Particle Excursion Data Analysis: From Trend Plot to Predictive Maintenance

Predictive maintenance for cleanroom particle excursions is the difference between a program that responds to events and a program that anticipates them. The data is already being collected. The analysis is what turns the data into a maintenance schedule, and the maintenance schedule is the input to the cost of operation. A program that analyzes the data and acts on the signal is a program that runs at a lower excursion rate, and the lower excursion rate is the input to the customer’s confidence.

This article is a working guide to building a predictive maintenance program from the particle trend plot. It assumes the cleanroom is already running with a defined excursion response procedure, a working data integrity program, and a risk-based sampling plan. The predictive maintenance program is the analysis layer that sits on top of those three programs and turns the data into action.

What Predictive Maintenance Is and What It Is Not

Predictive maintenance is the use of trend data to schedule a maintenance action before the action is required by an event. In a cleanroom, the trend data is the particle concentration over time, and the maintenance action is a filter change, a fan replacement, a gasket inspection, or a cleaning rotation. The action is scheduled based on the rate of change in the trend, not on a calendar or on a failure.

Predictive maintenance is not a calendar-based maintenance program. A calendar-based program replaces a filter every 12 months regardless of the trend. A predictive program replaces the filter when the trend shows the filter is approaching the action limit, which may be 8 months or 18 months. The cost saving is real, and the cost saving is the input to the program’s justification.

The Three Signals in the Trend

The particle trend has three signals, and the three signals are the input to the maintenance schedule. The first signal is the absolute level. The second signal is the rate of change. The third signal is the variance. Each signal is a different kind of maintenance trigger, and each signal is a different kind of cost saving.

1. The absolute level. The absolute level is the particle concentration at a point in time. The absolute level is the input to the action limit. A trend that is consistently above the action limit is a signal that the maintenance is overdue, and the maintenance is the filter change or the fan replacement.
2. The rate of change. The rate of change is the slope of the trend over a defined period. A trend that is increasing at a rate of 10% per month is a signal that the filter is aging, and the maintenance is the filter change before the trend reaches the action limit. The rate of change is the input to the predictive schedule.
3. The variance. The variance is the spread of the data around the trend. A trend that has a high variance is a signal that the system is unstable, and the maintenance is the gasket inspection, the duct inspection, or the door inspection. The variance is the input to the root cause analysis.

The three signals are linked, and the link is the trend plot. A trend plot that has the three signals annotated is the input to the maintenance schedule, and the maintenance schedule is the input to the cost of operation.

The Rolling Average: Smoothing the Noise

The raw particle data is noisy, and the noise is the enemy of the predictive analysis. The corrective action is the rolling average, which is the mean of the data over a defined window. A working rolling average has the following elements:

1. The window size. The window size is the number of data points that are averaged. A 7-day window is typical for ISO 5, a 30-day window is typical for ISO 7. The window size is defined in the data analysis SOP, and the SOP is the input to the trend plot.
2. The data source. The data source is the LIMS, not the spreadsheet. The LIMS is the system of record, and the rolling average is generated from the LIMS, not from a separate spreadsheet.
3. The visualization. The visualization is the rolling average plotted on the trend plot, with the raw data plotted as a scatter plot behind the rolling average. The two are linked, and the link is the input to the analysis.
4. The annotation. The annotation is the rolling average annotated with the maintenance events, with the events linked to the corresponding record.

The four elements are linked, and the link is the trend plot. A trend plot that has the four elements is the input to the predictive analysis, and the predictive analysis is the input to the maintenance schedule.

The Rate of Change: The Predictive Trigger

The rate of change is the predictive trigger, and the trigger is the slope of the rolling average over a defined period. A working rate-of-change calculation has the following elements:

1. The period. The period is the number of days over which the slope is calculated. A 30-day period is typical for ISO 5, a 90-day period is typical for ISO 7. The period is defined in the data analysis SOP, and the SOP is the input to the trigger.
2. The calculation. The calculation is the linear regression of the rolling average over the period. The slope is the change per day, and the slope is the input to the trigger.
3. The threshold. The threshold is the slope that triggers a maintenance action. A threshold of 5% per month is typical for ISO 5, a threshold of 10% per month is typical for ISO 7. The threshold is defined in the data analysis SOP.
4. The action. The action is the maintenance action that follows the trigger. The action is documented in the maintenance SOP, and the SOP is the input to the work order.

