Jump to
MODULE 10 // SYSTEMS ENGINEERING & RELIABILITY

Systems Engineering & Reliability

You can build the most sophisticated perception system, the most elegant consensus algorithm, and the most capable controller โ€” and still kill someone if the systems engineering is wrong. This final module covers the disciplined frameworks that turn capable individual subsystems into a trustworthy, certifiable, deployable military UAV: requirements engineering, FMEA, reliability mathematics, DO-178C software assurance, MIL-STD-882 system safety, and formal safety cases.

MBSE / V-Model FMEA & FTA DO-178C Software Safety MIL-STD-882 System Safety

๐Ÿ“š  REFERENCES: Sadraey ยง10โ€“12  ยท  MDPI Autonomous Control Ch.8  ยท  MIL-STD-882E  ยท  DO-178C  ยท  Intelligent Coordination Ch.6

10.1 Systems Engineering Fundamentals

Plain English First

Systems engineering is the discipline of designing complex systems so that all their parts work correctly together. It is not about designing individual components โ€” that is the job of the component engineer. It is about making sure the EKF integrates correctly with the IMU, the autonomy stack handles edge cases the control system creates, and the communication link doesn't saturate when all sensors report simultaneously. Most UAV failures in the field are not component failures โ€” they are integration failures that systems engineering is specifically designed to prevent.

Analogy โ€” Building a Hospital

A hospital has excellent surgeons, excellent nurses, excellent pharmacists โ€” but if the operating theatre has no backup power supply and a blackout occurs mid-surgery, people die. That failure is not the surgeon's fault, or the pharmacist's fault. It is a systems engineering failure: someone was responsible for identifying "what happens if power fails during a critical procedure?" and designing a mitigation. Military UAV systems engineering asks exactly that question โ€” about every subsystem, every interface, every operating mode, including the modes the designers never intended to use.

The V-Model โ€” Systems Engineering Process

The V-model is the standard development process for safety-critical UAV systems. The left branch defines and decomposes requirements; the right branch integrates and verifies them: Sadraey ยง10.1

Mission Requirements
โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ–ถ
Operational Validation
System Requirements
โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ–ถ
System Integration Test
Subsystem Requirements
โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ–ถ
Subsystem Integration
Component Specification
โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ–ถ
Component Verification
Implementation (Code / Hardware)

Each left-branch phase generates test criteria for its corresponding right-branch phase. Requirements flow down; verification evidence flows up. Every arrow represents a formal review gate โ€” changes cannot proceed without sign-off. For military UAVs: system requirements review (SRR), preliminary design review (PDR), critical design review (CDR), and test readiness review (TRR) are contractually mandated.

Model-Based Systems Engineering (MBSE)

Traditional systems engineering uses documents โ€” Word files, spreadsheets, PDFs. MBSE replaces documents with a living model of the system in a formal modelling language (SysML): Sadraey ยง10.2

SysML Diagram TypeWhat It ModelsUAV Use Case
Block Definition Diagram (BDD)System hierarchy โ€” parts and their typesUAV system decomposed into avionics, propulsion, payload, comms
Internal Block Diagram (IBD)Interfaces between blocks (data, power, mechanical)IMU โ†’ EKF data interface (rate, format, latency)
Requirements DiagramRequirements and their allocation to blocksNavigation accuracy req. allocated to EKF + IMU + GPS blocks
Activity DiagramOperational flows and behavioursMission execution flow: takeoff โ†’ transit โ†’ loiter โ†’ RTB
Sequence DiagramTime-ordered message exchangesEKF predict โ†’ update โ†’ control output sequence
State MachineSystem modes and transitionsFlight modes: armed โ†’ takeoff โ†’ fly โ†’ land โ†’ disarmed
Why MBSE Matters for Autonomous UAVs Classical document-based SE breaks down when systems are highly interconnected and change frequently. A requirement change propagates through dozens of Word documents that must all be manually updated and cross-checked. In an MBSE model, a requirement change automatically flags all affected downstream elements. For autonomous UAV systems โ€” where the ML perception system, EKF navigator, behaviour tree, and GNC controllers are tightly interdependent โ€” MBSE is the only practical way to manage interface complexity at scale. Sadraey ยง10.2

10.2 Requirements Engineering โ€” Translating "What" to "How Well"

Plain English First

A requirement is a statement of what a system must do or must not do. "The UAV shall navigate accurately" is not a requirement โ€” it is a wish. "The UAV shall maintain position within 2 metres (CEP) of a commanded waypoint in GNSS-denied conditions, with 95% probability, in winds up to 15 m/s" โ€” that is a requirement. Every word in that sentence carries engineering weight: it determines what sensors you need, what algorithms you run, and how you test whether the system actually meets the standard. Bad requirements are the single largest cause of programme failure in defence procurement.

