How is H(t) different from traditional vibration threshold alarms?

Threshold alarms are binary — alert above the limit, ignore everything below. H(t) is continuous — 0.95 and 0.45 are both "below the alarm threshold," but 0.45 means severe degradation requiring derated operation and scheduled maintenance. A continuous scalar lets you choose "slow down" rather than "full speed or full stop."

Why not just use a deep learning model for everything?

Industrial edge devices have hard constraints — RK3588 has 6 TOPS NPU, 32 channels must be processed within a 200ms window. CNN (25KB) + VAE (300KB) + Isolation Forest (50KB) delivers total inference under 5ms. No need for a large model. The right tool for the constraint.

What's the relationship between A[H(t)] and OEE?

A[H(t)] from PHM directly replaces the traditional Availability A in the OEE formula when reported to Online. Traditional A = runtime / planned time. A[H(t)]-driven A automatically derates when equipment health declines. OEE becomes a health-aware efficiency metric, not just a time-based one.

Aren't Isolation Forest and VAE redundant?

They operate on different inputs. VAE processes raw time-series signals (learning normal patterns). Isolation Forest processes handcrafted features (statistical + frequency-domain). They cross-validate from different perspectives. When both agree, confidence increases. When they disagree, the system is more cautious (lower confidence). Two angles, one conclusion.

PHM: Equipment Health Isn't Good or Bad. It's a Number Between 0 and 1.

TL;DR

Traditional maintenance knows two states: running and broken. But most equipment operates in the gray zone — still turning, efficiency declining. PHM digitizes health into H(t), a continuous scalar that drives adaptive capacity control from full-load to derated to forced shutdown.

What Does "Run to Failure" Actually Cost?

A wind turbine main bearing seizes. The bearing itself costs tens of thousands. But add tower disassembly, crane logistics, and weeks of lost generation — the total bill can reach hundreds of thousands. Bearings degrade over months before they seize. That window exists. But without quantitative monitoring, it passes unused.

Most factory equipment isn't simply "healthy" or "broken." It lives in the gray zone — elevated vibration below alarm thresholds, rising temperature within limits. Traditional threshold alerts either stay silent (degradation continues) or false-trigger (unnecessary downtime).

PHM tri-layer fusion architecture

Tri-layer fusion: CNN identifies known fault types → VAE+Isolation Forest cross-validate and discover unknown anomalies → Rule engine computes H(t) via weighted fusion

H(t): Turning "Health" Into a Computable Number

H(t) = Σ wᵢ · pᵢ. The CNN outputs fault-type probabilities pᵢ. The rule engine assigns dynamic weights wᵢ based on severity and development stage — a bearing fracture (sᵢ=0.9) is weighted far more heavily than minor poor lubrication (sᵢ=0.3). EWMA exponential smoothing removes transient noise.

H(t) ∈ [0, 1]. H=1: perfectly healthy. H=0: mandatory shutdown. It's not "vibration exceeded threshold → alarm." It's "health declining steadily, approximately 48 hours until H(t) reaches the stop threshold."

H(t) computation flow

H(t) computation: Probability distribution × dynamic health weights → weighted sum → EWMA smoothing → continuous health scalar

A[H(t)]: One Number, Three Capacity States

H(t) is the diagnostic result. A[H(t)] is the control signal. A segmented function maps H(t) to equipment capacity commands:

H(t) < H_stop (e.g., 0.3) → A=0. Forced shutdown. Trigger maintenance work order.
H_stop ≤ H(t) < H_full (e.g., 0.8) → A scales linearly from 0 to 1. Derated operation — reduce speed and load. Schedule preventive maintenance.
H(t) ≥ H_full → A=1. Full load. No intervention.

A[H(t)] segmented function

Three-state capacity regulation: A=0 forced shutdown → 0
When A[H(t)] Enters the OEE Formula, Efficiency Gets Real
Traditional OEE Availability A = actual runtime / planned runtime. But this doesn't reflect equipment health. Sub-healthy equipment running at full load still reports A=1 — until it breaks.
PHM reports A[H(t)] to Online. Online replaces the traditional A with A[H(t)] in the OEE formula. OEE now reflects "health-aware efficiency" — not just "the machine is spinning."
Tri-Layer vs. One Model: It's Not About Accuracy. It's About Coverage.
CNN excels at classifying known faults but can't detect novel ones. VAE excels at discovering unknown anomalies but can't classify them. Isolation Forest performs well on handcrafted features but has limited scope.
Tri-layer fusion isn't "three models running together" — it's each layer correcting the previous one's misses. CNN provides first-pass screening. VAE+Isolation Forest cross-validate (catching patterns CNN has never seen). The rule engine turns all outputs into a computable health value. Remove any layer, and you create blind spots in specific scenarios.
Tech Snapshot

Dimension Spec

Models CNN 25KB/0.3ms + VAE 300KB/3ms + Isolation Forest 50KB/0.1ms
Inference 204.8ms window (2048 pts @ 10KHz), 32 channels parallel, total <5ms
Quantization INT8 (CNN) + INT16 (VAE), RK3588 NPU 6 TOPS
Output H(t) ∈ [0,1] · A[H(t)] ∈ [0,1] · Normal/Abnormal/Unknown

Let the Equipment Speak
Vibration, temperature, current — equipment has been generating these signals all along. Nobody was listening. PHM turns them into two numbers: H(t) tells you how much time is left. A[H(t)] tells you how fast you can still run. The gap between "run to failure" and "know when to maintain" isn't about sensors. It's about the digitization step.

Dimension	Spec
Models	CNN 25KB/0.3ms + VAE 300KB/3ms + Isolation Forest 50KB/0.1ms
Inference	204.8ms window (2048 pts @ 10KHz), 32 channels parallel, total <5ms
Quantization	INT8 (CNN) + INT16 (VAE), RK3588 NPU 6 TOPS
Output	H(t) ∈ [0,1] · A[H(t)] ∈ [0,1] · Normal/Abnormal/Unknown