Weibull – Analyses www.crgraph.com Visual-XSel
General Foundations

It has proven advantageous to assume the cumulative distribution of the failures as the basis for calculations. The distribution used in the Weibull calculation is especially suited to this field of application. In general the Weibull distribution is obtained through an exponential distribution. Calculations are done in this way because :

· Many forms of distribution can be created through the Weibull distribution.

· The Weibull functions are user-friendly.

· Time-dependant fault mechanisms are depicted on a line graph.

· The method has proven itself to be reliable.

When presented on a linear diagram, the failure probability of the entire vehicle life span will be shown in a complicated "S"-form. Distortion of the vertical scale (double logarithms) and the horizontal scale (logarithm) will enable this "S" to be shown as a straight line (fitting line). The Weibull distribution function in this simplified 2-variable form is known as Equation 1.

H = failure probability or failure frequency (standardised at 1 (%) in 100), t = lifetime variables (distance driven, length of time in operation, load cycles, etc.), T = typical service life, the point at which 63.2% of the units were faulty (for t = T, H =100% (1-1/e) = 63.2%) and b = the parameter form, i.e. the rise (slope) of the fault line within the Weibull chart. In the 3-parametrical form, the so-called fault-free time period to is taken into consideration (Equation 2).

In most cases to = 0 can be assumed. This corresponds with the 2-parametrical form. Despite stress, some component faults behave in such a way that they are first detected after the operation-time to. In this situation the co-ordinates on the Weibull graph will lie above the service life line most often curving to the right. When analysing the steep and falling left-hand curve, it is possible to pre-position the point of intersection of the curve with the infinite "zero-line" using to .

Data Preparation

Generally the reliability of components, assemblies and vehicles can be determined only once failures exist, i.e. once the service life of the unit in question is available. In order to be able to comment on reliability, one must first establish the service life e.g. during testing, in the laboratory or in the field. The service life line (variable t) is in most cases :

· Distance driven

· Length of time in operation

· Load cycles

· Number of load changes

In order for reliability analysis to be possible, data for at least one of the above must be available for the "damaged part".

This information is charted on the X-axis of the Weibull graph. For a spot check it is usually necessary to carry out classification of the failures from n>50, resulting in certain service life groups being placed together. Such classification results in a roughly even curve on the Weibull graph. The range of classification can be estimated using Sturges Equation 3.

In practice it is often appropriate to use a range of ‘000’s, rounding either up or down. This is especially useful in cases where field data in conjunction with kilometres is used for analysis, e.g. 1000 km, 2000 km, 5000 km etc.

Information is also lost through classification which means that there are slight deviations in the results between the different classes. The same parameters and or methods of classification should therefore be selected when attempting a comparison between various analyses.

The Weibull Diagram

If the damaged parts are sorted by increasing service life then it is very simple to establish the corresponding failure probability H using Equation 4 for values n < 50.

This means I = the ordinal number of the damaged parts, n = the spot check or number of defective parts. For n ³ 50 the mid-range value is used for calculations in accordance with Gumbel (Equation 5).

The simple relation H=100*i/n can frequently be found in the literature, but Equation 5 is more suitable to the lower area. This data can be entered directly into the Weibull chart .

Determining Weibull Parameters

In this classical interpretation, Weibull parameters are provided through defining the adjustment line on the Weibull failure probability line graph. The points for the adjustment line are obtained by altering the 2-parametrical Weibull-function. The relationship for the X-axis and the axis of ordinates is provided in accordance with Equation 6.

The scale is calculated using the acknowledged method of taking the lowest failure quadrant and the values for X and Y. This scale corresponds to Equation 7.

b is the rise and b.In(T) is the vertical position of the adjustment line.

In practice b and T are often calculated using the Gumbel method, in which the points on the Weibull chart are weighted differently (Equations 8 and 9).

Through the Gumbel method, the values provided for b are greater than those given when using the standard method. This should be noted when interpreting the results.

A further possible way of determining points b and T is the Maximum-Likelihood Estimate (Equation 10):

In order to ascertain b Equation 10 must be repeatedly solved. Once b is obtained T can be calculated (Equation 11)

Interpretation of results

Due the regular dispersion in statistics for the service life, it is clear to see that there is little point in inputting the average figure for the "running time". The Weibull evaluation reaches a sufficient conclusion regarding the cause of failures of the relevant part. The figure at which the average lies is usually taken as the typical-service-life T, at which point 63.2% of the parts show faults. This is most often illustrated with the perpendicular in the graph.

