LOPA and Human Reliability

GCPS 2010 __________________________________________________________________________
LOPA and Human Reliability Human Errors and Human

IPLs

By: William Bridges
Process Improvement Institute, Inc. (PII)
1938 Riversound Drive
Knoxville, TN 37922 USA
+1-865-675-3458
[email protected]

2010 Copyright reserved by Process Improvement Institute, Inc.

Prepared for Presentation at
American Institute of Chemical Engineers
2010 Spring Meeting
6th Global Congress on Process Safety
San Antonio, Texas
March 22-24, 2010

UNPUBLISHED

AIChE shall not be responsible for statements or opinions contained
in papers or printed in its publications
GCPS 2010 __________________________________________________________________________

LOPA and Human Reliability Human Errors and Human
IPLs

By: William Bridges
Process Improvement Institute, Inc. (PII)
[email protected]

Keywords: human factors, Layer of Protection Analysis, LOPA, process safety, PSM, human
error, independent protection layer, IPL, SIS, SIL, safety instrumented system

Abstract

Layer of Protection Analysis (LOPA) is a simplified risk assessment method that
provides an order of magnitude estimate of the risk of a potential accident scenario.
Humans can be the cause on an accident scenario (the Initiating Event [IE]) or
humans can serve or participate as an independent protection layer (IPL). In either
case, estimating the likelihood of the human error and measuring the human error
rate at a site are troublesome tasks within the LOPA framework, which is one of the
reasons some companies do not give any credit for a human IPL.

Identifying and sustaining independent protection layers (IPLs) is the heart of
LOPA. Each IPL must be:

independent of the initiating event (IE) and the other IPLs
capable (big enough, fast enough, strong enough, etc.)
maintained or kept in practice/service
validated/proven to provide the probability of failure (PFD) chosen

and all of the above must be documented and audited periodically to ensure
compliance with these definitions.

There are many types of IPLs, and some are more trustworthy than others, hence
the difference in the PFD of IPLs. As just mentioned, one possible type of IPL is a
Human IPL. These include preventative steps that may stop a scenario from
progressing once it is initiated, but more typically the human IPLs are responses to
alerts or alarms or troubling readings and sample results.

This paper discusses the data needed for adequately counting the human in a LOPA
(and other risk assessments), and includes discussion of the theory of human
factors. One key focus of the paper is on practical means for collecting raw data in
GCPS 2010 __________________________________________________________________________
a plant setting for substantiating the error rates for the site, and especially for
crediting a human IPL. The method for data collection covers the training
requirements that should be met, proof drills for response to alarms, simulations and
tests, and frequency of proofs, and of course the effect of human factors on human
error rates. Actual plant data and tests are included in the paper to provide the
reader with some examples of how a simple data collection and validation method
can be set up within their companies.

Human Error Fundamentals

Human errors are sometimes mistakenly called procedural errors. This is not true anymore than
saying all equipment errors are due to design errors. Over the past 5 decades of research and
observation in the workplace on human error, we have come to know that human error
probability depends on many factors. These factors (described in more detail in Human Factors
Missing from PSM, LPS/GCPS 2010), include:

Procedure accuracy and procedure clarity (the number one most sited root cause of
accidents):
o A procedure typically needs to be 95% or better accuracy to help reduce human
error; humans tend to compensate for the remaining 5% of errors in a written
procedure.
o A procedure must clearly convey the information (there are about 25 rules for
structuring procedures to accomplish this) and the procedure must be convenient
to use.
o Checklist features These should be used and enforced either in the procedure or
in a supplemental document

Training, knowledge, and skills
o Employees must be selected with the necessary skills before being hired or
assigned to a department.
o Initial Training There must be effective, demonstration based training on each
proactive task and each reactive (e.g., response to alarm) task.
o Ongoing validation of human action is required and usually must be repeated (in
either actual performance or in drills/practice) at least once per year (as discussed
later in this paper). For human IPLs, the action must be demonstrated to be fast
enough as well.
o Documentation the performance of the humans must be documented and
retained to demonstrate the error rates chosen are valid.

Fitness for Duty Includes control of many sub-factors such as fatigue (a factor in a great
many accidents), stress, illness and medications, and substance abuse.

Workload management Too little workload and the mind becomes bored and looks for
distraction; too many tasks per hour can increase human error as well.

GCPS 2010 __________________________________________________________________________
Communication Miscommunication (of an instruction or set of instructions or of the
status of a process) is the second or third most common cause of human error in the
workplace. There are proven management systems for controlling communication errors.

Work environment Factors to optimize include lighting, noise, temperature, humidity,
ventilation, and distractions.

Human System Interface Factors to control include layout of equipment, displays,
controls and their integration to displays, alarm nature and control of alarm overload,
labeling, color-coding, fool-proofing measures, etc.

Task complexity Complexity of a task or job is proportional to the (1) number of
choices available for making a wrong selection of similar items (such as number of
similar switches, number of similar valves, number of similar size and shaped cans), (2)
number of parallel tasks that may distract the worker from the task at hand (leading to
either an initiating event or failure of a protection layer), (3) number of individuals
involved in the task, and (4) judgment or calculation/interpolation, if required. For most
chemical process environments, the complexity of the task is relatively low (one action
per step), but for response actions (human IPLs) there are almost always other tasks
underway when the out-of-bounds reading occurs or the alarm is activated.

In addition to the human factors listed, other considerations for use of a human as an IPL include
(1) time available to perform the action and (2) physical capability to perform the action.

When using human error data for IEs and IPLs, the site must ensure that the factors above are
consistently controlled over the long-term and that they are controlled to the same degree during
the mode of operation that the LOPA covers. For instance, if the workers are fatigued following
many extra hours of work in a two week period leading up to restart of a process, then the human
error rates can increase by a factor of 10 times or 20 times during startup.

