Developing an objective FMEA

Alessandro Vasta

04/24/2021

The problem of Risk Priority Numbers

The comments in this document would have the ambition to be an extension to the article written by Donald
J. Wheeler with title “Statistical Process Control”. This article finally point out a problem of many FMEA
developments where the classical data referred as Severity (S), Occurrence (O) and Detectability (D) have all
assigned numerical values but the related numerical fields are effectively simply considered as ordinal scale
data. In other words, it is only possible to establish an order relationship between each value af a numerical
field, (that is 2 > 1), but without ratio scale references to provide to this numbers an objective meaning.
Nevertheless, the numbers representing S, O and D are multiplied together to obtain a function called Risk
Priority Number RPN in this way:

RP N = S ∗ O ∗ D (1)

This is a problem, because as pointed out by the above mentioned article, if the variable S, O and D have no
ratio scale and without an absolute zero, the operations of addition, subtraction, multiplication and division
are all meaningless operations. Then, the next section deals with the possibility of defining all the components
representing the risk function in such a way to allow a risk calculation with a mathematical meaning.

The Risk function

In the standard ISO14971 “Application of risk management to medical devices”, the risk is defined in this way:
combination of the probability of occurrence of harm and the severity of that harm. The above standards
accepts various methods to estimate risks, where each component: severity (S) and probability (P) could
belong to a qualitative or, if possible to a quantitative scale. When a full qualitative approach is adopted,
both S and P are represented by a finite number of levels, each of them defined by a description, then finally
defining a category potentially collecting multiple cases. This is the situation where S and P belongs to
ordinal scale data fields, for this reason the risks are analyzed by a so called risk chart, where all combinations
of severities and probabilities are represented on a matrix and each combination of risks is analyzed case by
case. This approach is reasonable if the number of levels is low as in the following example.

1
The qualitative approach

Table 1: Severity of the final effect

Severity (S) Description
Significant Death or loss of function structure
Moderate Reversible or minor injury
Negligible Will not cause injury or will injurie slightly

Table 2: Probability of occurrence

Probability (P) Description
high Likely to happen, often, frequently
medium Can happen, but not frequently
low Unlikely to happen, rare, remote

Table 3: Risk Control Chart

Severity (S)
Probability (P) Negligible Moderate Significant
high R1 R2
medium R4 R5,R6
low R3

In the above example, a possible risk evaluation criterion could imply that only the Risk R1 and R3 are
judged acceptable while the remaining should require further mitigation. However, there are situations with
much higher complexity, where an higher resolution for severity and probability is required. In particular,
if the probability is split in two components O & D and also the severity S, are defined by 10 levels as in
the referred article, we should think to take a decision comparing 1000 different categories. Furthermore,
as correctly pointed out by D.J. Wheeler, the resulting RPN = S * O * D cannot be considered a priority
index, than even if S, O and D are represented by ordinal scale data fields, RPN is not an ordinal scale,
so if we want to perform a risk evaluation, we have to compare case by case all 1000 combinations!. For
some situations,this could not be the best approach, furthermore the comparison between close risks could be
affected by subjective interpretations.

A possible quantitative approach

Therefore, in order to achieve objectivity, a theoretical solution could be to adopt a full quantitative approach.
Concerning the probability of occurrence and referring it with P’, we could obtain it with a ratiometric
method. For instance, in medical contexts, the maximum likelihood estimation of P’ could be calculated
with the ratio n/N, where n = number of recorded failures/errors over N = number of treatments performed.
Moreover, always keeping the ratiometric properties, we could split P’ in two components as:
P’ = O’ * D’, where:
O’ is the frequency of occurrence of a failure mode and D’ is not defined just as a generic difficulty of detection
of the failure mode, but rather
D’ is the probability to fail the final effect prevention, given the failure occurrence.
In this way the computation of P’ is consistent, provided that the probabilities O’ and D’ are independent
and this is what generally happen.
Now, in order to achieve a full quantitative calculation of risk we should get a quantitative representation of
the severity, and this is what generally does not happen. A typical situation could be shown adding a column
to the previous Table 1, including an ordinal scale data column as in following Table 4:

2
 Table 4: Severity of the final effect
Severity Description S (ranking)
Significant Death or loss of function structure 3
Moderate Reversible or minor injury 2
Negligible Will not cause injury or will injurie slightly 1

