Back to Basics

Three Philosophies: Recalls, Statistics and NDT

by Emmanuel P. Papadakis^*

 

Recalls
In the manufacturing industries, the practice of recalling products is well known and (perhaps grudgingly) accepted. The philosophy behind recalls is less broadly disseminated but is very important for the consumer, the manufacturer and the nondestructive testing (NDT) expert asked to look into a problem. The philosophy of recalls, as explained to me by one legal specialist, is as follows: if a manufacturing problem is discovered in some part in the field - or even before it reaches the field - and if this problem could cause death or serious injury by failure of the part, then it is incumbent upon the manufacturer to take every action a reasonable person could be expected to take to assure that this one problematic part is unique. This philosophy is consistent with the requirements regarding safety of a number of fields and industries.

The first response of the manufacturer must be to determine what sets of parts might exhibit critical discontinuities or have other problems. These sets may be circumscribed by manufacturing dates, plant locations, supplier batches, "process out of control" occurrences or other parameters which could make a set of parts related. (The idea of "relatedness" was addressed in a previous "Back to Basics" article by me in the December 2002 issue of Materials Evaluation.) These related parts must then be located in the field (that is, the manufacturer must locate each product containing the parts) and replaced in a recall campaign. The replacement parts must be guaranteed to be free of problems. The cost of the work, including locating owners, notifying them, supplying new parts and repairing the endangered products, must be borne by the manufacturer. This can be extremely expensive: in the aircraft jet engine case cited in Papadakis (2002), the cost was $8 million in 1983 US dollars. Other examples of industrial recalls were described in Papadakis (1985).

---

There are several separate scenarios where NDT can be of help.

---

There are two timelines which are relevant when guaranteeing parts of adequate quality: an immediate timeline for the replacements and a long term (full life) timeline to assure that the one problematic part remains unique. For both timelines, the manufacturer is permitted to rely upon whatever technology is deemed most fit.

The manufacturer has the option to choose a mix of statistics and nondestructive testing to address any concerns regarding possible recalls. Choices always pose difficulties. One of the difficulties in this case is that both options have their advocates. These advocates tend to deprecate the qualities and/or efficacy of the other option. Management must understand both options as well as their concomitant limitations. The two options are explained here.

 

Statistics
Manufacturing engineers try to set up processes with adequate capability to provide parts which are free of any problems. Hidden in this statement, however, is a statistical limitation, because every process has a variability with a statistical bell curve of a certain width (Figure 1). This means that there will be some outliers beyond the acceptable specification limits even if the mean value during production is right on the design value. The current buzzword in statistical circles concerning process capability is "six sigma." Put simply, this means that the process engineers should strive to improve the process to achieve a capability which will make the bell curve of variability narrow enough such that 3σ (three process standard deviations) on each side of the mean will fit inside the specification limits of the part. Then, with the process centered, the fraction of nonconforming parts will be about one in a thousand on each end of the bell curve (usually one end is more detrimental than the other). Accomplishing 6σ is quite a feat inasmuch as manufacturing practice until recent years frequently operated on 2σ or 3σ capability with sorting of output to eliminate nonconforming material. 6σ seems to be the ultimate goal for process capability excellence. 6σ is what statisticians strive for. It should be noted that 1 in 1000 nonconforming is not the same as unique. In ten million parts during ideal production, the result is about 10 000 ± 100 bad ones.

Figure 1 - A bell curve plotting the numerical values in a population versus the number of times the value occurs, with indications for the standard deviation σ locations along the curve. In this, m is the mean for the population, with σ ± from m including 68% of the data, σ2± from m including 95% if the data and σ3± from m including 99.8% of the data. 

 

Production is not ideal to this degree, but has its own statistics. Indeed, production is often kept under control by statistical process control. In this discipline, samples of a few parts (typically five) are taken at frequent intervals (typically hourly) and measured. The mean of the production parameter being controlled varies up and down over time. This variability has its own bell curve and its own standard deviation, which is much smaller than that of the process capability. The σ for the manufacturing line is determined empirically while the line is running as accurately as possible. It is possible for the sampled manufacturing line process mean to fall outside one, two, or three standard deviations from the process control mean. There are "run rules" for this variability over time which indicate when a manufacturing process has gone out of control (Western Electric, 1956). "Out of control" means that the process must be stopped and repaired. The reason is that the process, upon going out of control, has begun producing inferior material at an alarming rate and will get worse.

However, the trigger point for detecting out of control processes is not the beginning of trouble. There are two other previous time periods posing difficulties for the "uniqueness" requirement. The first time period is the entire production run. As the sampled mean of the manufacturing process meanders up and down, the number of outliers beyond the specification limits invariably increases. If the process mean moves up so that one outlier beyond the lower specification limit is lost, then the curvature of the bell curve yields two or more extra outliers beyond the upper specification limit. (The whole bell curve moves with its mean, of course.) So, even while under statistical process control, the manufacturing process is producing two or three parts per thousand of nonconforming goods (or more) despite the "six sigma" process capability.