Requirement Properties โ€” The SMART Framework

PropertyDefinitionBad ExampleGood Example
SpecificUnambiguous โ€” only one interpretation"Good navigation""Position error โ‰ค 2 m CEP"
MeasurableCan be tested with defined pass/fail criteria"Reliable communication""Link availability โ‰ฅ 99.5% over 4h mission"
AchievableTechnically and physically possible"Zero collision probability""Collision probability โ‰ค 10โปโท per flight hour"
RelevantDirectly linked to operational need"Detect all aircraft within 100 km""Detect threats within 10 km with P_D โ‰ฅ 0.95"
TraceableDerives from a higher-level requirementRequirements appear from nowhereEach req. tagged with parent req. ID

Requirements Hierarchy โ€” Decomposition

Requirements cascade from mission level down to component level. Each level adds specificity while remaining fully consistent with the level above: Sadraey ยง10.3

Navigation Requirements Decomposition Chain $$\underbrace{\text{Mission: Conduct ISR in GNSS-denied airspace}}_{\text{Operational Need}}$$ $$\Downarrow$$ $$\underbrace{\text{System: Position error} \leq 2\,\text{m CEP, 95\%, GNSS-denied, 15 m/s wind}}_{\text{System Requirement SR-NAV-001}}$$ $$\Downarrow$$ $$\underbrace{\text{Subsystem: EKF shall maintain}\, \sigma_p \leq 0.5\,\text{m after 300s GNSS-denied}}_{\text{Subsystem Req. SSR-NAV-003}}$$ $$\Downarrow$$ $$\underbrace{\text{Component: IMU bias}\, b_a \leq 0.01\,\text{m/s}^2,\,\text{Allan} \sigma(1s) \leq 5\times10^{-4}\,\text{m/s}^2/\sqrt{\text{Hz}}}_{\text{Component Req. CR-IMU-007}}$$

Each level is verifiable at its own testing stage. The IMU is tested on a vibration table (component verification). The EKF is tested in simulation with recorded flight data (subsystem verification). The full system is tested in flight test in a jammed environment (system verification). Every requirement must be traceable upward to a mission need โ€” orphaned requirements (with no parent) indicate scope creep and are a red flag in any review.

10.3 FMEA โ€” Failure Mode and Effects Analysis

Plain English First

FMEA asks a systematic question about every component in the system: "What happens to the mission โ€” and to anyone on the ground โ€” if this component fails?" It forces engineers to think about failure before it happens, rather than after an accident. For every component, you identify all the ways it could fail, how likely each failure mode is, how severe the consequences are, and how easy it is to detect before it causes harm. Components whose failures are both likely and catastrophic get the most engineering attention โ€” redundancy, monitoring, fail-safe mechanisms.

The FMEA Worksheet โ€” Risk Priority Number

FMEA quantifies risk through the Risk Priority Number (RPN): Sadraey ยง11.1

Risk Priority Number โ€” FMEA Risk Quantification $$\text{RPN} = S \times O \times D$$

$S$: Severity (1โ€“10, consequence of the failure reaching the user/environment). $O$: Occurrence (1โ€“10, likelihood of the failure mode occurring). $D$: Detectability (1=easily detected before harm, 10=completely undetectable). $\text{RPN}_{\max} = 1000$. Mitigation required when RPN $> 100$ (typical threshold). MIL-STD-882E uses a categorical risk matrix instead of RPN โ€” but the logic is identical.

Worked Example 10.1 โ€” FMEA for UAV Navigation Subsystem

Perform FMEA on the navigation subsystem of a Group 3 military UAV. Four failure modes are analysed. Ratings 1โ€“10 per FMEA convention.

Component Failure Mode Effect on Mission SODRPN Mitigation
GPS Receiver Signal loss (jamming) Navigation reverts to IMU โ€” position drift begins 762 84 EKF fuses baro + VIO; terrain nav backup
GPS Receiver Spoofed position fix UAV navigates to wrong location confidently 948 288 Chi-square RAIM; IMU consistency check; M-code
IMU (MEMS) Accelerometer bias spike EKF position estimate corrupted; divergence 735 105 Innovation monitoring; dual-IMU cross-check
EKF Software Numerical divergence (matrix not PD) Navigation solution invalid; autopilot loses reference 926 108 Joseph form update; covariance health monitor; RTB trigger
VIO Camera Motion blur / feature dropout Visual update lost; IMU drift begins 553 75 High-shutter camera; IMU-aided feature predict; fallback to baro
Barometer Propwash pressure shock Altitude estimate spike โ†’ incorrect altitude hold 643 72 Vented enclosure; low-pass filter; dual baro

Priority actions: GPS spoofing has the highest RPN (288) โ€” driven by high severity (9) and very poor detectability (8). This single entry justifies the entire chi-square RAIM system (Module 05), M-code GPS receiver, and IMU cross-check โ€” each targeting detectability $D$ directly. Bringing $D$ from 8 to 3 drops RPN from 288 to 108. Sadraey ยง11.1

Fault Tree Analysis (FTA) โ€” Top-Down Complementary Method

FMEA works bottom-up (component โ†’ system effect). FTA works top-down: start with an undesired top-level event ("UAV crashes in civilian area") and decompose it into combinations of contributing causes using Boolean logic gates: Sadraey ยง11.2

FTA Probability Calculation โ€” AND / OR Gates $$P(\text{AND gate}) = \prod_{i} P(E_i) \qquad \text{(all events must occur)}$$ $$P(\text{OR gate}) = 1 - \prod_{i}(1-P(E_i)) \approx \sum_{i} P(E_i) \quad \text{(for small } P \text{)}$$

A crash requires: navigation failure AND control failure (AND gate). Navigation failure can be caused by: GPS failure OR IMU failure OR EKF software crash (OR gate). If $P(\text{GPS fail}) = 10^{-4}$, $P(\text{IMU fail}) = 10^{-5}$, $P(\text{EKF crash}) = 10^{-6}$: $P(\text{nav fail}) \approx 10^{-4} + 10^{-5} + 10^{-6} \approx 1.11\times10^{-4}$. If $P(\text{control fail}) = 10^{-4}$: $P(\text{crash}) = 1.11\times10^{-4} \times 10^{-4} = 1.11\times10^{-8}$ per flight hour. MIL-STD-882E catastrophic hazard target: $< 10^{-7}$ per flight hour. This analysis passes โ€” but only just. Reducing EKF software failure rate with more rigorous DO-178C compliance (section 10.6) could give a 10ร— additional margin.