The four elements are linked, and the link is the predictive trigger. A trigger that has the four elements is the input to the maintenance schedule, and the schedule is the input to the cost saving.

The Variance: The Stability Signal

The variance is the stability signal, and the signal is the standard deviation of the data around the rolling average. A working variance calculation has the following elements:

1. The window size. The window size is the same as the rolling average window. The window is the input to the calculation, and the calculation is the standard deviation.
2. The threshold. The threshold is the standard deviation that triggers a stability investigation. A threshold of 20% of the rolling average is typical for ISO 5, a threshold of 30% is typical for ISO 7. The threshold is defined in the data analysis SOP.
3. The investigation. The investigation is the documented root cause analysis, with the data reviewed, the system reviewed, and the procedure reviewed. The investigation is closed within the response time defined by the tier.
4. The corrective action. The corrective action is the documented CAPA, with the owner, the target close date, and the close-out evidence.

The four elements are linked, and the link is the stability signal. A signal that has the four elements is the input to the root cause analysis, and the analysis is the input to the corrective action.

The Filter Life: A Predictive Model

The filter life is a predictive model, and the model is the input to the filter change schedule. A working filter life model has the following elements:

1. The baseline. The baseline is the rolling average at the time of the filter installation. The baseline is the input to the model, and the model is the difference between the current rolling average and the baseline.
2. The projection. The projection is the rolling average projected forward to the action limit. The projection is calculated from the rate of change, and the projection is the input to the filter change date.
3. The confidence interval. The confidence interval is the range of the projection, calculated from the variance. The confidence interval is the input to the buffer, and the buffer is the time between the projected filter change date and the scheduled filter change date.
4. The trigger. The trigger is the buffer, with the buffer typically set at 30 days. A buffer of 30 days means the filter is scheduled for change 30 days before the projected action limit date.

The four elements are linked, and the link is the filter life model. A model that has the four elements is the input to the filter change schedule, and the schedule is the input to the cost saving.

The Cost Saving: The Justification

The cost saving is the justification for the predictive program, and the saving is the difference between the predictive cost and the calendar cost. A working cost saving analysis has the following elements:

1. The calendar cost. The calendar cost is the cost of the filter, the labor, and the downtime, summed over a defined period (typically one year). The calendar cost is the input to the comparison.
2. The predictive cost. The predictive cost is the same items, but with the filter change date determined by the predictive model. The predictive cost is typically lower than the calendar cost, because the filter is changed when it is needed, not when the calendar says.
3. The excursion cost. The excursion cost is the cost of the production loss, the investigation, and the corrective action, summed over the same period. The predictive program reduces the excursion cost, because the filter is changed before the excursion.
4. The net saving. The net saving is the calendar cost plus the excursion cost, minus the predictive cost plus the program cost. The net saving is the input to the program justification.

The four elements are linked, and the link is the cost saving analysis. An analysis that has the four elements is the input to the program justification, and the justification is the input to the budget.

The Semiconductor Equivalent: Wafer Defect Correlation

Semiconductor R&D and pilot lines do not have a regulatory mandate to follow predictive maintenance, but the principle is the same. The trend data is the particle concentration over time, and the maintenance action is the filter change or the tool maintenance. The wafer defect correlation is the input to the cost saving, and the cost saving is the input to the budget.

A working semiconductor predictive program has the same four elements (rolling average, rate of change, variance, filter life model) as a pharmaceutical program, with the addition of the wafer defect correlation. The wafer defect correlation is the comparison of the particle trend to the wafer defect trend, and the correlation is the input to the maintenance schedule. A filter change that is correlated with a wafer defect reduction is a filter change that is justified by the data, and the data is the input to the budget.

Closing: The Trend Plot Is the Maintenance Schedule

The mental shift that makes the predictive program work is to stop treating the trend plot as a record of past events and start treating it as the maintenance schedule. The data is already being collected. The analysis is what turns the data into the schedule, and the schedule is what turns the data into the cost saving. A working rolling average, a working rate of change, a working variance, and a working filter life model are the four elements that turn a record into a schedule, and the schedule is the input to the next quarter’s budget.

If you are building a predictive maintenance program from scratch, or reviewing an existing one, we can share a draft trend analysis template, a rate-of-change calculation worksheet, and a filter life model template, typically within two business days. Reach out with your current LIMS, your current filter inventory, and the date of your most recent filter change.