S is an ordinal number assumption without any ratiometric meaning. So, if we perform now R = S * P’ or R
= S * O’*D’, we cannot in general achieve a ratio scale data for R, because there is no reason to assume that
the severity Moderate (S = 2) is two times the severity Negligible (S = 1).
Therefore, instead to assume an arbitrary ordinal scale for S, it can be calculated in this way:
• Given the lowest severity level S1 and assuming S1 = 1
• We indicate with Si all the remaining severity levels with i = 2,. . . ,m
• We determine P’i as the highest probability value that is still considered acceptable for all specific
severity level Si . Higher values are considered not acceptable, while lower or equal values are acceptable.
This probability value is the consequence of a decision that is in general context dependent. In reliability
contexts could be related to repair cost, while in safety context, like in medical devices risk analysis, P’i
could be in relation to the lowest probability technically achievable by the generally acknowledged state
of the art, provided that there is a therapeutic benefit to outweigh the risk. The determination of the
m values of P’i should be the crucial part for any development of risk analysis, on the basis of these
assumptions all risks are calculable and related to each other (see next point).
• We determine any Si with the following equation:

P10
Si = (2)
Pi0

In this way for any severity level Si , the corresponding risk at the probability P’i it will be always
aligned to:
R(Pi0 )i = Si ∗ Pi0 = P10 (3)
Therefore for all severity levels, the maximum acceptable risk is always the same value P’1 , while for all
the other probability levels, the risks are calculated in ratiometric way by R = S * P’, where S = Si are
defined by equation (2) and the acceptability criterion.
Replacing P’ with O’ * D’ and applying the above quantitative criterion, we can finally rewrite equation (1)
in this way:
RP N = S ∗ O0 ∗ D0 (4)

Now, RPN has finally this double meaning:

• is a quantitative risk estimation
• is a priority index
Unfortunately the full quantitative approach described so far could imply difficulties in accurate probabilities
estimations, especially in complex systems where failure and errors come from very different situations. In
some cases is not possible a numerical probability estimation, but rather only assumptions, like associating to
a potential fault probability only an order of magnitude. So a more suitable approach, also proposed by ISO
24971 guidance, is to use a semi quantitative scale for probabilities.

A possible semi quantitative approach

In this case, the Probability of occurrence could be defined by ranking numbers, where each value represent
an actual probability interval as reported in the following table:

3
 Table 5: Semi-quantitative Probability of occurrence
Common terms Probability range P (ranking)
Frequent 10−3 < P 0 ≤ 10−2 5
Probable 10−4 < P 0 ≤ 10−3 4
Occasional 10−5 < P 0 ≤ 10−4 3
Remote 10−6 < P 0 ≤ 10−5 2
Improbable 10−7 < P 0 ≤ 10−6 1

We define the semi quantitative table with the constraint that any probability category represent a different
order of magnitude, therefore indicating with P’MAX the maximum value of the generic interval category in
Table 5, we can write P = log10 (P’MAX )+7. Then, P is now a logarithmic representation of the maximum
value of the associated probability interval, no more only an ordinal ranking number. Each increment of
P represent a probability increase of an order of magnitude with base 10. This change is what, in many
situation is important to differentiate in terms of risk evaluation. For these data fields, sum and subtraction
are meaningful, representing changes in probability of as many orders of magnitude as there are increases or
decreases in P.
Given P, a conservative estimation of P’ is P’ = 10P-7 , while a possible average estimation could be
P’=0.5*10P-8
Thus, returning to the initial example based on 10 levels for O and D we could define the following tables:

Table 6: Semi-quantitative Occurrence and Detectability

Common terms Probability range O D
Almost sure 10−1 < O0 orD0 ≤ 1 10 10
very High 10−2 < O0 orD0 ≤ 10−1 9 9
high 10−3 < O0 orD0 ≤ 10−2 8 8
often 10−4 < O0 orD0 ≤ 10−3 7 7
probable 10−5 < O0 orD0 ≤ 10−4 6 6
occasional 10−6 < O0 orD0 ≤ 10−5 5 5
Remote 10−7 < O0 orD0 ≤ 10−6 4 4
Improbable 10−8 < O0 orD0 ≤ 10−7 3 3
Very Improbable 10−9 < O0 orD0 ≤ 10−8 2 2
Incredible 10−10 < O0 orD0 ≤ 10−9 1 1

Therefore given O & D, the conservative estimations for O’ and D’ are: O’ = 10O-10 and D’ = 10D-10 , then:

P 0 = 10O−10 ∗ 10D−10 = 10O+D−20 (5)

Therefore we can convey a semi quantitative representation for P with P ∈ [1,2,. . . ,20], where:

P = O + D, and 10O+D−21 < P 0 ≤ 10O+D−20 (6)

Then, given a 10 levels severity scale, and in similar way to the previous quantitative approach we can
determine for each severity level the maximum acceptable levels for P. In the next example, referring with
Pacc this vector, we suppose the maximum acceptable probability for the lower severity = 0.01 (P = 18) and
the following acceptable values obtained with unitary ranking decrements for the first 7 steps, and 2 units
decrements for the following 2 with the highest severities, that is:
Pacc = [18,17,16,15,14,13,12,11,9,7], so the severities can be calculated as: Si = 18/Pacci , with i = 1,2,. . . ,10.

