The Architecture

Chapter 3 — The Architecture

ISO 13374 Alignment

RAPID AI’s architecture aligns with ISO 13374 (Condition Monitoring and Diagnostics of Machines), which defines six processing levels for machine condition data. This alignment ensures that the system meets international standards for diagnostic completeness while extending beyond what the standard requires:

ISO 13374 Level	Description	RAPID AI Module	Pipeline Stage
L1 Data Acquisition	Sensor measurement	External (sensors/historians)	Input
L2 Data Manipulation	Feature extraction, quality gating	Module A	GUARD
L3 State Detection	Pattern recognition, fault identification	Modules B, B.2, B.3	SENSE
L4 Health Assessment	Severity scoring, fusion	Module C, AESF	FUSE
L5 Prognostics	RUL estimation, degradation projection	Module D	ACT
L6 Advisory Generation	Maintenance recommendation	Module E	ACT
Beyond ISO	Reliability modeling	Module F	INSIGHTS
Beyond ISO	Design-out engineering	Module G	INSIGHTS

Modules F and G extend beyond the ISO 13374 framework. Module F adds condition-adjusted Weibull reliability modeling — combining population statistics with real-time sensor evidence. Module G adds Contradiction Driven Engineering — a capability that no international standard covers because it bridges diagnostics and design, transforming field intelligence into engineering change.

Three Intelligence Layers

RAPID AI organizes its reasoning into three progressive intelligence layers. Each layer answers a deeper question about the machine’s condition, and each layer’s output becomes the next layer’s input.

Signal Intelligence (Modules A, B, B.2, B.3)

Question answered: What is the machine doing?

Raw sensor signals enter Module A, which validates data quality and extracts statistical features. Module B applies 119 physics-based fault rules across 12 component categories to detect failure initiators. Module B.2 analyzes trends over time windows to classify whether a parameter is stable, drifting, accelerating, stepping, or chaotic. Module B.3 computes Shannon entropy across spectral, temporal, and directional domains, producing a Stability Index (SI) that quantifies how freely energy flows through the system.

Together, these modules transform raw vibration into validated, classified, and entropy-scored fault evidence. Detailed technical reference: Chapter 4.

Stability Intelligence (Modules C, AESF)

Question answered: Is the system stable or drifting toward failure?

Module C fuses per-component evidence into a System Stability Index (SSI) using profile-weighted block aggregation. It combines fault detection confidence, trend severity, and entropy stability into a single bounded score representing system-level health. The AESF (Acceleration-Entropy Stability Framework) goes further, computing four indices — SI, EI, CSS, JII — that characterize the machine’s operating “climate” across coherence, disorder, coupling, and transient instability dimensions. The AESF assigns one of five stability states (S0 through S4), from Coherent Stable to Critical Transition.

Detailed technical reference: Chapter 5.

Engineering Intelligence (Modules D, E, F, G)

Question answered: What mechanism explains this, and what should we do?

Module D maps the fused evidence to specific failure mechanisms and health stages. Module E scores maintenance priority and selects actions from a catalog of 15+ interventions. Module F provides condition-adjusted Weibull RUL estimates and 30-day failure probability. Module G, triggered only for chronic or high-impact failures, applies Contradiction-Driven Engineering to identify design-level root causes that no amount of maintenance can fix.

Detailed technical reference: Chapter 6.

