Spiral bevel gears are used where a drivetrain must turn a corner under high torque, including the final drives of agricultural harvesters. Their curved teeth engage progressively, so contact moves across the tooth surface rather than across the full face at once. This contact geometry gives high load capacity and relatively smooth operation, but it also spreads the vibration signature over several contact lines and mesh components. When a root crack or surface spall appears, the resulting impact is modulated by the same meshing cycle and by the transmission path to the sensor. A single envelope peak therefore becomes an unreliable diagnostic indicator.
The diagnostic problem is not simply a rise in vibration energy. It is the separation of a weak, repetitive impact from mesh harmonics, speed fluctuations, bearing tones and structural resonances. Wavelet packet decomposition (WPD) is well suited to this kind of nonstationary signal because it divides both low- and high-frequency regions into subbands. Its weakness is equally clear: the mother wavelet and decomposition level are usually fixed before the fault is known. A physics-guided WPD strategy addresses this by using a dynamic model of the gear pair to generate a reference spectral signature, then selecting the wavelet basis whose decomposition best preserves that signature. The measured signal remains the object of diagnosis; the model supplies a prior for how the signal should be represented.
Why spiral bevel gears are a difficult diagnostic target
For a gear pair, the mesh frequency is set by tooth count and shaft speed:
$$ f_m = z f_s, $$
where \(z\) is the number of teeth and \(f_s\) is the shaft frequency. For a 37-tooth gear running at 600 rpm, \(f_s = 10\) Hz and \(f_m = 370\) Hz. A local tooth defect introduces a stiffness loss once per revolution. In the spectrum, that periodic stiffness change produces sidebands around the mesh frequency and its harmonics, spaced by the shaft frequency. In the envelope spectrum, it should produce a line at the shaft rate and at its harmonics.
In practice, the raw envelope often fails to make this pattern clear. The impact energy is distributed across several mesh sidebands, and random components from the gearbox and sensor path can be stronger than the shaft-rate line. The problem becomes more severe in agricultural machinery, where load varies with soil conditions and the gearbox may operate at low speed with high torque. Under these conditions, a fixed time-frequency representation can smear the impulse or spread it across too many subbands.
Wavelet packets and the basis-selection problem
Wavelet analysis, introduced by Morlet in 1981, uses a scalable, finite-duration window to match local signal features. A mother wavelet is shifted and scaled as
$$ \psi_{a,b}(t)=\frac{1}{\sqrt{a}}\psi\left(\frac{t-b}{a}\right),\quad a>0,\ b\in\mathbb{R}, $$
where \(a\) is the scale and \(b\) is the translation. The discrete wavelet transform is an inner product between the signal \(f(t)\) and shifted, scaled versions of the mother wavelet:
$$ W_f(m,n)=\langle f(t),\psi_{m,n}(t)\rangle. $$
WPD extends this by recursively splitting both the low-frequency and high-frequency subbands. For level \(i\), band index \(j\), and low- and high-pass filter coefficients \(h_0\) and \(h_1\), the coefficient recursion can be written as
$$ d_n^{i,2j}=\sum_l h_0(l-2n)d_l^{i-1,j},\qquad
d_n^{i,2j+1}=\sum_l h_1(l-2n)d_l^{i-1,j}. $$
The decomposition level determines the number of bands and the frequency resolution. A three-level WPD produces eight bands. The mother wavelet determines the shape of the analysis window: a basis that is too smooth will smear a sharp impact, while a basis that is too oscillatory can introduce artificial ringing. Neither choice can be optimized reliably from the vibration signal alone when fault data are scarce, which is the central limitation of conventional WPD in gear diagnosis.