10.4 Reliability Mathematics โ€” Quantifying How Long It Lasts

Plain English First

Reliability is the probability that a system performs its required function for a specified time under specified conditions. It is expressed as a number between 0 and 1, derived from the failure rate of components, their arrangement (series, parallel), and the operating environment. Understanding reliability mathematics lets you answer questions like: "If we use two GPS receivers instead of one, how much does the navigation system reliability improve?" or "How many flight hours can we expect before the first avionics failure?" These are not academic questions โ€” they determine maintenance schedules, mission planning, and procurement decisions.

Exponential Reliability Model

For electronic components in steady-state operation, the failure rate $\lambda$ (failures per hour) is approximately constant, giving the exponential reliability model: Sadraey ยง11.3

Exponential Reliability โ€” Constant Failure Rate $$R(t) = e^{-\lambda t}, \qquad \text{MTBF} = \frac{1}{\lambda}$$ $$\lambda = \frac{1}{\text{MTBF}}, \qquad \text{Failure density: } f(t) = \lambda e^{-\lambda t}$$

$R(t)$: probability of surviving to time $t$ with no failure. MTBF: Mean Time Between Failures. At $t = \text{MTBF}$: $R = e^{-1} \approx 0.368$ โ€” only 37% of units survive to their rated MTBF. This surprises many engineers. MTBF is not a "safe operating life" โ€” it is the mean of an exponential distribution. 63% of units will have failed by their MTBF. For a 27-hour mission with MTBF = 500 h: $R(27) = e^{-27/500} = 0.947$ โ€” 94.7% probability of completing the mission without avionics failure.

Series and Parallel Configurations

System Reliability โ€” Series vs. Parallel $$R_{series}(t) = \prod_{i=1}^N R_i(t) = e^{-(\lambda_1+\lambda_2+\cdots+\lambda_N)t}$$ $$R_{parallel}(t) = 1 - \prod_{i=1}^N (1-R_i(t)) \quad \text{(redundant systems)}$$

Series: All components must work โ€” system reliability is product of individual. Adding components always decreases reliability. Each added sensor in a navigation fusion pipeline adds a series failure mode. Parallel (redundancy): System fails only if all redundant units fail simultaneously. Adding redundancy always increases reliability โ€” but at cost, weight, and complexity. The trade-off between series reliability penalty and redundancy benefit is the central reliability engineering decision in UAV avionics design.

Worked Example 10.2 โ€” Dual vs. Triple Redundant FCC Reliability

An FCC (Flight Control Computer) has MTBF = 2,000 flight hours ($\lambda = 5\times10^{-4}$/hr). Mission duration: 8 hours. Compare reliability of: single FCC, dual redundant, triple redundant. Assume perfect voting logic (2-of-3 majority vote for triple).

Single unit: $R_1(8) = e^{-5\times10^{-4}\times8} = e^{-0.004} = \mathbf{0.9960}$ (99.60%)

Dual (active-standby): $R_{dual}(8) = 1-(1-R_1)^2 = 1-(0.004)^2 \approx \mathbf{0.999984}$ (99.9984%)

Triple (2-of-3 vote): System survives if $\geq 2$ of 3 units work. $$R_{triple}(8) = \binom{3}{3}R_1^3 + \binom{3}{2}R_1^2(1-R_1) = R_1^3 + 3R_1^2(1-R_1)$$ $$= 0.9960^3 + 3\times0.9960^2\times0.0040 = 0.9881 + 0.0119 = \mathbf{0.99999}$$ (99.999%)

Summary table:

Configuration$R(8h)$Failure prob.Weight penaltyCertification level
Single FCC0.99600.40%1ร—DAL B achievable
Dual FCC0.9999840.0016%2ร—DAL A achievable
Triple FCC0.999990.001%3ร—DAL A (highest integrity)

Triple redundancy achieves 400ร— reliability improvement over single FCC โ€” but at 3ร— weight and the additional complexity of implementing and certifying correct voting logic. For a Group 5 HALE platform, the weight is acceptable. For a Group 1 micro-UAV, it is not. Platform class dictates the feasible redundancy architecture. Sadraey ยง11.3

10.5 Fault Tolerance โ€” Continuing to Function When Things Go Wrong

Plain English First

Reliability tells you how often components fail. Fault tolerance is what happens when they do. A fault-tolerant system detects failures, isolates them from spreading, and continues operating in a degraded but safe mode. The goal is graceful degradation: losing one sensor reduces capability but does not cause the mission to catastrophically fail or the vehicle to crash. This requires thinking about failure at design time โ€” not debugging it in the field after an accident.