Revealed versus Unrevealed Errors for Human. As with equipment failures, human errors can
lead to a revealed fault in the system (the flow does not start, for instance) or to an unrevealed
fault (the block valve downstream of a control valve is left closed, but the failure is not revealed
until the control is needed/used). If the error is revealed, then the error can be corrected or
compensated for. If the restoration/correction time is sufficiently short, then the probability of
being in the failed state is much less than for an unrevealed failure that is only discovered upon
testing or inspection.

GCPS 2010 __________________________________________________________________________
General Relationship of Human Factors Consideration to LOPA

Every risk assessment must consider the likelihood and effect of human factors. For LOPA,
poor human factors can lead to higher human error rates that increase IE frequencies and that
increase the PFD of a human IPL. The table on the next page summarizes the human factor
issues that relate directly to LOPA. The table contrasts the impact of good and poor human
factors on initiating event frequency (IEF) and on the PFD of independent protection layers
(IPLs). If all of the factors were optimized, one would expect the IEF to be about 0.01 per task
for IEs and about 0.05 for response to alarms or call for actions (human IPLs), but such an
analysis must be made on a case-by-case basis, especially for the PFD of IPLs.

GCPS 2010 __________________________________________________________________________

Considerations for Getting Low Human Error Rates

Issue
Requirement for claiming a low
error rate for humans causing
(initiating event) an accident
(usually skipping a step or doing
one wrong)
Requirement for using a low probability of failure
for a human as a protection layer against an
accident (usually a response to alarm)
Comments
Time to
perform action
NA
Time to detect deviation, diagnose the problem,
decide on proper action, and take action is
recommended to be less than 1/2 the time required to
reach the consequence of interest.
The reliability of the indication (sample &
analysis in lab, field reading, etc.) or
annunciation (alarm of the deviation) may
limit the value for the human action IPL.
Operator must be reliably available.
Capable NA
It is physically possible to perform the control action
required to forestall the consequence of interest, and
the capability has been demonstrated and is not the
limiting factor in the human action (e.g., the operator
is strong enough to do the required task)
If the final control requires closing a manual
block valve, the action and strength
required must be verified in the field (some
valves take several operators more than 1
hour to close, even with a valve wrench or
cheater). This usually requires preventive
maintenance or equipment use as part of a
'critical action.'
Procedure
Step must be in procedure. Not
completing the step as stated and in
that sequence (most aspects of errors
of omission and commission) is the
IE. The step is a critical step and is
indicated as critical in the procedure
or with a warning or caution. The
procedure must follow "best
practices" for control of human error;
best practices relate to content
accuracy, page format, and step
presentation. See rules for
procedures presented elsewhere.
Step must be in a trouble-shooting guide or similar
contingency procedure (including emergency
shutdown procedure), that is in paper form or reliably
available on demand electronically (including via
linkage to trouble-shooting guides within a distributed
control system [DCS] for an alarm point) that
describes how to respond to a process deviation.
The response procedure follows best practices as
well; see rules for procedures development and
writing.
Note that we assume there is a 100%
chance of skipping the required step or
doing it substantially wrong if the step is not
in the written procedure or if the step
written is wrong or in the wrong sequence.
Checklist
Checklist in place and its use is
enforced.
NA
Without a checklist, the error rate goes up
by a factor of 3 to 5 times.
GCPS 2010 __________________________________________________________________________
Issue
one wrong)
Comments
Initial Training
There is initial training focused on this
step as a critical step
There are initial training and drills on the response to
a deviation; and the deviation is annunciated.
emphasized in initial training.
Ongoing
validation (or
practice) of
human action
Steps must be practiced, but for
seldom-used procedures, this
practice (normally in the form of a
talk-through for shutdown or startup
of a continuous unit) is not until just
prior to use. Practice rate increases
reliability per action (due to increase
in skill) but Actual Use Rate also
increases the number of opportunities
for failure. These factors can offset
each other completely. See
limitations on error rates that account
for "practice" versus "error rate per
task demand."
Steps must be practiced routinely enough to assure
the reliability of the action under the increased stress
caused by the alarm/indication. The time it takes to
complete the action must be measured during each
practice/drill, to ensure response time limits are not
exceeded. Each person who is expected to
implement the action "on demand" must perform the
practice/drill per the assigned frequency. The
minimum amount of practice or drills and type of
practice should be at least one per year. (Note: One
refinery practiced one scenario for a human IPL,
selected at random, every shift and found no issues
on shift workload and actual had improvement in
general worker performance.)

It is possible to group similar IPLs along the
classification of similar types of actions and similar
available response times. If such grouping is
possible, then practicing one action will in effect be
practicing all similar actions. For determination of
sample size for tests/drills based on number of
human IPLs and number of human responders, refer
to standard methods for sample size determination.
practiced as often as specified here.
Documentation
of validation of
human action
Document each practice (i.e., each
startup and each shutdown) and
document failures rates for 'critical
steps'. Keep the file of the validation
associated with the task and also with
each worker.
Document each practice (i.e., each response) and
document failures rates for 'critical response action.'
Keep the file of the validation associated with the
'action' and also with each worker.

