Tools › Reliability › Safety Analysis › SIL Verification and LOPA › Method
SIL Verification and LOPA: Method
Open the SIL Verification and LOPA tool
SIL Verification and LOPA: Method, Sources, Assumptions and Limits
What the tool does
Two calculations of process safety, linked. A layer of protection analysis (LOPA) takes several initiating causes of one consequence, multiplies each by the protection layers and modifiers credited against it, adds the results and compares the sum with a tolerable frequency; the gap is the probability of failure on demand (PFD) required of a safety instrumented function. A SIL verification then computes the average PFD of that function from its sensor, logic-solver and final-element subsystems, with proof tests that do not reveal every failure, and compares it with the requirement. Either part can be run alone.
The companion tool Hazard and Safety Analysis holds the MIL-STD-882E hazard log, a single-cause LOPA and a SIL verification with perfect proof tests, diagnostic coverage and the architectural constraints (hardware fault tolerance and safe failure fraction). This tool adds what that one leaves out: several causes with layers credited cause by cause, proof-test coverage and mission time, 2oo4 voting, the spurious trip rate and the sensitivity to the test interval. It does not repeat the architectural constraints.
Layer of protection analysis
For cause i with initiating frequency fi (per year), the mitigated frequency is fi Π pij, the product running over the layers credited against that cause: an independent protection layer enters with its PFD, a conditional modifier (probability of ignition, of a person present, of injury) and an enabling condition (fraction of time the hazard exists) with their probabilities. The causes lead to the same consequence, so their mitigated frequencies add: F = Σ fi Π pij.
A layer is credited only against the causes it is independent of. The opening case shows this: an operator's response to an alarm raised by the control system is credited against a valve closed in error and against loss of cooling water, and not against the failure of the control loop itself.
With Fp the mitigated frequency of the causes the function acts on and Fu that of the others, the function must satisfy Fu + PFD Fp ≤ ftol, so PFDreq = (ftol − Fu) / Fp and the required risk reduction factor is RRFreq = 1 / PFDreq. If Fu alone reaches ftol no function can close the gap.
SIL bands (low-demand mode). SIL 1 is a PFDavg of 10−1 to 10−2 (risk reduction 10 to 100), SIL 2 10−2 to 10−3, SIL 3 10−3 to 10−4, SIL 4 10−4 to 10−5. The bands are those of IEC 61508 and ANSI/ISA 84.00.01 as quoted in public U.S. Department of Energy contractor documents (Hearn and Suttinger, 2007, all four bands; Strelow, 2015, SIL 1 to 3). A required PFD of 0.1 or more is below SIL 1; one below 10−5 is beyond what a single function may claim.
Model of a subsystem
A subsystem has N identical channels voted MooN: M must work for the function to act, so it fails dangerously when k = N − M + 1 channels are down. Each channel carries three failure mechanisms, taken as independent with constant rates:
Undetected dangerous failures the proof test reveals, rate C λDU, where C is the proof-test coverage. They stay hidden until the next proof test, every TI, and are then repaired.
Undetected dangerous failures the proof test does not reveal, rate (1 − C) λDU. They stay until the end of the mission time TM, when the subsystem is overhauled or replaced.
Detected dangerous failures, rate λDD, announced by diagnostics and repaired at rate 1 / MTTR while the process runs. As an option they drive the channel to its trip state instead: they then leave the PFD and are added to the safe failure rate.
Common cause follows the β-factor model: a fraction β of each rate strikes every channel at once and the rest, 1 − β, strikes the channels one by one. β is ignored for 1oo1.
With τ the time since the last proof test and t the time since the start of the mission, the unavailability of one channel from its own failures is
q(t) = 1 − e−(aτ + bt) [1 − w(t)], a = (1 − β) C λDU, b = (1 − β)(1 − C) λDU, w(t) = g / (g + μ) [1 − e−(g + μ)t], g = (1 − β) λDD, μ = 1 / MTTR.
The common-cause part qcc(t) has the same form with β in place of 1 − β. The subsystem is unavailable when the common cause has struck or when at least k of the N channels are down:
U(t) = 1 − [1 − qcc(t)] [1 − Σj = k..N C(N, j) qj (1 − q)N − j].
Exact average
PFDavg = (1 / TM) ∫ U(t) dt over the mission time. This is the definition of the average as Bukowski and others (2014, equation 1) print it for a single device. The tool integrates numerically (ten-point Gauss-Legendre quadrature between consecutive test and renewal instants, where U is smooth). For the function, the integrand is 1 − Π [1 − Ui(t)] over the subsystems, so the overlap of the subsystems' failures is not counted twice. "Exact" means exact for the model above; it is the value carried forward and compared with the requirement.
Simplified equations
To leading order in rate × time (1 − e−x ≈ x, and the probability of k or more channels down ≈ C(N, k) qk):
PFDavg ≈ C(N, k) ⟨q1k⟩ + β [C λDU ⟨τ⟩ + (1 − C) λDU TM / 2 + λDD MTTR], q1 = (1 − β) [C λDU τ + (1 − C) λDU t + λDD MTTR].