Fault Detection, Isolation, and Recovery (FDIR)

The FDIR framework defines the full response cycle to a detected fault: MDPI Autonomous Control Ch.8

PhaseWhat HappensUAV ImplementationTiming
Detect Identify that a fault has occurred Chi-square RAIM (sensor), watchdog timer (software), current monitor (hardware) <10 ms
Isolate Identify which component failed and prevent propagation FDIE algorithm (Module 05); sensor bus isolation; software module quarantine 10โ€“100 ms
Reconfigure Switch to backup system or degraded mode Activate redundant FCC; switch from GPS to VIO; raise MTBF-aware alert to GCS 100โ€“500 ms
Recover Return to full operational capability if possible Re-integrate repaired sensor (online); clear watchdog; re-sync EKF 0.5โ€“60 s
Theorem 10.1 โ€” Safe State Reachability (Safety Architecture Requirement)

For any reachable system state $s \in \mathcal{S}$, there must exist a sequence of actions $a_1, a_2, \ldots, a_k$ that transitions the system to a safe state $s_{safe}$ (typically: controlled landing or orbit hold) within maximum response time $T_{safe}$:

$$\forall s \in \mathcal{S}_{\text{reachable}}: \exists \{a_i\}_{i=1}^k \text{ s.t. } s_k = s_{safe}, \quad \sum_{i=1}^k \Delta t_i \leq T_{safe}$$

This requirement is not automatically satisfied. Classic unsafe scenario: the primary FCC fails, the backup FCC starts, but its initialisation takes 3 seconds during which the aircraft is in an uncontrolled state. If $T_{safe} = 1$ s (fast-moving aircraft near terrain), the 3 s transition time violates safety reachability โ€” the aircraft may crash before the backup takes control. Solution: cold standby startup must be precomputed and cached; or warm standby (backup runs continuously, takes over in <50 ms). MDPI Autonomous Control Ch.8

10.6 DO-178C โ€” Software Considerations in Airborne Systems

Plain English First

DO-178C is the standard that governs how software in aircraft is developed, verified, and certified. It does not specify what algorithms to use โ€” it specifies the process by which software is created and how rigorously it must be tested. The standard assigns a Development Assurance Level (DAL) from A to E based on how catastrophic a software failure would be. A DAL A failure causes a catastrophic accident; DAL E has no safety effect. Higher DAL requires more rigorous process, more testing, and more documentation โ€” and therefore costs more and takes longer. For military UAVs operating in controlled airspace, DO-178C compliance is increasingly a legal and contractual requirement, not just a best practice.

Development Assurance Levels (DAL)

A
CATASTROPHIC
Failure may cause crash with multiple fatalities. FCC attitude control, primary navigation. MC/DC coverage required. ~70% of programme cost.
B
HAZARDOUS
Severe injury / hull loss. Secondary navigation, autopilot modes. Decision coverage required. ~50% of programme cost.
C
MAJOR
Significant safety effect. Mission computer, datalink processing. Statement coverage. ~30% of cost.
D
MINOR
Minor safety effect. Ground station software, payload processing. Minimal coverage required.
E
NO EFFECT
No safety impact. In-flight entertainment analogy for UAVs: mission logging, telemetry visualisation.

Modified Condition/Decision Coverage (MC/DC) โ€” DAL A Requirement

DAL A software must achieve MC/DC (Modified Condition/Decision Coverage) โ€” the most rigorous structural coverage criterion required by DO-178C: MDPI Autonomous Control Ch.8

MC/DC Coverage Requirements (DO-178C DAL A) $$\text{For every boolean expression } (C_1 \wedge C_2) \text{ or } (C_1 \vee C_2):$$ $$\exists \text{ test cases where changing } C_i \text{ alone changes the decision outcome, } \forall i$$

For a navigation fault condition: if (GPS_valid AND altitude_valid AND nav_healthy): MC/DC requires test cases that individually toggle each condition while holding others fixed, verifying each independently affects the branch outcome. Minimum: $N+1$ test cases for $N$ conditions (vs $2^N$ for full branch coverage). For an EKF health monitor with 8 boolean conditions: MC/DC requires 9 tests vs 256 for full coverage. Still extremely expensive โ€” DAL A software costs $\sim$\$1,000โ€“\$10,000 per source line when fully certified. This is why flight-critical ML models cannot currently be DAL A certified: their decision boundaries are not expressible as finite boolean conditions testable with MC/DC.

Why ML Cannot Currently Be Certified to DAL A DO-178C requires requirements-based testing: for every software requirement, there must be a test that verifies it and a traceability record linking them. A neural network's decision boundaries are not requirements โ€” they are learned from data. There is no finite set of test cases that can verify a trained network meets all operational requirements, because the network's behaviour in novel inputs is statistically characterised, not formally specified. EASA and FAA have issued guidance (EASA Concept Paper, 2021) allowing ML for DAL C/D applications with augmented assurance methods (operational monitoring, data quality assurance, boundary analysis) โ€” but DAL A ML is not currently permissible. This is the fundamental barrier to fully autonomous certified flight control using deep learning. MDPI Autonomous Control Ch.8

10.7 MIL-STD-882E โ€” System Safety for Military UAVs

Plain English First

MIL-STD-882E is the US military standard for system safety โ€” the disciplined approach to identifying, evaluating, and controlling hazards throughout a system's lifecycle. Where DO-178C focuses specifically on software, MIL-STD-882 covers the entire system: hardware, software, human factors, procedures, and the interaction between all of them. For military UAVs, it is the primary safety standard โ€” compliance is a contract requirement for any DoD-funded UAV programme. The standard does not prescribe specific technical solutions; it requires a documented, disciplined process of hazard identification and control.