If a sample plan method is used, the same data as
above is still required, and further, the data should be
checked each quarter to ensure against inadvertently
using the same IPL and same operator too often.
practiced as often as specified here or if
there is No documentation of the
completed refresher training or just-in-time
training for an IE, or of drills for an IPL.
GCPS 2010 __________________________________________________________________________
Issue
one wrong)
Comments
Control of
related
physiological
and
psychological
stressors
Fatigue Company policy and enforcement to limit fatigue, including limit of hours worked
per day and per week, and including restrictions on hours worked outside of work. Can affect
error rates by factor up to 20X.
Other fitness of duty issues Company policy and enforcement to limit effect of alcohol &
drug abuse, illness, prescription drug effects, and personal life stress effects. Can affect error
rates by factor up to 50X.
Workload Workload needs to be managed to optimize worker performance. Enough stress
(time/task load-based), but not overstress. (Workload is also related to the complexity of the
work. Complexity increases mental workload, even if the number of tasks per hour is not
affected.) Workload evaluation may be necessary. Can affect error rates by a factor up to
10X.
Communication The second most common cause of human error, the site must have a
policy and rules and follow-through for controlling verbal and visual communication. Rules
include repeat-back on instructions received and use of common jargon. Poor control of
communication rules can increase human error rates by 10X.
Work environment Temperature, lighting, noise, distractions, ventilation, etc., has been
optimized to improvement worker performance. Can affect error rates by factor up to 5X.
Human-System Interface Includes control-display-information integration, labeling, error-
proofing designs, color-coding, alarm management, etc. Poor human-system interface control
can increase human error rates by 10X or higher (especially if norms are violated, where a
negative effect of 20-50X is assumed).
Complexity Complexity of a task or job is proportional to the (1) number of choices
available for making a wrong selection of similar items (such as number of similar switches,
number of similar valves, number of similar size and shaped cans), (2) number of parallel
tasks that may distract the worker from the task at hand (leading to either an initiating event or
failure of a protection layer, (3) number of individuals involved, and (4) judgment or
calculation/interpolation, if required. For most chemical process environments the complexity
of the task is relatively low (one action per step), but for response actions there is almost
always other tasks underway when the out-of-bounds reading or the alarm is activated.
Complexity is difficult to predict (since it is not known when a human action will be needed),
but higher complexity can increase error rates by 2X to 10X.
All of the listed human factors (beyond the
procedure and training and skills listed
above) have a large, independent influence
on human error rates. A large negative
swing in any one of the factors can
increase human error rate by a factor of 3
to 50 times.
GCPS 2010 __________________________________________________________________________
Issue
one wrong)
Comments
Qualitative
hazard
evaluation
There has been a hazard evaluation of the
scenario for skipping the step and doing it
wrong and a hazard evaluation team has
evaluated the consequences and
safeguards.
The human action in this case is a safeguard that is called
into play once a sequence of events is initiated. The human
action (following the trouble-shooting guide or the
emergency shutdown procedure) is to make a change of
state in the system to prevent propagation of the event to
the stated consequence. The hazard evaluation has
evaluated the human action (and any annunciation device
unavailability) in judging that this "critical responsive action"
is capable, reliable, and audited.
It is essential that a company finds accident
scenarios that can arise during all modes of
operation, including startup, shutdown,
emergency shutdown, and online maintenance.
Use of multiple
successful
human actions
within a scenario
Making multiple mistakes (including
shortcutting multiple steps) is the same
probability as making a single mistake, if
the "safeguard" steps amount to ensuring
the same state in the system. The human
who makes the errors sees this as
performing the task by method B instead
of method A. Method B could have
different steps or less steps that method
A.
If one person in a workgroup is assumed to make the
mistake that is the initiating event, then it is assumed that no
one in the workgroup can be part of the protection layers.
Further, supervisors are also part of the workgroup, due to
the trust/relationships that build quickly.

Seemingly unrelated workgroups, such as maintenance and
operations, usually build enough trust to rule out counting as
separate groups.

For a group/individual to be counted as a safeguard against
the errors by another human error, the detection and
response group must not share staff, must have an
incentive to find mistakes and report/respond, and near miss
reporting and auditing both indicate that independent
measurement and action is occurring.

An exception that is normally allowed is to re-use a
workgroup or individual if the second action/response is
separated by distance (location) and/or separated by
enough time and/or there is a reminder alarm for the same
alarm (the alarm that was missed earlier) that recurs often
enough in the allowed time to still avoid the consequence.
It is nearly impossible to use a human response
safeguard, if a cause of the predicted accident
sequence is a human error.

GCPS 2010 __________________________________________________________________________

TYPES of HUMAN ERRORS of IMPORTANCE IN LOPA

Human error comes in many forms, but the primary ones of interest in LOPA are:

Errors in following proactive procedures, such as startup of a unit, that results in an
initiating event of a LOPA accident scenario.
Errors in responding to a call for action; if performed correctly such actions would
interrupt the scenario and prevent the consequence.

The examples below illustrate both types of human error as they relate to LOPA.

Example of I E caused by Human error

In NUREG/CR-1278 The Human Reliability Handbook (Dr. Alan Swain was the
primary author), error rates were tabulated from various sources, including some
empirical data. For errors that could be considered initiating events, these were
summarized in Table 15-3 of that document, which is extracted and included in slightly
amended form below:

Omission of Item
Human Error
Probability
When procedures with check-off provisions are correctly used:
(1) Short list, less than or equal to 10 items .001
(2) Long list, greater than 10 items .003
When procedures without check-off provisions are used, or when available
check-off provisions are incorrectly used:
(3) Short list, less than or equal to 10 items .003
(4) Long list, greater than 10 items .01
(5) When written procedures are available and should be used but
are not used:
.05

The table above shows the estimated probabilities of errors of omission per item of
instruction when use of a written procedure for a task is required. The probabilities in the
tables above assumes that human factors are optimized as well, which includes providing
an accurate and easy to follow procedure, good control of fitness for duty factors
(including controls against fatigue), good controls against miscommunication,
training/practice (or actual performance) at least once per year, etc., as described earlier.

