改进MKPCA方法及其在发酵过程监控中的应用

Application of an improved multi-way kernel principal component analysis method in fermentation process monitoring

针对间歇发酵过程缓慢时变和非线性等特点,提出一种基于滑动窗技术的多向核主元分析(MWMKPCA)方法。该方法结合了核主元分析(KPCA)和滑动窗口技术的优点,其中KPCA能有效解决过程数据的非线性问题,保证数据信息抽取的完整性;而滑动窗口技术能有效避免MKPCA在线应用时预报未来测量值所引入的误差,提高监控性能。对于已判断正常的新批次过程数据,将其加入模型参考数据库进行更新,从而提高间歇过程性能检测的准确性。将该方法应用到工业青霉素发酵过程的监控中,并与MPCA、MKPCA方法的监测性能进行了比较。结果表明:该方法能有效提取过程变量间的非线性关系,降低运行过程的误报率,对缓慢时变的间歇过程具有更可靠的检测性能。

Aiming at the complex nonlinear characteristic and slow time-varying behavior of batch process, a new method was developed based on a moving window MKPCA （multi-way kernel principal component analysis） for on-line batch process monitoring. The proposed method uses a moving window and does not require predicting the future value of the current batch; while the nonlinear characteristics within normal batch processes are cap- tured by using KPCA. It also enhances the reliability of the monitoring system through consecutively updating the database of normal batches. The proposed method is used to evaluate the industrial penicillin fermentation process data and is compared with traditional MPCA and MKPCA methods. Results show that the proposed method has better performance, can effectively extract the nonlinear relationships among the process variables, adapts to new normal operating conditions and decrease false alarm rate.