The second time period which poses difficulties is the few hours from the inception of the run rule which will detect the out of control condition until the actual detection of the condition when all the conditions of the run rule are met. One can only look at these run rules after the fact. One does not know if a beginning is true or a fluke. During this time, more nonconforming material (above 2 or 3 parts per thousand) can be produced but temporarily remain undetected. If it is shipped under the regime of "just in time" inventory control, the uniqueness requirement is flaunted. The proper procedure is to quarantine all production for at least the length of time of the longest run rule before shipment. This would put industry back into the currently deprecated condition of "just in case" inventory control.

For the replacement parts called for in the recall, the manufacturer relying upon statistics alone must make a judgment call about the small (but not negligible) number of nonconforming parts in the regular production which will be set aside to fulfill the recall. If 100000 parts are called for, how many of them will have problems even at the "six sigma" process capability level? What is the risk that a few more outliers will be added to the population which was supposed to have no outliers? How close will the manufacturer be to complying with the requirement that the one problematic part be unique? This is basically unknowable by statistics alone. Some sort of sorting regimen is needed.

 

Nondestructive Testing
Nondestructive testing and other types of metrology like laser gaging for dimensions can be used to address the issue of problematic parts numbered in parts per million. There are several separate scenarios where NDT can be of help. First, some background.

In general, NDT can be used on parts whose discontinuities or other problems can be correlated to NDT parameters. Examples are strength of nodular iron by ultrasonic velocity, hardness of iron and steel by eddy current impedance plane response, invisible internal discontinuities by ultrasonic echoes, adhesive bonding by ultrasonics or infrared and so on. Any of the so called "latent defects," to use W.E. Deming's terminology, can be addressed by NDT research to determine whether they can be actually detected by NDT (1982).

Several manufacturing scenarios where NDT may be utilized are given in the following.

	The replacement parts mentioned above may be amenable to sorting by NDT to eliminate the last few outliers.
	The measurements for statistical process control can be made by NDT techniques where physical properties and latent discontinuities are the parameters to be controlled. The machine operator can use the NDT equipment as any other caliper.
	In using NDT techniques in statistical process control, the NDT measurement can be automated. Then the measurement can be made on every part, not just on five every hour. Computer control can choose the five measurements at the end of every hour and carry out the statistical process control by algorithms. Feedback automatically goes to the manufacturing process. This process would eliminate the statisticians' objection to excessive reliance upon testing.
	With the NDT equipment and computers in place as in the above, the NDT measurements could eliminate all outliers. Shipment of product by the "just in time" process would fulfill the perfection requirement of uniqueness, there being no outliers escaping the factory. This regimen would be consistent with another current methodology, "in-process verification."
	All of these uses of NDT would be consistent with the requirements of ISO 9000:2000 concerning use of statistics, corrective action, remedial action and continuous improvement.

In the NDT realm, one question remains. What are the statistics of NDT? After all, NDT is governed by one crucial statistic, namely the probability of detection. Can the probability of detection be made high enough that the number of outliers escaping NDT detection would be much lower than the number of outliers shipped under ordinary statistical process control in a "six sigma" environment?

The answer to this must be determined experimentally by manufacturing feasibility studies on particular NDT systems proposed for the detection process. In certain cases already in production, the answer has been a resounding yes: NDT is better. One example which comes to mind is the assurance of nodular iron strength by ultrasonic velocity. This is the standard of the foundry industry. The history of this test is that there have been no product failures in hundreds of millions of critical parts. Another example is the detection of "chevrons" inside forward extruded steel parts. These invisible voids could break axles, for instance. Still another example is the use of low frequency eddy current response to detect soft iron in parking pawls in automatic transmissions.

In general with probability of detection, it is possible to make Type I errors as small as desired if a larger Type II error is acceptable. Type I errors are "calling bad material good" while Type II errors are "calling good material bad." Type I errors are desired to be zero, while Type II errors represent the economic burden of discarding good material. As it is not permissible to balance the cost of a life in an accident against the economic cost of a solution a reasonable person could undertake, the economic burden of discarding good material or salvaging it in some other way must not be the determining factor. NDT may well be the optimum solution. Experiments will determine the degree and cost of eliminating problematic material at a rate much better than the "six sigma" statistical methods.

Deming, W.E., Quality, Productivity, and Competitive Position, Cambridge, Massachusetts Institute of Technology, 1982.

Papadakis, E.P., "The Deming Criterion for Choosing Zero or 100% Inspection," Journal of Quality Technology, Vol. 17, No. 3, July 1985, pp. 121-127.

Papadakis, E.P., "Justification for Engine Parts Testing in Manufacture," Materials Evaluation, Vol. 60, 2002, pp. 1399-1400.

Western Electric Company, Statistical Quality Control Handbook, Newark, Western Electric Company, 1956.

 

* Quality Systems Concepts, Inc., 379 Diem Woods Drive, New Holland, PA 17557; (717) 355-2142; fax (717) 355-2142; e-mail <papadakis@desupernet.net>.

 

Copyright © 2003 by the American Society for Nondestructive Testing, Inc. All rights reserved.