Since startup is when most accidents occur due to human error (when human error is the
initiating event), it is appropriate to use that as an example. Startup procedures normally
have more than 10 steps, indeed more than 50 steps is more typical. So, if all human
factors are optimized, the probability of human error taken should be 0.003 per step. If
practice with the procedure is only done once per year or less often, then the IEF would
GCPS 2010 __________________________________________________________________________

be 0.005, which we would round up for Basic LOPA to 0.01. It may seem conservative
to round up, but consider that there are few (if any) chemical plant sites that have
optimized most of the human factors. Poor control of certain human factors can have a
multiplying effect on the human error of 3 to 50 times. For this reason alone, it is
dangerous to use the data above as representative of the error rates at a given site. Some
analyst believe that in the absence of site data, these values can be applied, but perhaps it
is best to scale the values upward (use higher base error rates) if the site determines that
the human factors described earlier in this paper and in other resources are Not
optimized.

Example human error related to human-based I PLs

Again, from NUREG/CR-1278 The Human Reliability Handbook, estimates were
provided for the probability of human error to respond correctly and in time following an
alarm (annunciation). The Figure below is extracted from the handbook for response by
control room personnel to an abnormal event. This represents the time it takes to
diagnose the problem given there are No other alarms and given that all human factors
are optimized, including practicing the response once per year or more often. The error
probabilities given are the joint probability of failing to notice the alarm and failure to
diagnose the situation correctly.

Note from the figure above that the more time to detect and diagnose the problem, the lower
the chance of error. (HEP refers to Human Error Probability. Pr refers to the probability of
making a diagnosis error.)

This figure only tells part of the story. After diagnosis, the operator, perhaps with help of
others, must then respond appropriately to prevent the scenario from propagating to the next
GCPS 2010 __________________________________________________________________________

protection layer or to the ultimate consequence. The sum of the time to detect, diagnose, and
complete the proper response must happen within the available process safety time (PST).
The process safety time is the interval from when initiation of the event begins until the time
when no further action can prevent the consequence. The PST must be known for a site to
judge if a specific human response can qualify as an IPL. Some organizations require that
the operators overall response must be accomplished in much less (normally half) of the
PST, for the human response to be considered an IPL (given other criteria of an IPL is met;
see LOPA 2001 for guidelines on qualifying an IPL). It is generally assumed that if the
problem is detected (such as by hearing an audible alarm) and if the diagnosis is correct, then
the action is likely to be more accurate than the detection/diagnosis (though this conclusion
does not account for the speed of the response action). The norm therefore is to use a PFD of
0.1 if the response can be performed within the PST (or within the PST for some
organizations) and the PST is equal to or greater than 20 minutes (so the complete detect-
diagnose-response steps are accomplished in 10 to 20 minutes, depending on the
organizations rules for LOPA). A value of better than this (lower error rate) is normally not
allowed in Basic LOPA. From an advanced topic perspective, however, Human Reliability
Analysis (HRA) will often use a PFD of 0.01 for response times greater than 40 minutes
(PST greater than 80 minutes). Performance of HRA and documenting the basis for lower
probability values requires a level of effort and a degree of education/experience far beyond
the required learning to perform Basic LOPA.

HUMAN ERRORS as related to IEs

Human errors that result in IEs are usually because of operator errors during startup (roughly
65% of all accidents in chemical-related process plants occur during startup) or errors during
normal shutdown or emergency shutdown (together these represent 10% of the IEs caused by
human errors). The remaining errors usually occur during normal operation and include such
issues as performing a step wrong during operation in manual mode (such as when a process
controller is bypassed) or by turning the wrong switch or opening the wrong valve.

To find these errors, a very good system for reporting and investigating near misses is needed
(but that topic has been covered in other research and papers, see Bridges 1997 and 2008). The
US Nuclear Regulatory Agency (NRC) in 10 CFR 55.45 and 55.59 requires that all power plant
operators be tested once per year on abnormal procedures. This testing is mostly related to
humans involved as IPLs. 10 CFR 55.45 and 55.59 also allude to the possibility of using a test
of a representative sample of human responses, but we address this option later in the paper.

To help minimize human errors that could be IEs, many companies will review a normal
shutdown or normal startup just before its next use. This has proven effective in controlling
human error in chemical-related industries. Nuclear power plants also perform such just-in-
time (JIT) refresher training on normal shutdowns and normal startup procedures.

If a site records the errors that are made during such tasks, then the company can validate the
average IEF for that task and so the site will have a better estimate of the error probability.
Further, if the measured rate is too high, they can take proactive steps to lower the error rate.
GCPS 2010 __________________________________________________________________________

Errors can be measured either in actual practice in the field, or via simulators, or roughly by
analysis of near misses reported (if the number of near misses reported is at least 20 times higher
than the number of loss events or accident; Bridges 1997 and 2008; CCPS 2003).

Below is an example of data from one site that was collected to estimate the IEF for mistakes
made during normal startup:

EXAMPLE - I nitiating event frequency data collection and use: In one company, the
operating group was able to provide the data on the following page for errors made
during startup of various units. Based on this data, the average probability of human
error was 0.0071 and the number of errors per year that could lead to a known accident
sequence was 0.17 (but note that if all errors are counted, the average error rate would
be 0.29/yr). For basic LOPA, this value is rounded up to an IEF of 1/yr, for initiating
events arising from these and similar procedures at the site. See the Table on the next
page for the raw data.

Chemical-related process plants should collect such data to understand and validate the initiating
event frequency caused by human error. This data would then be used to validate the values for
IEF used in LOPA and other risk assessments.

