Using FMEA for Autonomous Diagnostics, Predictive Maintenance and Cross-Layer Security: A Reference Architecture

When an on-board diagnostic or a predictive-maintenance model is listed as a control in a DFMEA or PFMEA, the main question is whether it really lowers Occurrence (prevention) or only improves Detection. The answer also decides whether the diagnostic itself, and the data path it depends on, needs its own failure and threat analysis. This article ties the 7-step method of the AIAG & VDA FMEA Handbook (1st ed., 2019) and IEC 60812:2018 to ISO 26262:2018 diagnostic coverage, ISO 13374 and ISO 17359:2018 condition-monitoring data flows, and ISO/SAE 21434:2021 TARA, so that each control credited in FMEA Tool can be justified and audited.

Prevention vs. detection, decision criteria

credit a prognostic model as a prevention control (lower O) only if it triggers a maintenance action before the failure mode can occur, and the lead time it provides is longer than the time needed to act on it (prognostic horizon > response time). If it only flags a failure that has already started, rate it under Detection. If the input data can be spoofed, delayed or missing, add the diagnostic as its own Structure element in Step 2 and analyse its failure modes there: stale data, a stuck sensor, model drift, or a manipulated message.

Pre-implementation checklist

before you rate a control, collect the sensor sampling rate and latency, the ISO 13374 processing layer where the decision is made (data acquisition → state detection → health assessment → prognostics), the expected false-alarm and missed-detection behaviour, any ISO 26262 diagnostic-coverage targets, and the cybersecurity goals from your ISO/SAE 21434 TARA. Record each of these as a reusable catalog entry so the same assumptions apply across every DFMEA and PFMEA.

Illustrative analytical scenario (not measured data)

a spindle-bearing failure mode has S=8. A vibration-trend model with an assumed 72 h prognostic horizon and a 24 h maintenance window justifies moving O from 6 to 3. The CAN path that carries its alerts has no authentication, so a separate failure-mode entry ('alert suppressed or forged') stays at D=7 until message authentication is added, for example AUTOSAR SecOC. Recovery plan: if the model's inputs are lost, fall back to fixed-interval inspection and revert to the original O rating. Re-run Action Priority, then export the Excel file as audit evidence.

How do I verify in FMEA Tool that a predictive-maintenance or diagnostic control deserves the lower Occurrence or Detection rating I gave it?

Write down the assumptions behind the rating: prognostic horizon versus response time, data latency, the expected false-alarm and missed-detection rates, and the TARA cybersecurity goals that protect the data path. Model the diagnostic as a separate Structure element and analyse its own failure modes, then confirm that Action Priority still makes sense once those failures are included. Validate with field or bench evidence that you actually collected; never use vendor claims as proof. If the evidence is missing, or monitoring degrades in operation, go back to the original rating and to time-based inspection, update the catalog entry, and export a new Excel revision so the change can be traced.

Back to FMEA Tool · Domains

← Back to FMEA Tool
Engineering & Architecture Brief

Using FMEA for Autonomous Diagnostics, Predictive Maintenance and Cross-Layer Security: A Reference Architecture

When an on-board diagnostic or a predictive-maintenance model is listed as a control in a DFMEA or PFMEA, the main question is whether it really lowers Occurrence (prevention) or only improves Detection. The answer also decides whether the diagnostic itself, and the data path it depends on, needs its own failure and threat analysis. This article ties the 7-step method of the AIAG & VDA FMEA Handbook (1st ed., 2019) and IEC 60812:2018 to ISO 26262:2018 diagnostic coverage, ISO 13374 and ISO 17359:2018 condition-monitoring data flows, and ISO/SAE 21434:2021 TARA, so that each control credited in FMEA Tool can be justified and audited.

Prevention vs. detection, decision criteria

credit a prognostic model as a prevention control (lower O) only if it triggers a maintenance action before the failure mode can occur, and the lead time it provides is longer than the time needed to act on it (prognostic horizon > response time). If it only flags a failure that has already started, rate it under Detection. If the input data can be spoofed, delayed or missing, add the diagnostic as its own Structure element in Step 2 and analyse its failure modes there: stale data, a stuck sensor, model drift, or a manipulated message.

Pre-implementation checklist

before you rate a control, collect the sensor sampling rate and latency, the ISO 13374 processing layer where the decision is made (data acquisition → state detection → health assessment → prognostics), the expected false-alarm and missed-detection behaviour, any ISO 26262 diagnostic-coverage targets, and the cybersecurity goals from your ISO/SAE 21434 TARA. Record each of these as a reusable catalog entry so the same assumptions apply across every DFMEA and PFMEA.

Illustrative analytical scenario (not measured data)

a spindle-bearing failure mode has S=8. A vibration-trend model with an assumed 72 h prognostic horizon and a 24 h maintenance window justifies moving O from 6 to 3. The CAN path that carries its alerts has no authentication, so a separate failure-mode entry ('alert suppressed or forged') stays at D=7 until message authentication is added, for example AUTOSAR SecOC. Recovery plan: if the model's inputs are lost, fall back to fixed-interval inspection and revert to the original O rating. Re-run Action Priority, then export the Excel file as audit evidence.

Technical FAQ: How do I verify in FMEA Tool that a predictive-maintenance or diagnostic control deserves the lower Occurrence or Detection rating I gave it?

Write down the assumptions behind the rating: prognostic horizon versus response time, data latency, the expected false-alarm and missed-detection rates, and the TARA cybersecurity goals that protect the data path. Model the diagnostic as a separate Structure element and analyse its own failure modes, then confirm that Action Priority still makes sense once those failures are included. Validate with field or bench evidence that you actually collected; never use vendor claims as proof. If the evidence is missing, or monitoring degrades in operation, go back to the original rating and to time-based inspection, update the catalog entry, and export a new Excel revision so the change can be traced.

Related Architecture Specifications (FMEATOOL): fmeatools.com · freefmea.com

More Technical Articles on fmeatool.com

DFMEA for Edge Compute Offloading: Deadlines, Protocol Conformance and SLA Evidence →Architecture Decision Guide: Local-First DFMEA/PFMEA Execution, Fault-Isolated Storage, and Zero-Trust Data Portability →Timing Failure Modes in DFMEA: Deterministic Telemetry, Sub-Millisecond Latency Budgets and HIL vs. Digital-Twin Detection Controls →FMEA Reference Architecture: Aligning DFMEA/PFMEA Data with AIAG-VDA, SAE J1739 and IEC 60812 →