space-rha.comShort courseChapter 9 · Rates and mission impact

Chapter 9 · Lesson 3 of 4 · about 16 minutes

Counts, probability and availability

A predicted count of one event is an average over many missions like this one. This spacecraft may see none, or several. The expected number, the probability of at least one event and the consequence of that event are three different quantities, and the report keeps them separate.

What you need to know

A predicted rate times a mission duration gives an expected number of events, μ. That number is a mean over many missions like this one. This spacecraft sees a whole number of events, which may be zero. For independent events at a constant rate the count follows a Poisson distribution, and the probability of at least one event is 1 − exp(−μ). For μ = 1 it is 63 %, not 100 %. For μ = 0.1 it is 9.5 %, close to μ itself. For μ = 3 it is 95 %. The figure shows the curve with the mission values of U1 and U2 marked. The conversion is done for each class of event separately, each with its own duration, population and duty state, and it says nothing about the consequence. An expected 28 interrupts is a different fact from an expected 0.001 destructive events, and neither becomes a probability of mission loss until the architecture has been applied.

A recoverable event still costs time. The interruption runs from the event to validated function (chapter 7 lesson 3). For rare interruptions that do not overlap, the unavailable fraction is the rate times the mean interruption. The two-state Markov model with exponential failure at rate λ and restoration at rate r gives a steady-state availability of r/(λ + r), which for small λ/r is the same number. The model leaves out events that overlap, restorations that fail, permanent damage and the event storms of lesson 9.2. Each of those needs its own state in a larger model. A small average unavailability also does not bound the chance of missing one critical window. That is a question about when the events fall, and it is answered with the rate over that window alone.

System counts add when the hazards are independent. The μ for a function is the sum over its parts of n_i λ_i t_i, with the number of devices, the rate and the time at risk for each part. The sum fails when one particle affects several nodes, when parts share a supply or a clock whose interruption takes them all down, and when a storm raises every rate at once. Those cases are modeled as common causes rather than as independent terms. Powered-off parts generally keep accumulating dose and can keep a vulnerable bias, so time at risk is not the same as time powered.

Figure 9.3. The probability of at least one event against the expected number μ, with the three-year mission values of the teaching parts marked: for U1 (best estimate and worst case) the probability and μ are the same number; U2's interrupts are certain, and their cost is the downtime.
Probability of at least one event against the expected number 0.001 0.01 0.1 1 10 0 0.2 0.4 0.6 0.8 1 expected number of events, μ probability of one or more 1 − exp(−μ) μ itself U1 high-current, 3 years, X/5 U1, X/100 U2 SEFI, 3 years
Try it: from a rate to mission counts and availability

Enter a rate, the number of devices, the mission length and the downtime per event. The table gives the expected count, the probability of at least one or k events, and the unavailable fraction.

Worked example 9.3. Three years in a sun-synchronous orbit for U1 and U2

Event classRate per day (SSO)μ over 3 yearsP(at least one)Question for the architecture
U1 high-current event, best estimate (X/5)1.3 × 10⁻⁶1.4 × 10⁻³0.14 %is the recovery path (limiter, independent supply) verified; is there latent damage
U1 high-current event, worst case (X/100)4.6 × 10⁻⁶5.0 × 10⁻³0.50 %as above; this is the number the risk decision quotes
U2 functional interrupt, heavy ions plus protons0.02628.5100.0 %what is lost in each interruption, and for how long
U2 upset, heavy ions plus protons1401.5 × 10⁵100 %do the code and the scrub interval keep uncorrectable words rare (chapter 10)

For U1 the destructive-class probability is of order 0.5 % in the worst case. The project will weigh that number in chapter 11 against the cost of a different part. U2's interrupts are certain, and their cost is the interruption of 87.5 s each (chapter 7 lesson 3). A rate of 0.026 per day times 1.46 minutes gives an unavailable fraction of 2.6 × 10⁻⁵, or 0.23 hours per year, and the two-state availability is 99.9974 %. If the mission's science depends on one recording window of ten minutes per orbit, the question becomes the probability of an interrupt inside that window, 0.018 % per window, times the number of windows that matter.

Exercise

For each event class in your own design, report μ, the probability of at least one event and the consequence at the function as three separate fields, so that nobody reads one of them as another.

Trace every dependency in the recovery path (detection, fault isolation, the reset or power cycle, the boot, reconfiguration, calibration and reacquisition) and measure the time to validated function under a realistic fault, instead of taking the nominal restart time.