Site data on Human Errors related to Initiating Event Frequency (IEF)

Task for which
a Potential
Initiating
Event exists
Number of
steps in the
procedure
Number of
Uses of the
procedure
Number of
Human Errors
Noted
Number of
Years of Data
Number of
Errors That
Initiated a
Potential
Accident
Sequence
Human Error
Probability per
Step
Error Rate
(probability of
error per year)
1. Startup of
Process A
46 4 2 6 1 0.0109 0.17
2. Startup of
Process B
71 5 3 6 1 0.0085 0.17
3. Startup of
Process C
55 4 2 6 2 0.0091 0.33
4. Startup of
Process D
56 4 0 6 0 0.0000 0.00
Average: 0.0071 0.17

IEF for LOPA
for site:
1/yr

GCPS 2010 __________________________________________________________________________

HUMAN ERRORS as related to HUMAN RESPONSE IPLs
A 10
-1
PFD value for a human response indicates that the correct response occurs at least 9 out of
10 times (or no more than 1 wrong response for every 10 attempts). Most organizations will
have identified many human responses involving a number of personnel, as part of their LOPA
studies. Some organizations believe that if they have a procedure and a training program in
place, that they can claim the PFD value of 0.1 for a Human IPL. This is no truer for a human
IPL than it is for an active component IPL. A Human IPL must be tested and validated the same
as a component IPL. On the next few pages are the options for validating human IPLs; these (1)
include 100% testing for each human responder and for each action to be taken and (2) a sample
plan testing of random combinations of human responders and actions.

Using a 100% Individual Test Plan Approach for Validation of Human IPLs.

One method to prove the operators will reliably respond for each Human IPL trigger is to have
each operator demonstrate they can individually respond to each alarm (or other trigger). This
response can be demonstrated by walkthroughs in the field or using simulators of the process
unit. The nuclear power industry uses this approach for validating response by control room
operators, but many in the chemical industry perceive this will take too much time. As an
example of the effort required, one nuclear power plant allocates about 200 hours per year per
operator for refresher training activities, including 60 hours per year per operator for
demonstration of skill in responding to critical alarms (from Tennessee Valley Authority [TVA]
internal requirements to meet the performance-based requirements of NRC regulations 10 CFR
55.45 and 55.59). But is this time too much? Well explore that answer in the rest of the
paper.

Consider the following as an example of the application of this method of testing for 100% of the
responses by 100% of the operators at an actual chemical plant:

Background: The operating area being evaluated has 25 operators and 130 identified
(from a hazards analysis and LOPA) human response IPLs.

Validation Test: The tests are documented on a set of index cards that call out various
alarm conditions; these are the events (triggers) for a scenario identified in the hazards
analysis (and LOPA) to be protected by the given human response IPL.

Demonstrated Result: A correct answer (success of the IPL) is the desired response to
the alarm scenario. A wrong answer (failure of the IPL) is any other response than the
one desired or a response that takes too long.

Estimation of Resources Requirements for 100% Testing Scheme: How much test
effort is needed to ensure that training and retraining programs are sufficient to validate a
10
-1
value for the PFD of all of these identified human response IPLs? An example
estimate is as follows:

GCPS 2010 __________________________________________________________________________

1. Determine the number of test to be performed. This would be the number of
human response IPLs multiplied by the number of people who are expected to
respond (perform as the human IPL) at some point in the future during their own
shift. This example would yield 2750 (25 x 130 = 2750) discrete tests for one test
of each combination of trigger and human responder (which makes up the 2750
IPLs).

2. Determine the test frequency. It is difficult to get consensus on this value. One
documented example is from the nuclear industry. The US NRC regulation for
control room operators (10 CFR 55.45 and 55.59) requires annual demonstration
of proficiency in response to critical alarms and signals. This in fact would likely
Not be enough testing to give a 90% chance of proper response to every alarm,
except that response to one alarm is normally similar to response to other alarms,
so the operator is in essence getting more practice on similar alarms as they
perform each demonstration. Operators under this regimen of recertification have
shown a 90% chance or better of proper response (detection, diagnosis, and
action) within 10 minutes of an annunciation of an event of interest (from internal
power plant records and also inferred from, Swain, 1983 and NUREG 1278 data,
since the basic control room regimen of testing each operator with each action
each year was essentially the same in the 1970s as they are now). For this
example, a frequency of 1 validation per year is chosen.
3. Determine the time required to perform each test. Assuming 10 minutes of

allowed response time for success (per alarm/trigger), then the test time for the
organization would be 27,500 minutes or about 460 staff-hours (22 hours per
operator per period). With a frequency (test period) of once per year per alarm,
this equates to about 1% of the normal staff-hours for a worker in the USA.
(Note: An equal amount of time would also be required to record/document the
training record for the test [once the tracking system for such demonstrations is
set up], but this is likely not a load the operator will have to bear.) An investment
of 1% of each operators time seems relatively inexpensive for the relative risk-
reduction afforded by a Human IPL. (Note that testing only a sample of these
responses, discussed later, would reduce the load considerably.)

EXPERIMENTAL DATA related to Human Response IPL.

Many companies have suggested that collecting data on human responses in order to validate that
a human IPL is justified will be very difficult. However, the actual effort to collect such data is
low and the benefits are great, as demonstrated by several chemical plants and refineries and
nearly all nuclear power plants that have tried this.

To help demonstrate this, PII designed a simple test for use in measuring the response to an
alarm condition. The test was not meant to measure the probability of detection of the alarm, but
rather was meant to measure the time and success in determining and accomplishing the proper
GCPS 2010 __________________________________________________________________________

response to critical alarms as part of human IPLs. Two chemical plants belonging to large
organizations (one in Malaysia and one in the USA) performed the test.