The Complete Pipeline

                        REQUEST (JSON payload)
                            |
                            v
            +-------------------------------+
            |  MODULE 0 -- Config (CONTEXT) |
            |  Asset profile, sensors,      |
            |  baselines, thresholds        |
            +-------------------------------+
                            |
                            v
            +-------------------------------+
            |  MODULE A -- GUARD            |
            |  Signal validation, DSP,      |
            |  feature extraction,          |
            |  quality gating (16 rules)    |
            +-------------------------------+
                            |
               (blocked if hard gate fails)
                            |
          +-----------------+-----------------+
          |                 |                 |
          v                 v                 v
   +------------+   +------------+   +---------------+
   | MODULE B   |   | MODULE B.2  |   | MODULE B.3    |
   | Fault det. |   | Trend      |   | SEDL Entropy  |
   | 121 rules  |   | analysis   |   | SE, TE, DE    |
   | 12 types   |   | 5 classes  |   | SI, states    |
   +------------+   +------------+   +---------------+
          |                 |                 |
          +-----------------+-----------------+
                            |
                            v
            +-------------------------------+
            |  MODULE C -- FUSE (SSI)       |
            |  Profile-weighted fusion,     |
            |  7 BSR rules, override logic  |
            +-------------------------------+
                            |
                   +--------+--------+
                   |                 |
                   v                 v
            +-----------+   +---------------+
            |   AESF    |   | MODULE D      |
            | SI/EI/    |   | Health stage, |
            | CSS/JII   |   | mechanism ID, |
            | S0-S4     |   | RUL estimate  |
            +-----------+   +---------------+
                   |                 |
                   +---------+-------+
                             |
                             v
            +-------------------------------+
            |  MODULE E -- ACT             |
            |  Priority scoring,           |
            |  maintenance planning,       |
            |  action catalog (15+ acts)   |
            +-------------------------------+
                             |
                             v
            +-------------------------------+
            |  MODULE F -- RUL & Reliability|
            |  Condition-adjusted Weibull,  |
            |  P-F interval, risk index     |
            +-------------------------------+
                             |
                  (conditional trigger)
                             |
                             v
            +-------------------------------+
            |  MODULE G -- CDE             |
            |  Contradiction-Driven         |
            |  Engineering (design-out)     |
            +-------------------------------+
                             |
                             v
                    RESPONSE (JSON)
              Scores + Insights + Actions

Each module reads only the previous module’s output. No module overlaps another. Data flows strictly forward. If any module fails, the pipeline returns partial results with degraded confidence rather than nothing.

The Rule System (451+ Rules)

RAPID AI’s intelligence is encoded in 451+ parseable rules organized into four layers:

Layer	ISO Level	Purpose	Count	Pipeline Stage
GUARD	L2 Data Manipulation	Data quality gates	16	Module A
SENSE	L3 State Detection	Signal intelligence	275+	Modules B, B.2, B.3
FUSE	L4/L5 Health + Prognostics	Fusion and staging	66	Modules C, D
ACT	L6 Advisory Generation	Action and priority	38	Module E

Additionally, the ACT layer includes 300 imperfection rules encoding the Theory of Imperfections across eight equipment types (see Chapter 8).

Rules Are Data, Not Code

A critical architectural decision separates RAPID AI from most diagnostic systems: rules are data, not code. Every rule is a parseable expression stored in JSON/YAML/CSV, never embedded in application logic. The rule engine is a custom recursive-descent parser that evaluates these expressions safely — no eval() is ever called on rule content.

This means:

Rules can be audited by reliability engineers without reading source code
Rules can be versioned and tracked in the same way as any data artifact
Adding new fault detection capability means adding rows to rule tables, not modifying service code
Every recommendation traces back through a chain of evidence to specific measurements and physical principles

Rule Schema

Every SENSE rule follows a consistent schema:

Field	Description
`rule_id`	Unique identifier (e.g., AFB01)
`initiator`	Root cause description
`detection_rule`	Ratio and amplitude conditions
`supplementary_indicators`	Additional confirming signals
`severity_logic`	Early / Mid / Late stage progression
`physics_basis`	Energy flow explanation per Theory of Imperfections
`diagnosis`	Human-readable diagnosis
`recommended_actions`	Corrective steps

Example — AFB01 (Excessive Preload):

Detection rule: A/H > 1.2; V/H < 0.8
Physics basis: “High preload increases axial stiffness, so axial vibration rises first”
Supplementary: crest_factor > 1.4
Severity cascade: Early (A/H > 1.2) -> Mid (RMS > 1.3x) -> Late (crest_factor > 2.0)
Diagnosis: “Bearing preload — check torque and shims”

The physics basis field is the critical differentiator. It is not just a label — it is the engineering explanation that transforms an alert into an understanding.

v2 Architecture (Current)

The v1 architecture had 8 separate modules with independent interfaces. The v2 architecture consolidates these into 5 pipeline stages while preserving the same underlying physics:

v2 Stage	v1 Modules	Function
GUARD	Module A	Validate, extract, gate
SENSE	Modules B + B.2 + B.3	Detect, trend, entropy
FUSE	Module C + AESF	Aggregate, assess stability
ACT	Modules D + E	Diagnose, plan, recommend
INSIGHTS	Modules F + G	RUL, reliability, design-out

Confidence Propagation

Confidence propagation is the thread running through the entire pipeline. Each stage computes an effective severity score (S_eff) and passes it downstream:

S_eff = (Sum of severity_i * confidence_i) / (Sum of confidence_i)

This weighted average ensures that high-confidence evidence dominates over uncertain signals, and that poor data quality (from Module A) degrades every downstream result proportionally. The confidence scoring standard is detailed in Chapter 12.