A reduced-order gear model as a physical prior
The physical prior used here is a six-degree-of-freedom lumped-parameter model of the gear pair:
$$ M\ddot{x}+C\dot{x}+Kx=F, $$
where \(M\), \(C\) and \(K\) are the mass, damping and support-stiffness matrices, and \(F\) is the load vector. The contact force and friction force are represented as
$$ F=k(t)\gamma_0 e(t)+C_m\gamma_0\dot e(t),\qquad F_f=\mu F, $$
with the transmission error
$$ e(t)=R_{bp}\theta_p-R_{bg}\theta_g+x_p-x_g. $$
Here \(k(t)\) is the time-varying mesh stiffness, \(C_m\) is the mesh damping, \(\gamma_0\) is a nonlinear clearance function, \(\mu\) is the friction coefficient, \(R_b\) is the base-circle radius, and subscripts \(p\) and \(g\) denote the pinion and gear. The mesh stiffness is obtained from the series combination of bending, foundation and contact compliances:
$$ \frac{1}{k(t)}=\frac{1}{k_{pb}}+\frac{1}{k_{gb}}+\frac{1}{k_{pf}}+\frac{1}{k_{gf}}+\frac{1}{k_h}. $$
Damage is introduced as a periodic stiffness reduction:
$$ k(t)=k_n(t)+\Delta k(h_q,t), $$
where \(k_n(t)\) is the normal mesh stiffness and \(\Delta k(h_q,t)\) is the change caused by a defect of size or severity \(h_q\). In the reported simulation, slight, moderate and severe damage levels were represented by different stiffness reductions. The dynamic response was solved with a Runge–Kutta method using the parameters in Table 1.
| Parameter | Value |
|---|---|
| Number of teeth | 37 |
| Module | 2.5 mm |
| Face width | 20 mm |
| Pressure angle | 20° |
| Elastic modulus | 203 GPa |
| Gear mass | 1.133 kg |
| Rotary inertia | 2.813 × 10−3 kg·m2 |
| Radial support stiffness | 6.56 × 108 N/m |
| Radial support damping | 1.8 × 103 N·s/m |
| Mesh damping coefficient | 67 |
| Friction coefficient | 0.02 |
| Input speed | 600 rpm |
The simulated response showed the expected progression. As damage severity increased, the time waveform developed growing impact peaks. For a single cracked tooth, the modulation appeared at the 10 Hz shaft rate, producing sidebands spaced by 10 Hz around the 370 Hz mesh frequency and its higher harmonics. The model is not a high-fidelity digital twin, and it is not intended to reproduce every resonance in the gearbox. Its role is narrower: it captures the mechanisms that matter for basis selection, namely time-varying mesh stiffness, clearance, friction and a once-per-revolution stiffness loss.
Matching the model spectrum to select the wavelet basis
For each candidate mother wavelet, the measured signal is decomposed and reconstructed, and its power spectral density is compared with the model spectrum. Let \(G_M(f_n)\) be the model power spectral density and \(G(f_n)\) the measured or reconstructed density at frequency bin \(f_n\). Two error measures are used:
$$ e=\frac{1}{N}\sum_{n=1}^{N}\left|G_M(f_n)-G(f_n)\right|, $$
$$ \mathrm{RMSE}=\sqrt{\frac{1}{N}\sum_{n=1}^{N}\left(\left|G_M(f_n)-G(f_n)\right|-e\right)^2}. $$
The two terms are combined with correlation-based weights \(R_e\) and \(R_{\mathrm{RMSE}}\), giving the objective
$$ \min_{d_n^{(m)}} J(d_n^{(m)}),\qquad
J=\frac{e}{R_e}+\frac{\mathrm{RMSE}}{R_{\mathrm{RMSE}}}. $$
The search space includes Daubechies and Symlet families, such as db4, db8 and sym5. After the basis is selected, the WPD coefficients are refined against the same objective and the reconstructed signal is used for envelope analysis. The objective is deliberately not a point-by-point waveform match. A reduced-order model will contain phase and amplitude errors that carry little diagnostic information; comparing spectral envelopes instead emphasizes the sideband structure that indicates a repeating impact. The model does not replace the measured signal. It ranks representations by how well they preserve a physically plausible fault signature.
Bench validation on a cracked spiral bevel gear
The bench test used a 37-tooth spiral bevel gear with a root crack introduced by wire electrical discharge machining. The test rig consisted of a drive motor, coupling, gearbox, torque sensor, magnetic-powder brake and data-acquisition system. A three-axis accelerometer was mounted on the gearbox housing. Vibration acceleration was sampled at 10,240 Hz for 3 s, with an input speed of 600 rpm and a brake load of 4 N·m.
