Troubleshooting an Industrial Batch Process
Batch processes play key roles in many industries including pharmaceutical, polymer, specialty chemicals and food. Multivariate methods are an effective way to troubleshoot their costly operating issues. With its batch alignment tool and easy to interpret graphs, ProMV helps users effectively troubleshoot batch data.
A herbicide manufacturer was producing a large number of out-of-specification batches in their batch dryer. The goal of this work was to troubleshoot why unacceptable chemical structural changes were occurring in so many batches.
Plant engineers and chemists had been convinced that the off-spec product was the result of the properties of the raw materials. However, the multivariate analysis showed that the operating conditions were actually the significant contributors to the quality problems. An operating strategy was identified that would lead to the production of good quality batches.
The ProSensus Approach
A batch analysis usually contains a large volume of data. In this batch dryer the data available include initial conditions and chemical properties (Z1), time-varying process measurements (X), and final quality data (Y). All of this information is analyzed jointly using multivariate analysis.
The batches are all of different duration, so alignment must first be performed on the process measurements. This critical step has always been one of the stumbling blocks to performing a batch analysis. ProSensus Multivariate (ProMV) has an alignment tool that allows users to synchronize their batch data using dynamic or linear warping. For this particular data, 3 stages are identified, as shown. In the first stage, the tank level is used to linearly align the data, in the second stage the dryer temperature is used, and in the last stage time is used. The amount of time each batch spends in each phase, and other features of the process data such as temperature slopes are recorded in a new data block (Landmarks, Z2).
Model Building and Analysis
Three of the models created to troubleshoot the process will be shown here. The first is a PCA model on the Y space. The second is a PLS model that includes the two Z blocks to predict the Y block, and the third contains all of the data.
PCA on Y Block
A PCA model on the Y space shows that there is clear separation between good, bad and high solvent batches. This indicates that the measured final properties sufficiently characterize the quality problem since good and bad batches can be distinguished effectively.
PLS on Z and Y Block
A PLS model to predict final properties using the two Z blocks is a good start for this analysis. The Landmark block summarizes the trajectory data and in some batch analyses this provides enough information to avoid the need to use the whole X block (i.e., the raw trajectory data). This model will show which variables have the greatest impact on poor quality outcomes.
When building multivariate models, the first step is always to exclude significant outliers. Using the Hotelling's T2 and SPE (squared prediction error) statistics, this step is made simple in ProMV.
After outlier removal, a model was built to predict the final quality from the two Z blocks. The model had a R2 of 55% and a Q2 of 29%. This is in fact a reasonable model for "happenstance" historical process data. The VIP and block importance plots show that process variables are the root cause of the bad batches rather than the initial chemistry, in contrast to the expectations of the plant personnel. Though this analysis shows us that the process variables are the culprit, it does not provide enough insight on where in the process the problems are occurring. An analysis incorporating the entire X block is needed for this.
PLS using all data
An analysis using all the data in one model is detailed here. The aligned trajectory data is unfolded batch-wise in ProMV. This means each observation (row in the dataset) contains all the information on a single batch. Given the large amount of data, 350 time-points for each of the 10 trajectories, it can be a challenge to interpret the data. However, several plots in ProMV make this possible. When looking at the score plot, it can be seen that good batches have a low value of T1. A loading plot of the trajectory variables shows the time behaviour of the measurements and which variables contribute to the quality problem and at which points in the batch progression. In order to have a batch with a low T1 (good quality), the batch should have a low tank level and low time duration for the whole batch, and high pressure, dryer temperature and jacket temperature in the first half of the batch. This informs the plant personnel on how to modify the standard operating procedure (i.e., the standard "batch recipe") to greatly reduce or eliminate the occurrence of bad batches.
1. S. Garcia-Munoz, T. Kourti, J.F. MacGregor, "Troubleshooting of an Industrial Process using Multivariate Methods", Ind. Eng. Chem. Res., 42, 3592-3601, 2003.