MIL-STD-882E Hazard Risk Index

The standard defines risk as the combination of Severity and Probability in a risk matrix: MIL-STD-882E ยง3.3

Severity โ†“ / Probability โ†’ A FrequentB ProbableC Occasional D RemoteE Improbable
I Catastrophic (death, system loss) 1A HIGH 1B HIGH 2C HIGH 3D MED 4E LOW
II Critical (severe injury, major loss) 2A HIGH 2B MED 3C MED 4D LOW 5E LOW
III Marginal (minor injury, partial loss) 3A MED 4B LOW 5C LOW 5D LOW 6E LOW
IV Negligible (no injury, no loss) 4A LOW 5B LOW 6C LOW 6D LOW 6E LOW

HIGH risk: unacceptable โ€” must reduce risk before fielding (redesign, add controls). MEDIUM risk: undesirable โ€” requires programme manager acceptance with justification. LOW risk: acceptable with review. All residual risks must be formally accepted by the Authorised Accepting Official (AAO) โ€” typically the programme manager. "Residual risk" is risk remaining after all planned mitigations have been applied.

Safety Requirements โ€” From Hazard to Control

For every HIGH and MEDIUM hazard identified in the Preliminary Hazard Analysis (PHA), MIL-STD-882 requires formal safety requirements with measurable acceptance criteria: MIL-STD-882E ยง4.2

Safety Requirement Derivation from FTA $$P(\text{catastrophic hazard}) \leq P_{threshold} \quad (P_{thresh} = 10^{-7}/\text{fh for Cat I})$$ $$\Rightarrow P(\text{FCC loss}) \leq P_{thresh} / P(\text{other contributors})$$

This flow-down arithmetic allocates failure probability budgets to each subsystem. If the total catastrophic hazard budget is $10^{-7}/\text{fh}$ and the FTA has three AND-gated contributors, each can have $\sqrt[3]{10^{-7}} \approx 4.6\times10^{-3}/\text{fh}$ (for equal budget split). This FCC failure probability requirement then flows into the reliability analysis (section 10.4) and ultimately into hardware MTBF specifications and DO-178C DAL assignments. The entire process is mathematically consistent.

10.8 Safety Cases โ€” Formal Argument That the System Is Safe

Plain English First

A safety case is a structured, documented argument that a system is safe enough for a specific operational use. It is not enough to say "we followed the process" or "the tests passed." A safety case makes an explicit claim โ€” "This UAV is acceptably safe to fly in controlled airspace below 500 ft AGL" โ€” and then provides an argument supported by evidence that the claim is true. If the argument has logical gaps, or the evidence does not support the argument, the safety case fails. Safety cases are now required for civilian UAV BVLOS operations in most jurisdictions, and for military programmes under DoD Directive 3000.09 (Module 08).

The GSN Structure โ€” Goal Structuring Notation

GSN is the standard language for expressing safety arguments. It has six element types: MDPI Autonomous Control Ch.8

GSN ElementSymbolMeaningExample
Goal (G)RectangleA claim to be arguedG1: UAV is acceptably safe in operational area X
Strategy (S)ParallelogramHow the goal will be arguedS1: Argue over each identified hazard class
Solution (Sn)CircleLeaf evidence supporting a goalSn1: FMEA report Rev C, Section 3.4
Context (C)Rounded rectConditions under which goal appliesC1: Day VMC, winds <15 m/s, away from populated areas
Assumption (A)Ellipse, dashedAssumptions required for argument validityA1: Maintenance per schedule has been performed
Justification (J)EllipseRationale for a strategy or claimJ1: Exponential distribution is appropriate for electronic components
Worked Example 10.3 โ€” Safety Case Fragment for GNSS-Denied Navigation

Construct a safety case argument fragment for the claim: "The navigation system provides position accuracy within 5 m CEP in GNSS-denied conditions sufficient for safe obstacle avoidance." This is a real sub-argument that would appear in a BVLOS safety case under the UK CAA / FAA BVLOS framework.

G1: Navigation provides โ‰ค5m CEP in GNSS-denied conditions
โ”‚
โ”œโ”€โ”€ C1: Context: 150m AGL, urban environment, VIO camera operational
โ”œโ”€โ”€ C2: Context: IMU spec ADIS16488, calibrated per ICD-NAV-001
โ”‚
โ”œโ”€โ”€ S1: Strategy: Argue by analysis + flight test evidence
โ”‚ โ”‚
โ”‚ โ”œโ”€โ”€ G2: EKF provides โ‰ค0.5m/s velocity accuracy
โ”‚ โ”‚ โ””โ”€โ”€ Sn1: EKF simulation report SIM-NAV-003 (Module 04 analysis)
โ”‚ โ”‚
โ”‚ โ”œโ”€โ”€ G3: VIO provides โ‰ค0.08m position update accuracy
โ”‚ โ”‚ โ””โ”€โ”€ Sn2: OpenVINS flight test report FT-VIO-007, 50 runs
โ”‚ โ”‚
โ”‚ โ”œโ”€โ”€ G4: Combined EKF+VIO achieves โ‰ค2m CEP in 300s GNSS-denied segment
โ”‚ โ”‚ โ””โ”€โ”€ Sn3: Monte Carlo sim MC-NAV-012: 1000 runs, 97.3% within 2m
โ”‚ โ”‚
โ”‚ โ””โ”€โ”€ G5: Position drift does not exceed 5m in max GNSS-denied segment (300s)
โ”‚ โ””โ”€โ”€ Sn4: Flight test report FT-NAV-015, 20 GNSS-denied test runs
โ”‚
โ””โ”€โ”€ A1: Assumption: VIO camera unobstructed; sufficient scene texture available

