Generative AI–Enhanced Predictive Maintenance for Gearboxes: Physics-consistent modeling with GMF, sidebands and synthetic fault injection

Introduction

Gearbox failures are rare — but extremely expensive. From a Predictive Maintenance perspective, this creates a paradox: models are trained mostly with normal data, while their real value is detecting rare faults under changing operating conditions.

To address this, this work combines:

• gearbox vibration physics
• frequency-domain fault signatures
• Generative (Regenerative) AI for synthetic fault injection

Two-stage parallel gearbox model (Z₁–Z₄)

We consider a two-stage parallel gearbox:

• Stage 1: pinion Z₁ driving gear Z₂
• Stage 2: pinion Z₃ driving gear Z₄

Gear ratios

i₁ = Z₂ / Z₁

i₂ = Z₄ / Z₃

i_total = i₁ × i₂

Shaft speeds

Given input speed RPM_in:

RPM_int = RPM_in / i₁

RPM_out = RPM_int / i₂

Conversion to frequency (Hz)

f_in = RPM_in / 60

f_int = RPM_int / (60 × i₁)

f_out = RPM_int / (60 × i₁ × i₂)

Gear Mesh Frequencies (GMF)

The Gear Mesh Frequency is the dominant excitation in gearbox vibration.

Stage 1

GMF₁ = f_in × Z₁ (also GMF₁ = f_int × Z₂)

Stage 2

GMF₂ = f_int × Z₃ (also GMF₂ = f_out × Z₄)

Harmonics naturally appear at:

2×GMF, 3×GMF, 4×GMF, …

Sidebands and fault modulation

Localized faults (tooth crack, broken tooth) introduce amplitude modulation of GMF.

The vibration signal can be interpreted as:

x(t) = A(t) · cos(2π · GMF · t)

with modulation:

A(t) = A₀ · [1 + m · cos(2π · f_mod · t)]

This generates sidebands in the spectrum at:

GMF ± n · f_mod (n = 1, 2, …)

Practical interpretation

• Stage 1 faults → sidebands around GMF₁ spaced by f_in or f_int
• Stage 2 faults → sidebands around GMF₂ spaced by f_int or f_out

These components are often weak, load-dependent and easily hidden by noise.

Vibration signal composition

The simulated acceleration signal is built as:

x(t) = Σ A_k · cos(2π · f_k · t) + impulsive events + broadband noise

Where:

• f_k includes shaft orders (1×, 2×), GMF₁, GMF₂, harmonics and sidebands
• impulsive events increase kurtosis (cracks, broken teeth)
• broadband noise represents lubrication and environment effects

Why traditional AI struggles

In real plants:

P_train(x) ≠ P_test(x)

because speed, load and noise change over time.

This leads to:

• missed early faults
• unstable anomaly thresholds
• excessive false alarms

Generative AI as synthetic fault injection

Instead of simple augmentation, Generative AI learns the conditional behavior:

p(x | fault)

Using a Conditional Variational Autoencoder (CVAE):

• encoder learns a latent representation of vibration patterns
• decoder generates new fault-specific signals

Synthetic samples are generated from a normal latent distribution and conditioned by fault type.

The training dataset becomes:

D_aug = D_real ∪ λ · D_synthetic

where λ controls the amount of synthetic fault exposure.

Critical constraint: synthetic data must respect GMF position, harmonic structure and sideband spacing.

Engineering conclusion

Generative AI does not replace vibration analysis. When constrained by physics, it extends AI’s experience with rare gearbox failures, improving robustness under data scarcity and regime change.

Final thought

If your predictive model has never seen a gearbox tooth crack at GMF ± shaft speed…

why would it recognize one in the field?

Julio Panzera - Failure Modes Specialist | Vibration Analysis Expert | Applied AI Developer for Industrial Reliability