Off-line analysis and troubleshooting of a batch process.
This case study describes the batch analysis and troubleshooting performed on an agricultural chemical. The work was done with FMC corporation1 and the analysis used the BatchSPC software2. BatchSPC has since been replaced by ProSensus MultiVariate.
The process and objectives
The batch process evaporates solvent from an initial wet cake charge. A certain chemical structure is desired at the end point, characterized by 11 quality variables. The data analysis aimed to uncover the reason for off-spec production.
The batch reactor system and the solvent collector tank.
The data available
The data are divided into three main categories:
- Z: Initial information: the weight of material charged to the reactor and nine chemical properties of this material;
- Y: Final quality: 11 measurements that characterize the final product; and
- X: There are 71 batches, each batch of variable duration, measuring 10 tags. These include measurements such as tank level, differential pressure, agitator power, jacket set point temperature and actual temperature, all collected in X.
The raw batch data: before and after aligning the trajectories in matrix X
It is helpful to first visualize the data before proceeding to align them into a cube. The alignment step reshapes each trajectory so that all batches can be compared on a scale where they all evolve similarly. By plotting out the data it becomes apparent how to align the trajectories of each measurement. Alignment is a critical step, and usually case specific.
The figure below shows some of the original data on the left and the aligned data on the right.
Alignment of the batch trajectories: before and after
Analysis of quality
There is a known quality problem and it can be quickly detected. The 11 quality measurements can be reduced down to a 2-D plot by performing PCA. This plot explains 70% of the quality measurements and clearly separates the on- and off-spec product. The company provided the on- and off-spec designation.
There is a clear separation between the batches when using the final quality data
Factors affecting final quality
The initial chemical information in the matrix Z and the batch trajectories in X are both sources that affect the final product quality (Y). The previous analysis used a multivariate technique, PCA, on a single matrix. Here we used PLS to relate two data matrices. The Z (chemical) and Y relationship is first examined, and then the X (trajectory) and Y relationship.
Results from both analyses were consistent and provided interesting operational insight:
- The operators used a variable to adjust the process. It was shown that this variable had little effect on the final outcome.
- Quality was strongly affected by the moisture of the initial charge and not so much by its chemistry. Also of importance in achieving an on-spec product was the way that the batch was operated, the recipe.
- Analysis of the trajectory information showed that high-quality batches are more likely if these guidelines were followed:
- Maintain a low solvent level in the collecting tank at all times
- Use high pressure and jacket temperature set points during the 1st stage
- Apply fast evaporation during the 2nd stage
- Operate the batch as fast as possible.
An analysis of the bad batches is possible with the use of contribution plots. These plots show which measurements contribute to the poor operation.
Conclusions
The batch analysis techniques that ProSensus uses have been shown to be powerful at isolating and diagnosing poor process behaviour. Having knowledge of events that lead to poor operation results in process changes to minimize poor production and associated wastage.
Footnotes
- The complete study is published in the open literature: Salvador Garcia-Munoz, Theodora Kourti and John F. MacGregor, A.G. Mateos, and G. Murphy, (2003). Troubleshooting of an industrial batch process using multivariate methods, Industrial and Engineering Chemistry Research, 42(15), 3592-3601.
- Batch data analysis was performed using BatchSPC, formerly available from the McMaster Advanced Control Consortium. BatchSPC has been replaced by ProSensus MultiVariate.
