type of modeling is performed between the rearranged X matrix which is augmented as a new data partition becomes available. As more data become available local models are developed (model no. 1.. .7 in Table 6.14) and explained variance in Y block increases with each local model as shown in Table 6.14.

Process variables for the new batch are sampled at percent increments of volume decrease for phase 1. After the completion of phase 1, the sampling rate is switched to percent completion of the amount of substrate added. New data vector xnew (1 x J) is monitored by using the following algorithm for each sampling point from k = 1 to k = K\, K2 for both phases

3. Calculate new batch scores, SPE, T2 and variable contributions to these statistics by using the information generated by MPLSV model

4. Compute zpred

5. Check MV control charts for abnormalities

Process monitoring and quality prediction stage: A small drift of magnitude -0.018% h~l was introduced into substrate feed rate from the start of fed-batch operation at 50 h until the end of the batch run as a test case. There are significant differences in fault detection times and out-of-control signal generation by different charts (Table 6.15). T2 detected the fault fastest (Figure 6.80). Second fastest detection is obtained by the linear score control chart of latent variable 2 (Figures 6.81 and 6.82). The last four latent variables give out-of-control signals for both linear and nonlinear score matrices. Although SPE is in-control throughout the course

Table 6.15. Fault detection times

Type % completed IV Time, h


Was this article helpful?

0 0

Post a comment