1.1 Overview of Current Situation in Methodological Aspects of Reliability Prediction
Because the problem of reliability prediction is so important there are many publications in the area of the methods of reliability prediction, mostly in the area of electronics. This is especially important when the reliability prediction is necessary to provide a high degree of effectiveness for the products.
Most of these traditional methods of reliability prediction utilize failure data collected from the field to estimate the parameters of the failure time distribution or degradation process. Using this data one can then estimate the reliability measures of interest, such as the expected time to failure or quantity/quality of degradation during a specified time interval and the mean time to failure (or degradation).
Reliability prediction models that utilize accelerated degradation data, rather than those models that utilize degradation data obtained under normal conditions, are often performed because the degradation process is very slow under normal field conditions and prediction modeling based on normal conditions would require excessive time.
One example of such practices is the procedure used by Bellcore for both hardware and software. O'Connor and Kleyner [1] provide a broad description of a reliability prediction method that emphasizes reliability prediction consisting of outputs from the reliability prediction procedure:
- steady‐state failure rate;
- first year multiplier;
- failure rate curve;
- software failure rate.
The output from the reliability prediction procedure is various estimates of how often failure occurs. With hardware, the estimates of interest are the steady‐state failure rate, the first year multiplier, and the failure rate curve. The steady‐state failure rate is a measure of how often units fail after they are more than 1 year old. The steady‐state failure rate is measured in FIT (failures per 109 h of operation). A failure rate of 5000 FIT means that about 4% of the units that are over a year old will fail during following one year periods. The first year multiplier is the ratio of the failure rate in the first year to that in subsequent years. Using these, one generates a failure rate curve that provides the failure rate as a function of the age (volume of work) of the equipment. For software, one needs to know how often the system fails in the field as a result of faults in the software.
The Bellcore hardware reliability prediction is primarily designed for electronic equipment. It provides predictions at the device level, the unit level, and for simple serial systems, but it is primarily aimed at units, where units are considered nonrepairable assemblies or plug‐ins. The goal is to provide the user with information on how often units will fail and will need to be replaced.
The software prediction procedure estimates software failure intensity and applies to systems and modules.
There are many uses for reliability prediction information. One such example is, it can be used as inputs to life‐cycle cost or profit studies. Life‐cycle cost studies determine the cost of a product over its entire life. Required data include how often a unit will have to be replaced. Inputs to this process include the steady‐state failure rate and the first year multiplier.
The Bellcore reliability prediction procedure consists of three methods [1]:
- Parts count. These predictions are based solely by adding together the failure rates for all the devices. This is the most commonly used method, because laboratory and field information that is needed for the other methods is usually not available.
- Incorporating laboratory information. Device or unit level predictions are obtained by combining data from a laboratory test with the data from the parts count method. This allows suppliers to use their data to produce predictions of failure rates, and it is particularly suited for new devices for which little field data are available.
- Incorporating field information. This method allows suppliers to combine field performance data with data from the parts count method to obtain reliability predictions.
Mechanical reliability prediction [2] uses various stress factors under operating conditions as a key to reliability prediction for different devices. This situation is more common in testing mechanical devices, such as bearings, compressors, pumps, and so on than with electronic hardware.
Although, the number of factors that appear to be needed in reliability testing calculations may appear excessive, tailoring methods can be used to remove factors that have little or no influence, or for which limited data are available. Generally, the problems encountered in attempting to predict the reliability of mechanical systems are the lack of:
- specific or generic failure rate data;
- information on the predominant failure modes;
- information on the factors influencing the reliability of the mechanical components.
The mechanical reliability prediction approach can be useful if there is a close connection with the source of information for calculation reliability during the desired time, whether that be warranty period, service life, or other defined period. Obtaining accurate initial information is a critical factor in prediction testing.
Unfortunately, the approaches described earlier, and other prediction methods, provide little guidance on how one can obtain accurate initial information that simulates real product reliability over time or amount of use (volume of work, or duty cycle). Without accurate simulation information, its usefulness is minimal.
Proper understanding of the role of testing and the requirement to do this testing before the production and use of a product is critical and can easily lead to poor product reliability prediction that will negatively impact financial performance.
Prediction is only useful if it reduces early product degradation and prevents premature failures of the product.
There are many recent publications addressing electronics, automotive, and other product recalls. While they usually address reliability failures in terms of safety matters that affect peoples lives by contributing to deaths or injuries, they may also consider economic impacts.
As was mentioned previously, such reliability and other problems are results, not causes. The actual causes of these recalls, and many other technical and economic problems, were a direct result of the inefficient or inadequate prediction of product reliability during the design, and prior to, the manufacturing process.
In the end, it is poorly executed prediction that negatively impacts the organizations financial performance.
Therefore, while many popular and commonly used approaches appear to be theoretically interesting, in the end they do not successfully predict reliability for the product in real‐world applications.
Consider the consequences of the recalls of Takata automobile air bag inflators [3]: