one estimate is generated for FPRNEW or current BPR run using information provided by the projection (528, 1220).
摘要:
A method for performing diagnosis on an engine includes the steps of obtaining data for a plurality of variables pertaining to the engine, transforming the data with respect to each of the plurality of variables using a wavelet transformation, to thereby generate initial coefficients for each of the plurality of variables, and aggregating the initial coefficients for each of the plurality of variables, to thereby generate an aggregate set of coefficients for the plurality of variables.