The averages ⟨τi tj⟩ over the mission are taken exactly, as a finite sum over the test intervals (⟨τ⟩ = TI / 2 when the mission is a whole number of intervals). The error is of relative order rate × time. Because 1 − e−x ≤ x and the union bound both err upward, the simplified value is never below the exact one. With a perfect test, no detected failures and no common cause it reduces to the familiar C(N, k) (λDU TI)k / (k + 1):
1oo1: λDU TI / 2; 2oo2: λDU TI; 1oo2: (λDU TI)2 / 3; 2oo3: (λDU TI)2; 1oo3: (λDU TI)3 / 4; 2oo4: (λDU TI)3.
The first and third of these, and the exact single-channel average 1 − (1 − e−λt) / (λt), are printed in NUREG/CR-6823 (equations 2.4 and 2.5, page 2-11). For 1oo1 with imperfect testing the simplified form is λDU [C TI / 2 + (1 − C) TM / 2] + λDD MTTR.
Spurious trip rate
A channel's safe failures (rate λS) put it in the trip state until it is repaired. A MooN subsystem trips the process when M channels are in that state at once. To leading order
STR ≈ N C(N − 1, M − 1) [(1 − β) λS]M MTTRM − 1 + β λS,
which is N λS for 1ooN voting, 2 λS2 MTTR for 2oo2 and 6 λS2 MTTR for 2oo3 without common cause. The exact value is the reciprocal of the mean time to the first trip of the same group (channels failing and being repaired independently, the group restarted as new after a trip), found by solving the small set of linear equations of that process. The rates of the subsystems add.
Sensitivity to the proof-test interval
Every proof-test interval is multiplied by the same factor, from 0.25 to 4, and the exact PFD of the function is recomputed. When a requirement exists the tool finds by bisection the factor at which the PFD equals it. Testing more often lowers only the part of the PFD that tests reveal: the part from failures no test reveals, the down time of detected failures and their common cause remain, and the tool says so when that floor is above the requirement.
What it assumes
Constant failure rates (no wear-out and no early failures); identical channels within a subsystem; all channels of a subsystem tested together; proof tests and repairs that are perfect for what they cover and take no time; a repair of a detected failure that does not reveal hidden ones; independent subsystems in series; the β-factor model with one β for every failure rate; a function that is as new at the start of the mission and after every mission time; the low-demand mode, in which demands are rare compared with the test interval. In the LOPA: layers that are independent of the cause and of each other, and causes that are independent of each other.
When it misleads
When the entered rates are not the plant's: generic or certificate failure rates can differ from field experience by an order of magnitude, and the result is no better than they are. When the proof-test coverage or the mission time is optimistic: with imperfect tests the untested failures often dominate, as the opening case shows. When the function is bypassed for testing or repair for a noticeable fraction of the time, or repairs after a failed test are slow: that time is not in the model. When common cause is not a single β (shared impulse lines, shared maintenance errors, shared power): the β term usually sets the floor of a redundant group, and it is the least certain input. When demands are frequent compared with the proof tests (the standards draw the line at about one demand a year): the high-demand and continuous-mode measures apply instead. When systematic failures, software errors and human error in maintenance matter: none is in a random-hardware PFD. When subsystems have different mission times that are not multiples of one another, the function value is an average over the longest one only. In the LOPA: when a layer is credited against a cause it depends on, or two layers share a sensor, an operator or a valve, the product of their PFDs overstates the protection; when the sum of several scenarios with the same consequence is what the criterion limits, each scenario's share of the tolerable frequency must be set by the owner.
A PFD in a SIL band is not a SIL: the standards also set architectural constraints (hardware fault tolerance, safe failure fraction or prior-use evidence) and requirements on systematic capability, which this tool names and does not check. The hazard analysis team and the governing standard decide.
Sources and what was not used
The PFD expressions are derived here from the model above; they are reliability mathematics. IEC 61508, IEC 61511, ISA-TR84.00.02 and the CCPS books on LOPA are cited for the method and their terms; no table, example or data value from them is used. The opening failure rates, probabilities and frequencies are illustrative round numbers chosen for this page, not taken from any source, and no typical values of initiating frequencies or layer credits are offered: none was found printed in a U.S. government document.
Verification
The exact averages are checked against an independent Markov-chain model (every channel as an eight-state chain, the channels and the common cause combined, the proof test as a state change, the state probabilities advanced by a matrix exponential) for 36 random subsystems and six random functions, to better than one part in a million, and against an event-by-event Monte Carlo simulation within its sampling error. The simplified equations are checked against numerical integration of the linearized form, the spurious trip rate against the full chain over all subsets of tripped channels, and the LOPA against separate arithmetic. Printed values reproduced: NUREG/CR-6823 page 2-12, a standby component with 0.02 failures per month tested monthly (λt / 2 = 0.01) and two such components tested together ((0.02)2 / 3); the Los Alamos SCRAM calculation, 10−1 per year against 5 × 10−3 per year giving a required PFD of 5 × 10−2, SIL 1. No public document was found that prints a numerical PFD example with proof-test coverage, so that part rests on the Markov and Monte Carlo checks alone.