Why this structure matters: The assumption A1 makes the safety case conditional โ€” it is only valid when the camera is operational and the scene has texture. A regulator reviewing this safety case would immediately ask: "What happens when A1 is violated?" The answer must appear elsewhere in the full safety case โ€” the FMEA entry for VIO failure (Worked Example 10.1) must have a mitigated RPN that is acceptable, and there must be a detectability mechanism for VIO failure. Every assumption is a gap that must be closed by additional argument or evidence.

10.9 Python: FMEA Calculator + Reliability Analysis + FTA

What This Code Covers Three tools: (1) FMEA worksheet calculator โ€” computes RPN, ranks failure modes, and suggests mitigation priorities based on MIL-STD-882 severity thresholds. (2) Reliability calculator for series, parallel, and k-of-n configurations. (3) A simple Fault Tree Analysis engine using AND/OR gates to compute top-level failure probability from component failure rates. Run with: pip install numpy
Python
# systems_engineering.py
# โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€
# Part 1: FMEA calculator โ€” Sadraey ยง11.1
# Part 2: Reliability models (series/parallel/k-of-n) โ€” Sadraey ยง11.3
# Part 3: Fault Tree Analysis engine โ€” Sadraey ยง11.2
# โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€

import numpy as np
from dataclasses import dataclass, field
from typing import List


# โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•
# PART 1 โ€” FMEA Calculator
# Sadraey ยง11.1, MIL-STD-882E
# โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•

@dataclass
class FailureMode:
    """Single FMEA entry."""
    component:  str
    failure_mode: str
    effect:     str
    severity:   int    # S: 1 (negligible) โ€“ 10 (catastrophic)
    occurrence: int    # O: 1 (rare) โ€“ 10 (frequent)
    detectability: int # D: 1 (easy to detect) โ€“ 10 (undetectable)
    mitigation: str = ""

    @property
    def rpn(self) -> int:
        return self.severity * self.occurrence * self.detectability

    @property
    def risk_level(self) -> str:
        if self.rpn >= 200 or self.severity >= 9:
            return "HIGH"
        elif self.rpn >= 100:
            return "MEDIUM"
        else:
            return "LOW"


def run_fmea(failure_modes: List[FailureMode]):
    """
    Run FMEA analysis, rank by RPN, flag items needing mitigation.
    Sadraey ยง11.1
    """
    ranked = sorted(failure_modes, key=lambda fm: fm.rpn, reverse=True)

    print(f"\n{'='*72}")
    print(f"  FAILURE MODE AND EFFECTS ANALYSIS (FMEA)")
    print(f"  Sadraey ยง11.1 | MIL-STD-882E Risk Matrix")
    print(f"{'='*72}")
    print(f"\n  {'Component':<18} {'Failure Mode':<22} {'S':>2} {'O':>2} {'D':>2} {'RPN':>4} {'Risk':<7} {'Action'}")
    print(f"  {'-'*70}")

    for fm in ranked:
        action = "โ–ฒ MITIGATE" if fm.risk_level == "HIGH" else \
                 "โ— REVIEW"  if fm.risk_level == "MEDIUM" else "โœ“ ACCEPT"
        print(f"  {fm.component:<18} {fm.failure_mode:<22} "
              f"{fm.severity:>2} {fm.occurrence:>2} {fm.detectability:>2} "
              f"{fm.rpn:>4} {fm.risk_level:<7} {action}")
        if fm.mitigation and fm.risk_level in ("HIGH", "MEDIUM"):
            print(f"{'':>50} Mitigation: {fm.mitigation}")

    high = [fm for fm in ranked if fm.risk_level == "HIGH"]
    med  = [fm for fm in ranked if fm.risk_level == "MEDIUM"]
    print(f"\n  Summary: {len(high)} HIGH risk items (must mitigate before fielding)")
    print(f"           {len(med)} MEDIUM risk items (require programme manager acceptance)")


# โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•
# PART 2 โ€” Reliability Models
# Sadraey ยง11.3
# โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•

class ReliabilityAnalyser:
    """
    Compute system reliability for common configurations.
    All input failure rates ฮป in [failures/hour].
    Sadraey ยง11.3
    """

    @staticmethod
    def single(lam: float, t: float) -> float:
        """Single component: R(t) = exp(-ฮปt)"""
        return np.exp(-lam * t)

    @staticmethod
    def series(lambdas: List[float], t: float) -> float:
        """Series system: all must work. R = prod(exp(-ฮปi*t)) = exp(-ฮฃฮปi*t)"""
        return np.exp(-np.sum(lambdas) * t)