A further important measurement is the parameter form b. This is nothing else but the rise in the line on the Weibull graph. Conclusions can be drawn on the behaviour of the faults by using rise b. Usually the Weibull chart will show a scale on the right-hand side which includes rise (slope) b. Taking the guiding line (bottom right) and the increase in the Weibull line, it is possible to make a simple estimate as to the real increase in the faults.

In practice it is often the case that the fault data has a variety of causes. This means that, after a certain "running time", there is a pronounced change in the fault increase due to other influences on the part also being present. The various sections should therefore be looked at separately and it is advantageous to connect the individual points on the graph (faults) to one another when an entire fitting line is available.

Before a part goes into production it is necessary to have a prediction indicating its reliability. In order to obtain this prediction, long-term tests and simulation tests are done under laboratory conditions, in which the part is put under a certain load stress in order to obtain an estimate of its life span. This is most often required due to safety reasons, e.g. factor 2..3. Entering the service life characteristics into the Weibull-Chart will result in the fitting line lying to the left of a test line in which a normal load had been used. If the progression of the fitting line is not parallel, but increases at a different rate, then this shows a variety of fault mechanisms in relation to the test and real applications. The test is therefore not suitable.

General evaluation problems

In cases where a calculation of faulty parts is required for parts already in production cars (so-called in-field faults), then the failure probability is calculated in the same way as previously shown. The relevant production figure is taken along with the production period for that figure. The fault figure is then used in relation with this information.

It is clearly necessary to presume that all faults that relate to this production figure have been recorded and that none have been exchanged due to incorrect findings. In the case of incorrect findings, parts are wrongly exchanged by customer services although they are not the actual reason for the fault. These parts have no failures are cannot be counted within the failure probability, therefore they are not included in the analysis. Apart from this it is important to note the service life. It is pointless to include parts in the analysis which have failed due to other influences (for example due to an accident). This means that damage analysis should always be carried out before the individual data analysis begins.

The confidence range

The Weibull calculation is based on what may be viewed as a random sample. This in turn means that the line on the Weibull chart reflects a sample. The points on the Weibull line will become increasingly scattered as the number of parts included in the calculation rises. It is however possible to make a statistical estimation as to the area on the chart which will make up the basic entirety of the failure probability. A so-called "confidence range" is taken in this case. This is generally given with a statement probability, mostly at PA=90%. This means that the upper confidence range is at 95%, and the lower at 5%. The confidence range is calculated using a Beta-distribution with the relevant rank numeration as parameter. Tabled values are generally available for the F-distribution. The confidence range may be confirmed through a transformation using this distribution (Equations 12 and 13).

The example shown in the Diagram shows the two limit lines, within which the 90% basic failure entirety may be found. There will, to a varying extent, be a split between the path of the failure probability line and the upper and lower lines which define the limits of the confidence range. This shows that the failure probability prediction is far less accurate within these areas than in the upper middle section of the chart. It is not possible to lengthen the confidence range, or the fitting line, so that either lies very far outside of the co-ordinate points plotted on the chart.

Confirmation of fault-free period t_o

There are various methods used to confirm to. One mathematical formula which is clearly best suited in this case does not exist. One can always say that to must lie between time point 0 and the time at which the first part became faulty. to usually falls very close to the first occurrence of a faulty part. The following method is possible : apply to in small, regular steps in the interval from 0 up to the first fault (Tmin) and calculate the coefficient correlation to the fitting line. The better the value of the coefficient correlation, the more precise the points on the Weibull chart line. to is the greatest value and therefore permits a good approximation with the fitting line. The confirmation of to is described later on in this text.

Comparison of 2 distributions

In cases where 2 different designs or systems are to be compared, perhaps with different finishing processes, then the typical service life can be examined. The relevant service life is often given at the point where a 10% failure probability has been reached. A comparison between the service life values alone is not enough because a statistical statement on the likelihood of further faults is not included. An example: In a test, design A is compared with design B which has a longer service life. For example: A design “A” around 5 units at 50% service life, whilst “B” has a considerably longer service life, showing ca. 6.5 at this point. A response to the failure probability in this case would therefore be that B is better than A. Additionally to this information, a relationship is shown as (Equation 14…17):

In this case z, y, y’ = help-functions for obtaining the statement of probability, q = sum-%-failure area, t_q,a = service life of design A at q-%-failures, t_q,b = service life of design B at q-%-failures, b_A = increase parameter of the Weibull distribution for design A, b_B = increase parameter of the Weibull distribution for design B, n_A = sample size for design A, n_B = sample size for design B. AW = Probability of different distributions. Is AW = 50% there is no difference between the two curves.

See also Consideration of faults not yet occurred

Software for Weibull -> Visual-XSel www.crgraph.com