Evidence Compounding

Module B uses an independent evidence compounding formula that allows multiple weak signals to compound into strong evidence:

confidence = 1 - (1 - c_m)(1 - c_t)(1 - c_s)

Where:
  c_m = rules confidence (from matched initiator rules)
  c_t = trend severity score (from B.2 trend analysis)
  c_s = 1 - SI (stability gap from B.3 entropy)

If any single evidence source is highly confident, overall confidence rises. Multiple weak signals compound into strong evidence. This mirrors how expert diagnosticians think: a bearing with slightly elevated kurtosis (c_m = 0.4), a drifting temperature trend (c_t = 0.3), and declining stability (c_s = 0.3) produces a combined confidence of 0.71 — above the RCM activation threshold of 0.70.

State Transition Hysteresis

State transitions use hysteresis to prevent oscillation between adjacent states:

Transition	S_eff Threshold	Required Windows	Hysteresis Band
-> Degrading	> 0.55	3	+/- 0.08
-> Unstable	> 0.70	3	+/- 0.08
-> Critical	> 0.85	2	+/- 0.08
<- Return (de-escalation)	< threshold - 0.08	5	—

Escalation requires S_eff exceeding the threshold for a sustained number of evaluation windows (typically 2-3). De-escalation requires S_eff dropping below (threshold - 0.08) for 5 windows. This prevents “alarm fatigue” from borderline readings that flicker between states. The asymmetry is deliberate: it is easier to escalate than to de-escalate, because false negatives (missing a real fault) are more costly than false positives (investigating a borderline condition).

IP Protection Boundary

The API enforces a strict separation between protected internal logic and exposed client outputs:

Protected (never sent to clients):

Entropy formulas and weights
Rule scoring logic and BSR rules
RUL model parameters and adjustment coefficients
Confidence computation math
Weight coefficients for SSI fusion
AESF index formulas

Exposed (client-facing outputs):

Health stage (Healthy/Degrading/Unstable/Critical)
Severity level with score 0-1
Risk index 0-100
Action plan with justifications
RUL in days
30-day failure probability
AI-generated narrative insights
Confidence score 0-1

The client receives actionable intelligence without gaining access to the proprietary diagnostic engine. This boundary is enforced at the API serialization layer — internal computation objects are never directly serialized to the response.

The next chapters provide module-by-module technical depth: Chapter 4 covers Signal Intelligence (Modules A, B, B.2, B.3), Chapter 5 covers Stability Intelligence (Modules C, AESF), and Chapter 6 covers Engineering Intelligence (Modules D, E, F, G).

Standards Alignment

Standard	Relevance to This Chapter
ISO 13374 — Condition monitoring and diagnostics of machines	The complete pipeline architecture (Modules A through G) explicitly implements and extends ISO 13374’s six processing levels (L1-L6), with Modules F and G going beyond the standard into reliability modeling and design-out engineering.
ISO 10816-1/ISO 20816-1 — Mechanical vibration evaluation	Module A’s ISO zone classification directly implements ISO 20816-1 vibration severity zones (A/B/C/D) for broadband vibration assessment across four machine groups.
ISO 17359 — General guidelines for condition monitoring	The three intelligence layers (Signal, Stability, Engineering) implement ISO 17359’s recommended progression from data collection through diagnosis to prognosis and advisory generation.
MIMOSA OSA-CBM — Open System Architecture for CBM	The v2 pipeline stages (GUARD, SENSE, FUSE, ACT, INSIGHTS) map to OSA-CBM’s layered architecture for interoperable condition-based maintenance systems.

Changelog

Version	Date	Author	Changes
2.1.0	2026-03-17	Rick D	Added standards alignment, living doc metadata, changelog
2.0.0	2026-03-17	Rick D	Enriched with production codebase content
1.0.0	2026-03-17	Rick D	Initial chapter creation

Next: Chapter 4 — Signal Intelligence Previous: Chapter 2 — The Philosophy