In the heavy engineering sector—spanning offshore rigs to mining belts—the gearbox is a critical point of failure. Despite advances in metallurgy and machining, drivetrain monitoring remains fundamentally reactive.
Traditional Condition-Based Monitoring (CBM) relies heavily on Fast Fourier Transforms (FFT) to track Gear Mesh Frequency (GMF) and RMS vibration thresholds. This approach has an important limitation: it generally waits for degradation to become energetic enough to produce a measurable vibration signature. By the time a conventional vibration sensor flags a significant anomaly, micro-pitting may already have progressed toward macroscopic spalling, at which point the metallurgical damage is irreversible.
A more sensitive prognostic approach is to move beyond vibration amplitude and examine the geometric structure of the machine’s acoustic and vibration signature. Topological Data Analysis (TDA), combined with nonlinear dynamics and multivariate anomaly detection, provides a framework for investigating changes in the underlying dynamical state of a transmission.
Consider a heavy-duty crane hoist operating under a 50-metric-ton load, corresponding to approximately 490.5 kN of wire-rope tension. The gearbox uses a 1:30 reduction ratio. A 1500 RPM motor therefore produces an output speed of approximately 50 RPM at the bull gear.
Under this loading, Hertzian contact pressure at the pinion–bull-gear interface can reach the gigapascal range. The system relies on elastohydrodynamic lubrication (EHL) to maintain separation between the contacting surfaces, with ISO VG 150 gear oil (HV-150) assumed for this illustrative case.
At elevated local mesh temperatures, the kinematic viscosity of the lubricant decreases substantially. The temperature dependence can be represented using the ASTM D341 / Walther relationship (often referred to as the Walther equation). At approximately 85°C, the viscosity may fall far below its nominal 150 cSt value, depending on the actual lubricant formulation and viscosity index.
Within the highly loaded contact zone, pressure can significantly increase the effective dynamic viscosity of the lubricant. This pressure–viscosity behaviour is commonly represented by the Barus relationship. The resulting EHL film is essential to maintaining surface separation and preventing excessive asperity interaction.
Water ingress is a common contamination mechanism in industrial gearboxes. When water enters the lubricant, agitation by the meshing gears can produce a water-in-oil dispersion or emulsion.
Water contamination can reduce lubricant performance and compromise film formation, although the precise mechanism is more complex than a simple change in bulk modulus. Loss of film thickness, additive degradation, corrosion, foaming, and altered rheology can all contribute to increased surface interaction.
As the lubricant film becomes inadequate, asperity contact can increase. Repeated micro-contacts can initiate or accelerate micro-pitting and surface fatigue. Progressive surface damage changes the effective contact geometry and can introduce additional modulation, impulsive components, and changes in the dynamics of the gear mesh.
A key premise of the proposed approach is that these early changes may first appear as changes in the structure of the measured signal rather than as a large increase in its overall RMS amplitude.
To investigate whether gear wear can be detected before a conventional vibration-amplitude alarm is reached, the raw acoustic or vibration signal can be reconstructed into a multidimensional representation of the underlying dynamics.
For an illustrative 1:30 crane-hoist gearbox sampled at 10 kHz, consider two operating states: State A (healthy lubricant and gear mesh) and State B (water-contaminated lubricant with early micro-pitting). The example assumes that RMS vibration remains below a nominal 4.5 mm/s alarm threshold in both states.
The following methods form a complementary analytical chain: delay-coordinate embedding reconstructs the state space; nonlinear-dynamics metrics quantify trajectory stability; Mahalanobis distance evaluates multivariate deviation; and persistent homology measures changes in geometric structure.
Rather than representing the measurement only as a one-dimensional time series or frequency spectrum, the signal can be reconstructed into a multidimensional phase-space representation using Takens’ delay-coordinate embedding.
For a scalar signal x(t), the reconstructed state vector is:
Once a reconstructed state space is available, local dynamical behaviour can be studied through an estimated Jacobian matrix. The Jacobian describes how small perturbations in the reconstructed state evolve locally.
From an appropriate dynamical model or local linear approximation, Lyapunov exponents characterise the average rate at which nearby trajectories converge or diverge. A positive maximal Lyapunov exponent is commonly associated with sensitive dependence on initial conditions and chaotic dynamics.
In this illustrative scenario:
• State A (healthy): λmax = 0.12
• State B (micro-pitting): λmax = 1.84
These values should be treated as simulated/example outputs, not as experimentally established thresholds. Their value is in demonstrating the proposed diagnostic logic: a change in dynamical stability may be detectable even when RMS amplitude remains below an alarm limit.
A mechanical anomaly should be distinguished from normal operating changes, such as load or speed variations. Mahalanobis distance provides a multivariate measure of how far an observation lies from a reference distribution while accounting for covariance among variables.
For a feature vector containing quantities such as motor RPM, acoustic amplitude, and structural vibration, the distance is:
Persistent homology provides a way to characterise the shape of point clouds in reconstructed state space across multiple spatial scales. In particular, H1 features represent loop-like structures in the filtration.
The persistence diagram of a healthy reference state can therefore be compared with that of a potentially worn state. Wasserstein distance provides one quantitative measure of the difference between the two persistence diagrams.
For the illustrative example:
• Baseline Wasserstein-distance variation: 0.02–0.05
• State B Wasserstein distance: 2.84
Again, these numerical values are simulation/example values and must be calibrated against experimental data before being used as an industrial alarm or prognosis threshold.
In the illustrative scenario, a large Mahalanobis distance and a substantial Wasserstein distance coincide with a change in the reconstructed dynamical structure. At the same time, conventional RMS vibration remains below the nominal alarm threshold.
The engineering proposition is therefore not that topology replaces FFT, but that topology and nonlinear-dynamics features can complement conventional condition monitoring by exposing structural changes that amplitude-only indicators may miss.
The practical objective is to identify the transition from normal operation to early degradation as close as possible to the onset of the underlying physical change. If validated experimentally, such a framework could support a shift from reactive replacement toward earlier intervention, improved maintenance planning, and higher equipment availability.
The central hypothesis is simple: before a gearbox produces a large vibration alarm, its measured dynamics may already have changed shape. Detecting that change may provide an earlier window for action.
Takens’ Delay-Coordinate Embedding
xᵥ(t) = [x(t), x(t + τ), x(t + 2τ), …, x(t + (d − 1)τ)]
MacCoull–Walther Equation (Kinematic Viscosity)
log₁₀(log₁₀(ν + 0.7)) = A − B · log₁₀(T)
Barus Equation (Dynamic Viscosity)
ηₚ = η₀ · e^(αP)
Mahalanobis Distance
D_M(x) = √[(x − μ)ᵀ · Σ⁻¹ · (x − μ)]
Wasserstein Distance
Wₚ(D_healthy, D_worn) = [inf_γ (Σ ||x − γ(x)||ᵖ)]^(1/p)
Anureet Das
Marine Engineer
Shipping Corporation of India