The damaged gear produced a larger overall vibration amplitude than the healthy gear, but amplitude alone was not a selective indicator. The amplitude spectrum showed sidebands around the higher harmonics of the mesh frequency, yet the raw envelope spectrum did not separate the shaft-rate line clearly; its amplitude was close to neighboring components. A three-level WPD produced eight bands. The seventh band had the highest energy share, but the other bands still retained significant energy. Simply selecting the highest-energy band would therefore not have been sufficient.
After physical-model-driven basis selection and coefficient refinement, the damage-related energy became concentrated in the seventh band. The reconstructed time waveform showed clear modulation pulses, and the envelope spectrum contained a 5 Hz shaft-rate component together with its third, fourth and fifth harmonics. The amplitudes at these lines exceeded the shaft-frequency amplitude in the raw envelope. This is consistent with suppression of mesh modulation and random noise, and with enhancement of the once-per-revolution impact from the damaged tooth.
Field validation on a harvester gearbox
The method was also applied to the main transmission gearbox of a Chinese onion harvester during field operation. The gearbox had a 2:1 ratio, and the driven gear was damaged. An accelerometer was mounted above the driven-gear bearing housing. Vibration acceleration was sampled at 4,096 Hz for 8 s, with a tractor driveline speed of approximately 500 rpm.
After optimized WPD processing, the sixth band had the largest energy share. Its time waveform showed pulse modulation after interference was suppressed. Envelope spectra of the fifth, sixth and seventh bands all contained shaft-frequency components, but the sixth band had the largest shaft-frequency amplitude and a clear second harmonic. This ranking agreed with the energy distribution and supported a fault call on the main transmission gearbox.
| Item | Bench spiral bevel gear | Chinese onion harvester gearbox |
|---|---|---|
| Signal | Vibration acceleration, three-axis sensor on housing | Vibration acceleration, sensor above driven-gear bearing housing |
| Sampling frequency | 10,240 Hz | 4,096 Hz |
| Record length | 3 s | 8 s |
| Shaft speed | 600 rpm input | Approximately 500 rpm tractor input |
| Load and ratio | 4 N·m brake load | 2:1 gear ratio; field load |
| Fault condition | Wire-cut root crack | Damaged driven gear |
| WPD analysis | Three levels, eight bands; seventh band dominant after optimization | Subband analysis; sixth band dominant after optimization |
| Envelope signature | 5 Hz component and third to fifth harmonics | Shaft-frequency and second-harmonic components |
What the method does and does not establish
The evidence supports a specific claim: a dynamic model can be used to constrain wavelet basis selection so that fault-related impact energy is concentrated into fewer subbands and the envelope signature becomes clearer. This was observed on a controlled bench gear with a known root crack and on a field gearbox with a damaged driven gear. In both cases, the optimized representation produced a clearer shaft-rate signature than the raw signal.
The evidence does not establish universal superiority over other diagnostic methods. The field test was a detectability check, not a controlled study of crack size, load level or speed sweep. The dynamic model requires gear geometry, mesh stiffness, load and a description of the damage mode; a new gear pair requires a new model. The basis search adds computation, and the candidate set is finite. The method should therefore be treated as a physics-guided feature-enhancement step ahead of envelope analysis or a classifier, not as a complete automatic diagnosis system. Its assumptions should be checked against the measured spectrum, especially when the gearbox contains strong bearing tones or resonances not represented in the reduced-order model.
Practical implications
For harvester drivetrain monitoring, the useful contribution is not a universal wavelet. It is a procedure for using known gear physics to constrain a signal-processing choice. Where a gear pair is critical and its geometry is known, a one-time dynamic model can generate a reference spectrum before fault examples accumulate. The WPD basis can then be selected to preserve the sideband pattern expected from a repeating tooth impact. This is particularly relevant in agricultural machinery, where fault data are scarce and operating conditions vary.
Spiral bevel gears will remain difficult to diagnose by raw vibration amplitude alone. Their contact geometry distributes energy across mesh components, and local damage produces a modulated impulse rather than a simple broadband event. A physics-guided WPD approach offers a practical way to make that impulse more visible, provided the model is treated as a prior rather than a substitute for measurement.