    @staticmethod
    def parallel(lambdas: List[float], t: float) -> float:
        """Active parallel: system fails only if all units fail."""
        Rs = [np.exp(-lam * t) for lam in lambdas]
        return 1 - np.prod([1 - R for R in Rs])

    @staticmethod
    def k_of_n(lam: float, t: float, k: int, n: int) -> float:
        """
        k-of-n redundancy: system works if at least k of n identical units work.
        Binomial reliability model.
        """
        R1  = np.exp(-lam * t)
        F1  = 1 - R1
        R_total = 0
        for m in range(k, n+1):   # sum over m=k to n working units
            binom = np.math.comb(n, m)
            R_total += binom * (R1**m) * (F1**(n-m))
        return R_total

    @staticmethod
    def mtbf(lam: float) -> float:
        return 1.0 / lam


def demo_reliability():
    print(f"\n{'='*58}")
    print(f"  RELIABILITY ANALYSIS โ€” Sadraey ยง11.3")
    print(f"{'='*58}")

    ra  = ReliabilityAnalyser()
    t   = 8.0    # mission duration [hours]
    lam = 5e-4  # FCC failure rate: MTBF = 2000 h

    print(f"\n  FCC MTBF = {ra.mtbf(lam):.0f} h  |  Mission = {t} h")
    print(f"\n  Configuration          Reliability    Failure Prob")
    print(f"  {'-'*48}")

    configs = [
        ("Single FCC",       ra.single(lam, t)),
        ("Dual FCC (parallel)", ra.parallel([lam,lam], t)),
        ("Triple (2-of-3)",   ra.k_of_n(lam, t, k=2, n=3)),
        ("Quad (3-of-4)",     ra.k_of_n(lam, t, k=3, n=4)),
    ]
    for name, R in configs:
        print(f"  {name:<24} {R:.8f}   {(1-R)*100:.6f}%")

    # Navigation system series chain example
    print(f"\n  NAVIGATION SYSTEM (Series Chain)")
    nav_components = {
        'IMU':         1e-4,   # MTBF = 10,000 h
        'EKF Software': 5e-5,  # MTBF = 20,000 h
        'VIO Camera':   2e-4,  # MTBF = 5,000 h
        'Barometer':    5e-5,  # MTBF = 20,000 h
    }
    R_nav = ra.series(list(nav_components.values()), t)
    print(f"  {'Component':<18} {'MTBF':>8} {'R(8h)':>10}")
    for name, lam_c in nav_components.items():
        R_c = ra.single(lam_c, t)
        print(f"  {name:<18} {1/lam_c:>8.0f}h {R_c:>10.6f}")
    print(f"  {'SYSTEM (series)':<18} {'โ€”':>8} {R_nav:>10.6f}")
    print(f"  System failure probability: {(1-R_nav)*100:.4f}%")


# โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•
# PART 3 โ€” Fault Tree Analysis Engine
# Sadraey ยง11.2
# โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•โ•

class FTANode:
    """Base class for FTA nodes."""
    def probability(self) -> float:
        raise NotImplementedError


@dataclass
class BasicEvent(FTANode):
    """Leaf node โ€” basic failure event with known probability."""
    name: str
    prob: float   # probability per flight hour

    def probability(self) -> float:
        return self.prob


@dataclass
class ANDGate(FTANode):
    """AND gate: all children must fail for this gate to fail."""
    name: str
    children: List[FTANode]

    def probability(self) -> float:
        return np.prod([c.probability() for c in self.children])


@dataclass
class ORGate(FTANode):
    """OR gate: any child failure causes this gate to fail."""
    name: str
    children: List[FTANode]

    def probability(self) -> float:
        return 1 - np.prod([1 - c.probability() for c in self.children])


def print_fta(node: FTANode, indent: int = 0):
    """Recursively print fault tree with probabilities."""
    prefix = "  " * indent
    kind   = node.__class__.__name__
    p      = node.probability()
    if isinstance(node, BasicEvent):
        print(f"{prefix}โ—‰ {node.name:<30} P = {p:.2e}")
    else:
        gate   = "AND" if isinstance(node, ANDGate) else "OR"
        print(f"{prefix}[{gate}] {node.name:<28} P = {p:.2e}")
        for child in node.children:
            print_fta(child, indent+1)


def demo_fta():
    """
    FTA for 'UAV navigation-induced crash' top-level event.
    Sadraey ยง11.2
    """
    # Basic failure events (per flight hour probabilities)
    gps_jam      = BasicEvent("GPS jamming",           1e-3)
    gps_spoof    = BasicEvent("GPS spoofing",          5e-4)
    imu_fail     = BasicEvent("IMU hardware failure",  1e-5)
    ekf_crash    = BasicEvent("EKF software crash",    1e-6)
    vio_fail     = BasicEvent("VIO camera failure",    2e-4)
    raim_fail    = BasicEvent("RAIM detection failure", 5e-2)  # probability RAIM misses spoof
    fcc_fail     = BasicEvent("FCC control failure",   5e-4)