Test setup: The test involved having multiple operators in one unit of one plant/site perform
responses to critical process alarms. These alarms were related to human IPLs. The actual
response and time of response was measured, but essentially the tests were setup as a pass or
fail in other words, the test were to determine if the operators were able to respond as
desired/expected.

Each test took 10-15 minutes to administer and less than 1 minute to record the data. The data
was administered to multiple operators on multiple shifts. The tests were administered by a shift
supervisor, a shift engineer, or in some cases, a process safety coordinator; though in the future,
it is likely fine for a lead operator to administer the test. The overall effort to administer the test
was deemed minimal, though there was resistance to such performance/validation test at first (as
would be expected).

A human response IPL failed if the operator could not perform the required action to prevent
the hypothetical outcome within the process safety time (definer earlier).

To run each test, the plants printed a data card (the size of an index card) and handed it to an
operator chosen at random. Below is an example of such an index card.

Note that the card contains an estimate of the PST as defined earlier in this paper and
elsewhere, this is the time an operator has to perform the task once the alarm is received until it
is too late to take any further action. The time it took to print and hand out the index card was
minimal. The person administering the test then timed the operator response and recorded the
results. Any time less than the PST was a successful validation. (Again note that these tests did
not validate the probability of an operator failing to detect an alarm.) The total time for the test
varied, but the two sites that performed the test considered the administration time to be minimal;
the largest time effort is simply for someone other than the operator to be there to
independently measure the operators response time (i.e., time to administer the test). Perhaps
in the future, another operator could likely administer most proof tests and then the site
GCPS 2010 __________________________________________________________________________

management could audit some percentage of the tests to help ensure against bias. If the
operators test each other, then the time to administer a test is likely not significant enough to
measure since they have to be there regardless for their other duties.

Aside: Currently the Safe Automation guidelines by CCPS and the upcoming guideline for IPLs
and IEs (pending, 2010) both stipulate that the action must be completed within of the PST.
But, the PST rule was arbitrarily chosen and has no statistical basis that the author is aware of.

For the most part, the tests came with little warning and occurred on all shifts. Several critical
alarms were tested using various operators, all randomly selected. (It is anticipated that unless a
sample plan is used, each operator will perform roughly one response related to each human IPL
each year.) The time to respond was recorded on the index card.

Based on such raw data, the site was able to (1) evaluate the degree to which they were
controlling human factors for the units, (2) identify which human responses qualify as IPLs, and
(3) validate that the response is accurate enough and quick enough to qualify for the PFD used as
the credit for the IPL (which for basic LOPA, the PFD is limited to a value of 0.1).

The table on the next page provides a sample of the site-specific data for several similar Human
IPLs from the two sites (one in Malaysia and one in Texas, USA). For the Malaysia site, the data
was from the same operating area consisting of multiple operators across 4 rotating shifts of 8
hours per shift and for the Texas site, the shift was 12 hours.

The sites believed that it is probably reasonable to use a sampling of human error for similarly
qualified humans doing similar response or proactive tasks. A sample plan of perhaps only 5%
to 10% of the number of human-task pairs may be necessary to have a valid statistic for human
error for a type of action. Sampling is valid for human actions because of how the mind
processes information and actions for similar situations. Obviously, sampling can greatly reduce
the measurement and documentation load for validation of human error rates. Human
performance sampling is discussed in more detail next in this paper.

GCPS 2010 __________________________________________________________________________

Human Error Data from Two Sites (USA and Malaysia) related to Independent Protection Layers (IPLs)

IPL
No.
Response Task
Number of
Test
Performed
Average
Response Time
(minutes)
Process
Safety Time
(minutes)
Number
Failures
PFD
(average)
LOPA PFD
Company A (USA) Site 1
IPL 1 ABP: High Temp in Generator 6 2.25 10 0 0 0.1
IPL 2 ABP: Loss of Acid Flow to Generator 12 2.166666667 10 0 0 0.1
Company B (Malaysia), Site 1
IPL 1 LICA-HL-05007 10 5.054 15 0 0 0.1
IPL 2 TI 09010 low alarm on 09-K001 discharge 10 4.943 15 0 0 0.1
IPL 3 LAL-21004 10 6.962 15 0 0 0.1
IPL 4 LAH 22002 10 6.058 15 0 0 0.1
IPL 5 PIAH 21006 10 5.036 15 0 0 0.1
IPL 6 PIC-29014 10 5.352 15 0 0 0.1

GCPS 2010 __________________________________________________________________________

Below are responses from the sites to questions related to the measurements:

How much labor did each performance test take (on average)?
o The majority of the labor was with the actual operator going out to the field to take
action on the alarm. Prep time was 1 hour for the initial recording of 10 data points
for each of 2 to 6 IPLs, the prep time for each test was minimal as the cards were
prepared beforehand and the recording takes less than 10 seconds.
o The test time was about 15 minutes per operator per test.
o Recording time was negligible.

How much work do you think this would be longer-term (after getting in the habit and
working any bugs out)?
o Not much difference in time as the testing process was picked up after the 4
th
to 5
th

data point in the unit, so it did not take long to get to minimal effort. One benefit of
the test was the re-enforcement of what to do with all parties involved (the person
doing the test, the person recording the test, and the other operators who noticed the
test in progress).
o The hard part is getting that up-front commitment; after doing a few tests, doing more
was no issue.