Going deeper

Independence and consequence decide what the count means

Adding the hazards of the parts gives a system count when the hazards are independent, or independent once the environment is fixed. A shared supply, one particle hitting several nodes and a storm that raises every rate at once break that assumption. Going from an event to a probability of loss takes a model of the architecture and the consequence of each event given where it lands; the Poisson conversion alone does not do it. Modes that were not identified stay outside the calculation.

For independent constant hazards, μ = sum of nᵢ λᵢ tᵢ. Say where that simplification fails.

The repair model has limits

The ideal two-state continuous-time Markov model, with exponential transitions up and down, has a steady-state availability of r/(λ+r), where r is the restoration rate. A reset sequence of fixed length, a finite mission window, permanent damage and event storms each need a different treatment. A small average unavailable fraction does not bound the chance of missing one critical operation.

Report the availability for each mission phase, and the probability of getting through a critical window, separately from the steady-state figure.

The common mistake

The Poisson formula assumes independent events at a constant rate. It does not make the events independent, and it knows nothing about failure modes that were not counted.

The two-state result leaves out permanent failures, correlated interruptions and recovery times that are not exponential, unless those are added to the model.

Takeaway

  • An expected event count read as a probability of mission loss distorts both the requirement and the mitigation trade.
  • A less sensitive part and a faster validated recovery are two different ways to raise the availability of a function.

Check yourself

A rate of one event per mission is quoted. What is the probability that the mission sees none?

exp(−1) = 36.8 %, and the probability of two or more is 26.4 %. "One event" is a mean over many missions, not the count this one will see.

The U2 interrupt downtime is budgeted as the 2 s reset command. What does that do to the availability figure?

It understates the unavailable fraction by the ratio of the validated recovery (87.5 s) to the command (2 s), a factor of 44. The interruption runs from the event to validated function.

Two redundant recorders share U1. Can their interrupt counts be summed as independent?

Their own interrupts can. The shared U1 is a common cause, since a U1 event takes down both recorders, so the hazard for the function includes the U1 rate once rather than twice.

The U1 worst-case line gives μ = 5.0 × 10⁻³ over three years and a probability of at least one event of 0.50 %. A reviewer reads the 0.5 % as the probability of losing the mission. What is wrong, and what else is needed?

The 0.5 % is the probability that U1 sees at least one high-current event in three years; for small μ the probability and μ are nearly the same number. Whether that event becomes a loss depends on the architecture: whether the recovery path (limiter, independent supply) is verified and whether there is latent damage, which is the question the table attaches to that line. An expected count becomes a probability of mission loss only after the consequence of each event, given where it lands, has been applied, and modes that were not identified stay outside the calculation.

The two-state availability of the recorder is 99.9974 %, and the project reads it as the chance of completing each ten-minute recording window. Why is that a different question?

The availability is a steady-state average: 0.026 interrupts per day times 1.46 minutes gives an unavailable fraction of 2.6 × 10⁻⁵, or 0.23 hours per year. A small average unavailability does not bound the chance of missing one critical window; that is a question about when the events fall, and it is answered with the rate over the window alone, 0.018 % per ten-minute window, times the number of windows that matter. The two-state model also leaves out events that overlap, restorations that fail, permanent damage and the event storms of lesson 9.2, each of which needs its own state in a larger model, so the availability is reported per mission phase and the critical-window probability separately from the steady-state figure.

A designer sets the time at risk t_i of a part that is powered for only part of each orbit equal to its powered time. When is that scaling right, and when is it not?

The μ for a function is the sum over its parts of n_i λ_i t_i, and t_i is the time during which the part can suffer the event class in question, which is not the same as the time powered. Powered-off parts generally keep accumulating dose and can keep a vulnerable bias, and the duty cycle can change the susceptibility, what is observed, or the consequence (lesson 9.4). The scaling is generally defensible only when the off state has been shown to remove the hazard for that event class, and the conversion is done for each class separately, with its own duration, population and duty state.

Further reading

  • NASA, "Avionics Radiation Hardness Assurance (RHA) Guidelines," NASA/TM-20210018053, 2021. NTRS 20210018053.
  • K. A. LaBel et al., "Single event effect criticality analysis," NASA Goddard Space Flight Center, 1996.
  • S. Mukherjee, Architecture Design for Soft Errors, Morgan Kaufmann, 2008.
  • K. A. LaBel et al., "Emerging radiation hardness assurance (RHA) issues: a NASA approach for space flight programs," IEEE Trans. Nucl. Sci., vol. 45, no. 6, pp. 2727-2736, 1998. doi:10.1109/23.736521.