Environmental monitoring with Process Screening

[#1]

Every pharmaceutical manufacturer needs to monitor its manufacturing facilities for possible contaminants from the environment. Mold from airborne spores and bacteria on surfaces are two of the threats that can ruin a batch. Other kinds of factories, like in the semiconductor industry, are threatened by airborne particulates. These facilities have specialized collection devices for environmental monitoring. Each monitoring device reports many times a day, and there are usually many different devices across each facility. It becomes a challenge to process all this data in a meaningful way.

Handling large amounts of process data efficiently is the special talent of the Process Screening platform in JMP, and in JMP 19, this platform has been empowered to handle the kinds of non-normal data produced from environmental monitoring.

Prefer video? Watch it here.

Environmental monitoring data generally occurs in two types: count data and non-negative continuous data, neither of which has a normal distribution. With these types of data, the platform looks at the data and fits several distributions appropriate to each type, selecting the best fit. For count data, it considers the Poisson, negative binomial, and the zero-inflated variations of these distributions. For non-negative continuous data, it considers exponential, Gamma, Weibull, lognormal and the zero-inflated variations of these four distributions.

After choosing the best-fitting distributions, it then estimates two quantiles of that distribution to use for alarms. An action limit, which typically uses the 99.85% quantile, signals if immediate action is needed. An alert limit, which typically uses the 97.5% quantile, produces a warning for paying close attention to determine if a problem needs to be ameliorated. These alerts correspond to the 3-sigma and 2-sigma limits traditional to control chart design. There are also tests for increasing contamination past the alert limit.

Historical data can be processed to fit the distribution, calculate quantile estimates, and set proposed limits to store in a limits table to be used in production runs on new data.

For example, consider a set of simulated environmental monitoring data with 7,740 samples from 86 different sampling sites of several different types.

The scrollable summary table shows each process, sorted by the action count. The best-fitting distributions are a mix of negative binomial and zero-inflated negative binomial. Selecting a process shows the control chart with an action alarm in late January. There are many zeroes in the process, which explains why it chose a zero-inflated distribution. The action limit was set to 16, the 99.85% quantile of the fitted distribution. An alarm graph shows the six processes with alarms by time, including two that are close the same time as the selected process, a hint that there may have been cross-contamination between monitoring sites.

The Process Screening platform can publish its report to JMP Live in an especially interactive mode, where clicking on the process in the summary table or the alarm graph brings up the control chart for that process.

The features that make Process Screening in JMP 19 particularly valuable are:

• It fits a variety of distributions, not just the Poisson distribution typical in c-charts.
• It uses the standard language of Action Limits and Alert Limits for environmental monitoring.
• It can use a separate table of Action and Alert Limits, and if so, calculates a non-normal Ppu capability index.
• It can sort the processes by the features of most interest, such as action count, bringing into focus where attention is needed.
• Clicking on a process immediately brings up the chart of that process.
• An alarm graph shows all the alarms by time, allowing you so see patterns where several processes signal close to the same time.
• The report can be published to JMP Live in fully interactive form.
• The platform is very fast, able to analyze thousands of processes almost instantly.