Do you feel such data is necessary for each human response I PL? (I n other words, do you
think a 5-10% sample of related type of responses would suffice?)
o The response for all the IPLs were the same when the same operator acts on different
alarms or when different operators act on the same alarms. If this is the case, we
probably can use a sample to test a general response to a type of alarm.
o The site should first screen which response can be grouped together into general types
of response IPLs. Then a lesser percentage will need individual validation drills.
o Perhaps do just 1 or 2 drills per shift per group per year for the simpler ones on some
periodic basis; that gives you a chance to test feasibility (make sure valves not locked,
make sure valve wrench or other tools are available, etc).

What improvements should be made to the tests/validation scheme?
o If we make the simulation a little bit sophisticated, we may be able to induce stress
and surprise (maybe create some disorientation or carry out the tests in not-so-smooth
situations such as when Gene Hackman commenced a missile drill at the end of a
minor crisis in the movie Crimson Tide) so that we can get closer to the actual critical
situations. This would obviously need to be under very controlled conditions.
o Perhaps with simulators or other methods, we could also include detection and
diagnosis of the alarm in the same test.

Who administered the test?
o Company 1. Plant production engineer
o Company 2. As agreed upfront, about 80% was by the Shift Supervisor; 20%
managed to be administered by the PSM Engineer or Shift Executive (Engineer)
although more in an observer capacity.
GCPS 2010 __________________________________________________________________________

Who do you think should admin the tests in future?
o Same or a Lead Operator, which is basically an expert from the operator ranks.
o Shift Supervisor level. In the Malaysia company, this is the highest level for non-
degree holders in operations (or it can also be a newly graduated engineer).

What unit was this in?
o This was a small batch plant in Texas
o Ammonia Plant and Urea Plant in Malaysia

Overall feeling of the exercise?
o The staff said it was easy to do and didn't have much difficulty in setting it up.
o Initially, probably due to not fully seeing the whole picture, it was seen as a burden to
do. Convincing was required to explain the potential benefits. But the plant staff
rose to the occasion and the actual implementation was not so bad judging by what
was able to be achieved even with the unplanned shutdowns that happen to occur in
the month when the validation tests were run.

Using a Statistical Sample Plan Approach in the Future for Validation of Human IPLs.

It is important to ensure that a 10
-1
value is indeed valid before using a human IPL. Testing
every involved person on every identified response is not that burdensome, as shown in the test
cases presented earlier that was run at two chemical plant sites. But, this burden could also be
wasteful if testing the response to one alarm by one individual is essentially the same as testing
the response of similarly trained and drilled individuals to the same or similar alarms.

Rather than testing 100% of the responses by 100% of the operators, it may be valid to use a
sampling plan. This may be especially true for groups of responses that are similar in action and
response time. US NRC alluded to a sampling of human response in 10 CFR 55.59, indicating
that this may be acceptable for validating human response to triggers (i.e., for validating Human
IPLs).

Statistical techniques developed decades ago are used to establish a basis of acceptance of all
kinds of products, raw materials, components, etc. These methods can also be used to validate
Human IPLs.

The sample plan approach must group similar type of actions and similar response time
requirements. For a sample plan approach, choice of responder and trigger must be chosen at
random. The lot size is the product of the number of responders multiplied by the number of
similar response actions in the group.

As a rough rule of thumb, the sample should be about 10% of your total population of data, but
not smaller than 30 and not greater than 350 to 500. The sample size and number of failures
before the PFD is invalid is related to the confidence level and margin of error that is acceptable
to the organization. A confidence level of 95% and an error margin of 5% indicate that the result
GCPS 2010 __________________________________________________________________________

of your testing (validation of a Human IPL) will be within +/-5% of the true PFD 95% of the
time the validation testing is performed.

The correct sample size is a function of those three elements your universe (how many people
multiplied by the number of actions; each pairing makes up a Human IPL), the desired error
margin, and the preferred confidence level. For IPL validation purposes, it is likely reasonable to
use a 5% error margin at 95% confidence. Below are typical sample sizes (the first at a 10%
error margin, the second at 5%):

50 in the population, sample 33 or 44

The trend above approaches a limit for a 10% error margin that hardly moves above 350 in the
sample size no matter how large the population (and approaches a limit of 500 for a 5% error
margin). The sample size can also be approximated using the equation below:

Sample Size, ss:

ss (for infinite population) =
Z
2
* (p) * (1-p)

c
2

Where:
Z = Z value (e.g. 1.96 for 95% confidence level)
p = percentage picking a choice, expressed as decimal (0.9 used for Human IPL
sample size determination)
c = confidence interval, expressed as decimal (e.g., .05= 5%)

This calculation then must be corrected to account for a finite population (validation SS):

Validation SS =
Ss
1 +
ss 1
population

Meeting acceptance criteria for a human response means that the procedures, training, retraining,
communication control, and all other human factors are achieving the desired result of no more
than 1 wrong response in 10 demands. Rejection means that the PFD of 10
-1
is not valid (or that
the control of human error needs improvement for the PFD to remain valid). When applying this
sampling plan to a validation of administrative IPLs, the acceptance criteria (the desired
response) is the response that prevents the consequence that is being considered in the LOPA,
e.g., the operator response to the alarm prevents the overflow.

GCPS 2010 __________________________________________________________________________

Sampling plan schemes rely on developing a valid definition of a group to be the sample. For
Human IPLs, the group is a combination of similar operators (likely all operators are treated
equally for this purpose) and similar responses to similar triggers. The latter grouping takes
careful consideration by a multiple-disciplinary team (with heavy emphasis on the operators on
the team composition), to ensure the grouping of IPLs makes sense (i.e., use expert judgment).
Although it is likely that all operators can be lumped into the same statistical group (if they are
selected at random for validation drills of IPLs), the triggers and responses will need to be
grouped to ensure that validation of one trigger/response is essentially the same as validating
other triggers/response within the same group.