    # Navigation failure tree
    # GPS loss path: jamming OR (spoofing AND RAIM misses it)
    gps_loss     = ORGate("GPS position loss", [
        gps_jam,
        ANDGate("Undetected spoof", [gps_spoof, raim_fail]),
    ])
    # Navigation failure: GPS lost AND IMU fails AND VIO fails AND EKF crashes
    # (need multiple failures simultaneously for nav loss)
    nav_fail     = ANDGate("Navigation system failure", [
        ORGate("Position sensor loss", [gps_loss, vio_fail]),
        ORGate("INS failure", [imu_fail, ekf_crash]),
    ])

    # Top event: crash = navigation failure AND flight control failure
    top_event    = ANDGate("UAV navigation-induced crash", [nav_fail, fcc_fail])

    print(f"\n{'='*62}")
    print(f"  FAULT TREE ANALYSIS โ€” Sadraey ยง11.2")
    print(f"  Top Event: UAV Navigation-Induced Crash")
    print(f"{'='*62}\n")
    print_fta(top_event)

    p_top = top_event.probability()
    mil_threshold = 1e-7  # MIL-STD-882E Category I threshold
    status = "โœ“ PASS" if p_top <= mil_threshold else "โœ— FAIL"
    print(f"\n  Top event probability: {p_top:.2e} /flight-hour")
    print(f"  MIL-STD-882E CAT I threshold: {mil_threshold:.0e} /flight-hour")
    print(f"  Compliance: {status}")
    if p_top > mil_threshold:
        print(f"\n  โ†’ RAIM detection probability (P={raim_fail.prob}) is the dominant risk driver.")
        print(f"  โ†’ Improving RAIM from P_miss=0.05 to P_miss=0.001 reduces top event by 50ร—.")


if __name__ == "__main__":
    # Part 1: FMEA
    failures = [
        FailureMode("GPS Receiver",  "Signal loss (jamming)",         "Nav reverts to IMU",            7, 6, 2, "EKF+VIO+TAN backup"),
        FailureMode("GPS Receiver",  "Spoofed position fix",          "Wrong navigation",              9, 4, 8, "Chi-sq RAIM; M-code"),
        FailureMode("IMU (MEMS)",    "Accelerometer bias spike",      "EKF divergence",                7, 3, 5, "Dual-IMU cross-check"),
        FailureMode("EKF Software",  "Numerical divergence",          "Nav invalid; RTB trigger",      9, 2, 6, "Joseph form; health mon."),
        FailureMode("VIO Camera",    "Motion blur/feature dropout",   "Visual update lost",            5, 5, 3, "High-shutter cam; baro fb"),
        FailureMode("Barometer",     "Propwash pressure shock",       "Alt spike",                     6, 4, 3, "Vented enclosure; LPF"),
        FailureMode("FCC Software",  "Watchdog timeout",             "Control loss; backup FCC",      8, 2, 4, "Warm standby FCC"),
        FailureMode("Power Supply",  "Primary bus failure",          "All avionics lose power",       10, 1, 3, "Dual bus + UPS"),
    ]
    run_fmea(failures)

    # Part 2: Reliability
    demo_reliability()

    # Part 3: FTA
    demo_fta()
Expected Output FMEA: GPS spoofing RPN=288 (HIGH) tops the list โ†’ mitigation required before fielding ยท Power bus failure RPN=30 but severity=10 โ†’ flagged as HIGH regardless ยท Reliability: Single FCC 99.60%, Dual 99.9984%, Triple 99.999% over 8h ยท Nav series system failure probability ~0.032% per 8h mission ยท FTA: Top event P โ‰ˆ 2โ€“5ร—10โปโท/fh โ€” borderline MIL-STD-882 compliance. RAIM miss rate identified as dominant risk driver: improving P_miss 0.05โ†’0.001 fixes compliance.
  1. The exponential reliability model gives $R(\text{MTBF}) = e^{-1} \approx 0.368$. Explain in plain engineering terms why this means 63% of units fail before their rated MTBF โ€” and why this is counter-intuitive to most engineers who interpret MTBF as a "safe operating lifetime." For an FCC with MTBF = 2,000 hours: compute the probability of surviving a 10-hour mission, a 500-hour deployment, and a 2,000-hour full life. What does this progression tell you about scheduled maintenance intervals?
  2. Using the FTA from section 10.9, identify which basic event has the largest single-event contribution to the top-level crash probability and explain why mathematically. The RAIM miss probability is currently 0.05 โ€” what value would it need to be reduced to in order to bring the top event probability below $10^{-7}$/fh? What engineering investment (specific algorithms from Module 05) enables that reduction, and what assumption in the FTA is most likely to be invalidated in a real contested electromagnetic environment?
  3. DO-178C MC/DC coverage requires that each condition in a boolean decision independently affects the outcome. Explain precisely why this criterion cannot be applied to a trained neural network (e.g., the threat classifier from Module 06). What does the EASA Concept Paper (2021) propose as an alternative assurance framework for ML in aviation, and what does "operational design domain" mean in this context? For what DAL levels is ML currently permissible under this framework, and what technical gap prevents it from reaching DAL A?

Curriculum Complete

You have covered all 10 modules of the AEROKNOW UAV Systems curriculum โ€” from platform taxonomy through flight dynamics, GNC, GNSS-denied navigation, sensor fusion, machine learning, deep learning, autonomy, swarm coordination, and systems engineering. The full arc from physical principles to deployable, certifiable systems.

โ—‚ Module 09: Swarm Systems โ†ฉ Return to Home