Variance doesn’t always mean that adjustments should be made.

by: Dr. Jeff Gwirtz

In my last article from the Third Quarter issue of Milling Journal, “Monitoring an Automated Tempering System,” I wrote about the challenge of limited manually collected data availability.

This article explores the importance of multiple measures to better estimate the average moisture and the standard deviation, which helps to better represent moisture uniformity in an automated or manual tempering system.

A word about the data presented in this as well as the previous article: The observations reported are based on the collection of more than 30 samples from a commercial mill. The sample population data was used to calculate the average and standard deviation for the sample.

Using a random number generating function, the sample statistics were used to generate 1,500 observations for a normal distribution. The number 1,500 was selected as it allows plotting of a complete looking, normal frequency distribution as presented in Figure 2 of the previous article. Using too few generated numbers results in a fragmented graph as the numbers presented are really pseudo-random numbers because of a formula seeding.

Returning to a plot of 1,500 dry wheat moisture observations shared in the previous article, Figure 1 shows those observations with an average and standard deviation of 10.20% and 0.23%, respectively. Each point on the graph represents a single dry wheat moisture observation. The red horizontal red line shows the average dry wheat moisture for the 1,500 data points shown. Observations are concentrated along the average dry wheat moisture (the horizontal red line at 10.20% moisture) with the frequency or concentration of observations thinning out the further below or above the average dry wheat moisture value.

The green lines show ±1 standard deviation from the average in which approximately 68% of the observations are observed. The green lines show ±2 standard deviations from the average in which approximately 95% of the observations lie. Assuming these observations were collected over time, there appears to be no systemic shift in the average as frequency concentration or density of observations consistently centers along the average, no moving up or down in a patterned form. Such a pattern indicates perhaps a change in the dry wheat mix moisture that should be investigated and resolved quickly to maintain uniformity.

Understanding Moisture Variance

Dry wheat mix variation occurs for many reasons. In every wheat field, there is wheat moisture variation due to field location and topography.

Weed components in harvested wheat may contribute to elevated moisture content as well as dockage and foreign material.

Field topography and surrounding growth impacts uniformity of drying due to temperature, soil moisture, sunlight, and wind exposure variability during wheat ripening.

Once harvested and stored, wheat from various production locations is combined for storage or shipment. Despite what some believe, the combining of various production lots as wheat moves though the transportation and storage chain does not necessarily constitute an intentional blending step with the goal of enhanced wheat uniformity. The inherent variability of wheat delivered to the mill is real and management essential for process control and efficiency.

Given inherent wheat moisture variability, blending of multiple wheat lots containing different moisture content can be impacted by wheat location within the bin. It is well known for example that wheat discharging from the bottom of a storage bin can bring significant variation and challenges to preparation of a mill mix or grist. Often, the wheat at the bottom of the mass has been shattered by a significant fall against the bin bottom resulting in broken wheat, fines, and screenings accumulation due to segregation and undoubtedly moisture content.

It has been observed that wheat discharged from the bin bottom can impact flow control devices to the point of stopping wheat flow, thus changing the wheat mix proportions unless quickly discovered and corrected. Even when pulling from multiple bins of the same wheat lot, millers are encouraged to blend from the bins at different rates to avoid reaching the bottom of all bins selected simultaneously.

Understanding Variance

Figure 2 displays the same data as Figure 1, but presented with a solid line drawn between the observations. With a connecting line drawn, space is filled in along the average value due to observation frequency close to the mean or average value. Values observed further way appear as spikes and are less frequent away from the center line.

Some millers behave as if a single moisture measurement at one point in time represents the overall average since the precious measurement. Should the measurement be higher or lower, then target process adjustments are made resulting in even greater product variability.

The practice of relying on a single measurement is flawed in that it does not consider variability that is a processing reality. A single measurement offers no indication of the population moisture from which the sample was drawn.

Multiple measures are needed to properly estimate the mean and standard deviation of any product attribute. The estimate of mean and variation or range is a critical step in establishing upper and lower control limits for average or mean and range a measure of variation, to establish upper and lower control limits needed to determine if the process is in control or not. In a flour mill such analysis is often employed in the packaging room to control and manage scale performance to prevent over and underweight flour packages.

Two more simple methods of data analysis are shown in Figure 3 at left and Figure 4 on p. 18. Using the same data shown in Figure 1, the running average from in Figure 3 shows at each point a running average of five. The average of the first five individual observations (observation 1-5) is calculated and reported as the first running average value.

The next running average value reported is calculated by dropping the very first individual observation used in the first running average and picking up the next individual observation beyond the last individual observation used in the first running average (observation 6) and calculating the average of the new running set of five.

That process is repeated along the entire data set as shown in Figure 3. As can be clearly observed, plotting a running average provides a smoother, more tolerable view of dry wheat moisture measurement centered around the average identified in red.

Figure 4, which also uses the same data from Figure 1, shows the average based on discrete groups of five in which there is no overlap of individual observations used to calculate the mean or average value reported. The average of the first five individual observations (observation 1-5) is calculated and reported as the first group of five. The average of the second group of five reported is calculated using the next five values (observations 6-10).

The reported average values calculated unlike those of the running average are independent one from the other as the groups contain no shared individual observations. As suggested earlier, the difference between the high and low value within the group provides an estimate of variability. The estimate of variability can be used to establish upper and lower control limits for both the mean value and the range to assess process control. In the case of dry what moisture movement in the average value is adequate to assess clean dry wheat for tempering. Control of the tempering process is best monitored using both average and variation.

Figure 4 showing the average of five observations in group appears to provide a plot even closer to the mean line in red than what it is shown in Figure 3. Clearly, either the running average of five or groups of five provide a better estimate of the mean moisture content passing through the system.

It is expected that individual observations appear scattered as they are part of a population whose average is not known without collecting additional observations and performing simple calculations which can be easily handled using software.

As indicated earlier, the observations reported were not shown to have been taken over a specific length of time. Table 1 below provides some thought for consideration when looking over the 1,500 data points shared. The top portion of the table identifies a potential sampling frequency for an online dry wheat moisture measuring system, ranging from taking a measurement every 3 seconds to 60 seconds in various time increments.

Manual vs. Online Monitoring

The 1,500 observations shown cover anywhere from 75 minutes to 25 hours of operation if measurements are made once per minute or every 60 seconds. Undoubtedly, manual moisture measurements could not possibly be made at such a rate; however, online monitoring at such a frequency is essential if there is to be any hope of monitoring and managing short runs of uniquely different wheat mixes.

The bottom portion of Table 1 identifies a potential manual sampling schedule of every 0.5, 1, 2, 4 and 8 hours during an eight-hour shift. The 1,500 observations shared cover the production hours shown, given the sampling schedule.

The manual collection of moisture data for control purposes especially during short run times is inadequate to offer reasonable control. Perhaps manual sampling frequency can be used to check for gross fluctuations, discrepancies, and to manage moisture addition off-set for tempering system set up. Consider the amount of dry wheat passing through the system between samples. Determine what frequency of sampling affords some degree of comfort with product uniformity.

Recording and responding to single observations as a process control technique is flawed and likely contributes to increased process variability. Additionally, the technique fails to maximize benefit from the cost of collecting data either through automation or manually.

Analysis is an essential step in the reporting process needed to provide an opportunity to properly respond and maximize resource potential. How are you using your data collection?

Dr. Jeff Gwirtz is CEO of JAG Services, Inc., an international consulting company in Lawrence, KS; 785-341-2371;
He also is adjunct professor in the Department of Grain Science and Industry at Kansas State University, Manhattan.