The upcoming guideline for IPLs and IEs (CCPS, 2010, pending) provides guidance on choice of
sample size and on grouping of alarms and workers who respond to them. Included in the
discussion is a method derived from ANSI Z1.4 (now incorporated by reference with US MIL
standards for lot sampling). Examples calculations and the resulting savings achievable by use
of sample plans are also provided in the guideline.

The table on the next pages provides an example of the relative savings possible by using a
sampling of human response IPLs instead of testing each operator for each IPL.
GCPS 2010 __________________________________________________________________________

GCPS 2010 __________________________________________________________________________

HUMAN ERRORS as related to Impact non-human IEs and IPLs (such as
SIS and Relief Systems)
In addition to having a direct impact as an IE or as failure of a human IPL, human error can of
course impact all other IEs and IPLs as well. This is because ALL safeguards are ultimately
controlled and maintained by humans. Typically, a base human error rate of 1 error in 50 steps is
reasonable, or for excellent control of human factors, an error rate of 1 error in 100 steps may be
achieved. Below are some examples of how human factors control the reliability of all IPLs and
IEs:

Example related to I E wrong materials of construction received: If the humans fail
to control or detect errors during selection and use of materials of construction, then
perhaps the wrong grade of steel or other material will be used in a highly corrosive
environment and then the IEF will be many times higher than expected. Industry data
indicates that the materials delivered are different that what we specify about 3 to 7% of
the time. Positive material identification (PMI) detect and correct such failures,
theoretically, but plant data suggests that using 100% PMI at receiving will lower the
composite error rate to 1% and that another 100% PMI of as-welded in the field
lowers the combined error rate to about 0.3%. This still leaves a large number of
components that may be of the wrong materials.

Example related to I PL high integrity SI S: Safety Instrumented Systems (SIS), can
be designed to provide risk reductions of 1, 2, or 3 orders of magnitude. However, even
the best designed SIS can be negated if (1) the root valves for the sensors are left closed
or (2) a bypass valve around an emergency isolation valve is left open (other errors are
also possible of course). The current versions of SIS standards do not explicitly require
validation of the control of human factors and the resulting human errors that can leave
the valves in the wrong position. (In addition, the humans are always capable of
revaluing trip points, etc., in the safety-rated PLC, given enough time to learn these
shortcuts and given enough spurious trips to make them want to use such a shortcut.)
Although an SIL 4 SIS is allowed by IEC 61508, with a PFD of <10
-4
to 10
-5
, looking
at it pragmatically, we do not believe it is possible to control the human errors
associated with in situ testing low enough to maintain a SIL 4 rating. Similar, due to
the complexity of the analysis of human errors associated with in situ testing, basic
LOPA only allows use of SIL 1. SIL 2 and SIL 3 rating is difficult to validate
statistically, due to the consideration of the human factor, as a result, SIL of 2 and 3
require human reliability event tree or similar methods to quantify, and so is considered
beyond basic LOPA.

Example related to I PL pressure relief systems: Many relief system IPLs can have a
PFD of 0.01 or better, assuming the rest of the process does not interfere or restrict the
flow. However, if you have a block valve upstream or downstream of the a relief
device, then leaving it closed will negate the entire value of the relief system. A human
reliability analysis of existing plant practices is needed to establish the PFD is better
than 0.1, and excellent sustained control of human factors is necessary to maintain a
PFD of .01 or better for a relief system; because of the relatively high probability that a
GCPS 2010 __________________________________________________________________________

human will leave a block valve closed upstream or downstream of a relief valve.
Common industry data for human error rates indicates that the error of leaving such a
block valve closed will likely be the dominating failure mode for the relief system, so
including block valve upstream or downstream of a relief device (though allowed by
ASME with administrative controls) will need to be evaluated by a site before assigning
a PFD for the combination system (comprised of the relief system and block valves).
The conservative PFD of 0.1 is therefore used as the default value for relief devices that
have blocks valves upstream or downstream of relief devices, unless a human reliability
for the plant proves otherwise.

Conclusion

Control of human factors is key to achieving high reliability of processes. Poor human factors
lead to higher initiating event rates and less or no value for IPLs. Proper control of human
factors can be achieved by following industry best practices (not minimal regulations and
codes/standards).

Failing to control human factors will result in the time invested in LOPA to be wasted, and more
importantly will lead to more accidents. Simply installing or claiming IPLs will not change the
critical dependency all plants face on controlling risk risk control always depends on human
factor control.

References

1. Tew, R. and Bridges, W., Human Factors Missing from PSM, LPS/GCPS, 2010.

2. US Nuclear Regulatory Commission, Training Requirements for Nuclear Power Plant
Operators, 10 CFR 55.45 and 55.59

3. NUREG/CR-1278 The Human Reliability Handbook; guidelines from the US NRC on
Human Reliability Analysis.

4. Layer of Protection Analysis (LOPA) Guideline, AIChE/CCPS, 2001.

5. Guidelines for Independent Protection Layers and Initiating Events, AIChE/CCPS, 2010
[pending].

6. Guidelines for Investigating Chemical Process Incidents, 2
nd
Edition, CCPS/AICHE,
2000.

7. Bridges, W., Gains in Getting Near Misses Reported, 8
th
Conference, ASSE-MEC,
Bahrain, 2008.

LOPA and Human Reliability

Uploaded by

Copyright:

Available Formats

LOPA and Human Reliability

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

LOPA and Human Reliability

Uploaded by

Copyright:

Available Formats

GCPS 2010 __________________________________________________________________________

LOPA and Human Reliability Human Errors and Human

3. Determine the time required to perform each test. Assuming 10 minutes of

You might also like