Sequential Time Slice Alignment Based Unequal-Length Phase Identification and Modeling for Fault Detection of Irregular Batches

Abstract

In practice, batch processes may have different durations, revealing unsynchronized operation events, due to the changes in operation conditions or control objectives. For batches with multiple operation phases, batch-wise process characteristics are irregular, and at the same time, phases are misaligned over batches, which have caused problems in phase analysis and modeling as well as online process monitoring. To solve the uneven-length problem in multiphase batch processes, this paper proposes a sequential time slice alignment based unequal-length phase identification and modeling method for fault detection of irregular batches. In comparison with previous work, the major contribution of the proposed method is as follows:(1) The irregular process characteristics are evaluated in sequence and directly related with the monitoring performance. (2) Multiple irregular phases are readily identified and modeled using sequential time slice alignment which avoids cumbersome postprocessing. (3) The sequential nature provides an easy way to real-time judge the phase affiliation of each new sample for online fault detection. Also, comparison is conducted between the proposed algorithm and clustering-based uneven-length phase division and process monitoring algorithm. The application to a typical batch process with varying durations illustrates the online monitoring performance of the proposed method.

Monitoring Uneven Multistage/Multiphase Batch Processes using Trajectory‐Based Fuzzy Phase Partition and Hybrid MPCA Models

“…The representative data‐driven phase partition methods include the multiphase (MP) algorithm, the sub‐PCA method, the stepwise sequential phase partition (SSPP) algorithm, the Gaussian mixture model (GMM)‐based method, and so on. Some data‐driven phase partition methods were also proposed to handle the uneven duration problem in multiphase batch processes . For example, Lu et al developed a revised sub‐PCA phase partition method for multiphase batch processes with uneven‐length operation phases.…”

“…Li et al extended the SSPP algorithm to the uneven batch case. The proposed method detects phase partition points sequentially by checking changes in variable correlations, while it does not distinguish phases and between‐phase transitions . To cope with the transitions between operation phases, Luo et al proposed a fuzzy phase partition method on the basis of the sequence‐constraint fuzzy c‐means (SCFCM) algorithm.…”

“…Therefore, to build high performance process models, operation phases should be treated differently according to their own process characteristics. However, most existing methods model all operation phases using the same modelling technique, ignoring the differences in process characteristics between operation phases. In fact, each modelling technique has a specific scope of application.…”

“…Some data-driven phase partition methods were also proposed to handle the uneven duration problem in multiphase batch processes. [16][17][18][19] For example, Lu et al [13] developed a revised sub-PCA phase partition method for multiphase batch processes with uneven-length operation phases. This method carries out phase partition by clustering the weighted loading matrices of time-slice PCA models via the K-means (KM) algorithm.…”

“…[17] However, it also ignores the between-phase transitions. Li et al [18] extended the SSPP algorithm to the uneven batch case. The proposed method detects phase partition points sequentially by checking changes in variable correlations, while it does not distinguish phases and between-phase transitions.…”