← Polyodos · Methods wiki · Industrial and technical risks
Method

🔁 STPA

System-Theoretic Process Analysis: a hazard analysis that starts from how a system is controlled, rather than from its failures, to find unsafe control actions and the scenarios that lead to them.

1What is STPA?

STPA (System-Theoretic Process Analysis) is a hazard analysis method developed by Nancy Leveson and her team at MIT. It is based on the STAMP accident model (System-Theoretic Accident Model and Processes), presented in Engineering a Safer World (MIT Press, 2011). The process is described step by step in the STPA Handbook by Nancy Leveson and John Thomas (2018), which serves as the reference.

Its starting point differs from that of traditional methods. FMEA and fault tree analysis look for component failures and their combinations. STPA treats safety as a control problem: an accident occurs when safety constraints are not enforced on the system, including when no component has failed. A command that is correct in itself but given at the wrong time, an operator who relies on incorrect information, two controllers that interact badly: these cases are at the heart of STPA.

It relies on four concepts:

ConceptDefinition
LossSomething stakeholders do not want to suffer: death or injury, environmental damage, loss of production, damage to reputation.
HazardA system state that, under worst-case environmental conditions, leads to a loss. It is neither a cause (“failed sensor”) nor a loss.
Control structureThe model of the controllers (people, automated systems, organisations), the control actions they issue and the feedback they receive.
Unsafe control action (UCA)A control action that, in a particular context, leads to a hazard.

2When to use STPA

STPA is used in particular in aviation, space, automotive and defence. It does not replace existing methods: it complements them, and its scenarios can then be rated or quantified with a PHA, a bow-tie or a LOPA.

3The STPA method, step by step

The STPA Handbook describes four steps.

Step 1 — Define the purpose of the analysis

List the losses (L1, L2…), then the system-level hazards (H1, H2…), each linked to the losses it can lead to. Derive the system-level safety constraints: each constraint is the hazard stated in reverse, i.e. what must remain true to avoid it.

Step 2 — Model the control structure

Represent the controllers and the controlled process, the control actions (downward arrows) and the feedback (upward arrows). Each controller acts according to its process model, i.e. what it believes the state of the system to be.

Step 3 — Identify unsafe control actions (UCA)

For each control action, examine the four types of UCA:

TypeDescription
Not providedThe action was needed and was not given.
Provided when unsafeThe action was given although it leads to a hazard in this context.
Wrong timing or orderThe action was given too early, too late, or in the wrong order relative to other actions.
Stopped too soon or applied too longFor a continuous action: it stopped before it was complete, or went on beyond what was needed.

A UCA is always written with its context (“the controller does X when Y”) and the hazards it leads to. Each UCA yields a controller constraint, its reverse statement.

Step 4 — Identify loss scenarios

For each UCA, look for the causal scenarios that explain it. There are two families: why would the controller issue the unsafe action (flaw in its control algorithm, incorrect process model, feedback missing, delayed or wrong), and why would a correct action be improperly executed or have no effect (actuator, transmission, unexpected behaviour of the process). These scenarios lead to design requirements and measures.

Feedback is often left out of the control structure. Without it, you miss a large share of the scenarios: those in which the controller acts logically on false information.

4The security extension: STPA-Sec

STPA-Sec applies the same approach to cybersecurity. It was proposed by William Young and Nancy Leveson. In addition to accidental causes, it considers control actions and feedback that are deliberately altered: injected commands, falsified measurements, masked feedback. The resulting constraints are security constraints in the sense of protection against malicious acts.

The benefit is to handle safety and cybersecurity on a single control structure: the same controller can receive a wrong command through a design flaw or through an attack, and the constraint that protects the process is often the same.

5Example: a road tanker loading station

A PLC controls the loading of 30 m³ road tankers with a flammable liquid: it starts and stops the loading pump and opens the inlet valve. It receives the loaded volume from a flow meter, the tank high level from a level probe, and the “hose connected” status. The driver connects the hose and starts loading from a control panel. Fictitious data.

Losses, hazards and constraints

Ref.StatementLink
L1Death or injury of a person
L2Soil or water pollution
H1The tank is filled beyond its capacityL1, L2
H2Product is released outside the closed loading circuitL1, L2
H3The truck leaves the station with the hose still connectedL1, L2
SC1The quantity loaded must never exceed the available capacity of the tankH1
SC2Product must flow only in a circuit that is connected and open from end to endH2

UCAs for the control action “start / keep running the loading pump” (PLC)

TypeUCA or justificationHazard
Not providedRuled out: without start-up there is no loading and no hazard identified (only a loss of production, not retained as a loss).—
Provided when unsafeUCA-1: the PLC starts the pump when the hose is not connected.H2
Wrong timing or orderUCA-2: the PLC starts the pump before the inlet valve is open.H2
Stopped too soon or applied too longUCA-3: the PLC keeps the pump running after the loaded volume has reached the available capacity.H1

Controller constraint derived from UCA-3: the PLC must stop the pump as soon as the loaded volume reaches the available capacity or the high-level probe is activated.

Causal scenarios for UCA-3

OriginScenarioPossible requirement
Controller (process model)The PLC believes the tank is empty whereas it still holds 4 m³ from a previous load: it aims for 30 m³ instead of 26 m³.Entry and check of the residual quantity before loading
FeedbackThe high-level probe is fouled and does not signal that the level has been reached.Probe tested at each loading, shutdown independent of the flow meter
Control pathThe stop command is issued but the pump contactor remains stuck closed.Back-up closure of the inlet valve, on a separate command
Controlled processThe product, warmer than expected, expands: the measured volume underestimates the filling.Temperature correction of the loaded volume
Malicious act (STPA-Sec)The flow meter reading is falsified through unauthorised access to the PLC network.High-level probe hard-wired directly to the shutdown, off the network

These scenarios can then be rated in a PHA or carried into a bow-tie, where the requirements become barriers.

In Polyodos, the STPA / STPA-Sec method is part of the Industrial risks module: losses, hazards and constraints; a control structure derived from the facility model; a UCA grid by control action and by type; then loss scenarios classified by origin, together with their malicious variants. Items can be sent to the PHA, the bow-tie, the fault tree or the LOPA. See the Industrial risks demo or try Polyodos free for 30 days.

6Common mistakes

MistakeWhy it is a problem
Writing causes instead of hazards“Failed level sensor” is a cause, not a hazard. Hazards describe a system state; causes come in step 4.
Leaving out the context of a UCA“The PLC starts the pump” is true of almost all normal operation. Without the “when…”, the UCA can be neither verified nor used.
Omitting feedbackA control structure without upward arrows hides the scenarios of missing or wrong information, which are often the most instructive.
Sticking to failuresReducing every scenario to a component failure amounts to redoing an FMEA: you lose what makes STPA worthwhile, namely interactions and design flaws.
Filling in the four types mechanicallyNot every cell yields a UCA. A cell ruled out with its justification is better than an artificial UCA.

7Further reading

Learn: the online course STPA on aurorisk.com.

Practise: the user guide for the Industrial risks module of Polyodos.

FR