4
The severity table become:
Common terms S
1 Catastrophic 2.57
2 Critical 2.00
3 Severe 1.64
4 Very High 1.50
5 High 1.38
6 Medium 1.29
7 Moderate 1.20
8 low 1.12
9 very low 1.06
10 negligible 1.00
Finally, we can calculate a real risk priority number in this way:

RP N = S ∗ (O + D) (7)

Also this RPN, even if based on ranking numbers, has respected by construction the multiplication and sum
meanings. For this reason, can be assumed as quantitative risk estimation and consequently a priority index.

Comparison with example reported in D.J. Wheeler article

In this section, comparing the example reported in D.J Wheeler article [D.J.W.], with the example of the
previous section for the semi-quantitative approach, some additional considerations are pointed out.
The two approaches have the same level of complexity concerning all RPN terms S, O & D that have 10 levels
each. The basic difference is due to the terms of the former method that are only qualitative ordinal data.
The next figure recall the distribution of RPN for [D.J.W.] example with RPN = S * O * D, plotting it with
100 bars rather than 1000 as in the original article.

Histogram of Risks RPN = S*O*D

70
50
Frequency

30
0 10

0 200 400 600 800 1000

RPN
Figure 1: RPN distributions for ordinal data method

5
Where:
[1] "Number of distinct classes = 120"
[1] "Max RPN values per class = 24"
In the following plot is shown instead the classes distribution of the proposed method.

Histogram of Risks RPN = S*(O+D)

70
50
Frequency

30
0 10

10 20 30 40 50

RPN
without rounding on S calculation

Figure 2: RPN distributions for semi-quantitative data method

[1] "Number of distinct classes = 149"

[1] "Max RPN values per class = 66"
The highest bar in the histogram is located at RPN = 18 that is the acceptability threshold as defined in
the semi-quantitative example. The RPN density is higher for RPN close to the acceptability threshold and
especially for the lower risks, that are those useful to evaluate risks and take decision on possible actions. It
is worth noting that all RPN classes above the acceptability threshold do not have a particular interest.
Concerning previous Figure 1, we can only say that the distribution is more concentrated on the lower risks,
but we cannot have any idea about the relevant RPN region where could be important taking decisions.
However, in general the information included in Figure 1 is not really meaningful since it is not possible to
identify any RPN reference because the RPN classes are not numerically comparable.
Comparing instead the number of classes, the proposed method has more classes, that are 149 against the
120 for the ordinal data method. Moreover, the class correspondent to the acceptability threshold, collects 66
different risk combinations against 24 for the most populated class in the ordinal data method.
It is worth noting that the risks distribution could be affected by the rounding effect during S calculation. In
the following Figure 3 is shown the histogram adopting the rounding to the second decimal digits as in the S
table shown above.
[1] "Number of distinct classes = 168"
[1] "Max RPN values per class = 30"

6
 Histogram of Risks RPN = S*(O+D)

70
50
Frequency

30
0 10

10 20 30 40 50

RPN
rounding S calculation on 2nd decimal digit

Figure 3: Effect of rounding in S calculation

In any case, for roundings with 3 or more decimal digits there is no practical difference with case without
rounding.

Conclusion
Scientific thinking should pursue knowledge and the consequent potential decisions, on objective measures.
I found in my experience, the proposed semi-quantitative approach a way to take effective design and
operating decisions based as much as possible on an objective process. I said “as much as possible”, since the
determination of the maximum acceptable probability limits can be still affected by subjectivity, nevertheless,
they can become more and more objective as new data and new experience are gained. The standard
subjective approach, being the consequent risk evaluation affected by subjective interpretation, could imply
consistent delays when is important to take a decision in a timely manner. In addition, there is a higher
probability to take the wrong decision.
Finally, I think that the power of the proposed method stands on its capability to force the practitioner
to define, before starting the risk analysis, some project-dependent acceptability limits, which are the core
concept of any possible further decision.