Gear Power Honing Monitoring

• Gear power honing is a hard-finishing process that uses controlled cross-axis infeed and axial oscillation to refine gear micro-geometry and surface quality for enhanced NVH performance.
• In-process monitoring employs dual MEMS accelerometers and STFT-based time-frequency analysis combined with machine learning techniques like PCA, PCA-LDA, and R-UMLDA to classify gear quality.
• The method achieves up to 100% test accuracy by linking transient vibration features to specific defect classes, enabling real-time quality control and process intervention.

Searching arXiv for the specified paper and closely related work on gear power honing monitoring. Gear power honing is a hard-finishing operation performed near the end of the gear manufacturing chain to achieve precise micro-geometry and surface quality on tooth flanks with the explicit goal of improving NVH (Noise, Vibration, Harshness) performance of assembled transmissions. It uses an abrasive honing stone/tool that engages the rotating gear under a controlled cross-axis contact: the workpiece and tool rotate synchronously, their axes are skewed by a cross-axis angle to generate the required relative motion, and the process applies a radial infeed with a superimposed axial oscillation. In the study "In-Process Monitoring of Gear Power Honing Using Vibration Signal Analysis and Machine Learning" (Capurso et al., 24 Sep 2025), gear power honing is examined not only as a finishing process but also as a monitored dynamical system whose transient vibrations can be linked to final gear quality categories, geometrical inspections, and end-of-line NVH behavior.

1. Process definition and manufacturing role

Gear power honing is positioned after hard-turning or grinding for shape and is used for micro-geometry corrections (profile and lead), ripple/waviness mitigation, and surface finishing (Capurso et al., 24 Sep 2025). Compared with continuous generative grinding or skiving, power honing has undefined cutting edges and generates tooth flank surfaces without parallel exciting structures, which is known to reduce excitation of gear mesh harmonics and lower perceived noise. The process is therefore directly tied to the NVH requirements of assembled transmissions rather than only to geometric compliance.

The honing cycle comprises three phases: roughing, finishing, and spark-out. Roughing uses higher radial infeed to remove allowance; finishing uses reduced infeed to hit dimensional targets and surface integrity; spark-out uses no infeed while improving surface finish. The study explicitly analyzes finishing and spark-out phases because they dominate final surface quality; the initial roughing segment is removed to avoid bias from input-part variability. This phase selection is consequential because it narrows the monitoring problem to the cycle segment most strongly connected to NVH-relevant surface generation.

A common misconception is that post-process inspection alone is sufficient for quality assurance in such finishing operations. The reported industrial motivation is the opposite: conventional post-process inspection and Statistical Process Control (SPC) cannot capture transient anomalies that occur during continuous operation, and end-of-line tests on assembled gearboxes do not provide immediate causality or real-time intervention (Capurso et al., 24 Sep 2025). In this setting, gear power honing is best understood as a process in which quality emerges from a time-varying interaction among tool, workpiece, machine structure, and fixturing.

2. Kinematics, excitation mechanisms, and NVH relevance

The honing head, workpiece spindle, tailstock, and abrasive tool are arranged so that tool and workpiece rotate synchronously under a cross-axis angle with axial oscillation superimposed on radial infeed (Capurso et al., 24 Sep 2025). The kinematic coupling of synchronous rotation, skewed axes, infeed, and oscillation creates the relative motion required for finishing, but it also creates a pathway by which dynamic instability can imprint quality-relevant patterns on tooth flanks.

The study states that NVH problems are often tied to dynamic instability, including chatter, self-excited vibrations, and variable contact conditions in honing. These can imprint periodic waviness, described as “ghost components,” on tooth flanks, increasing transmission error and producing spectral peaks at characteristic frequencies. Relevant frequencies include the workpiece rotational frequency,

$f_{\mathrm{rot}}=\frac{n}{60},$

with $n$ the workpiece rpm, and the gear mesh frequency in operation,

$f_m=z\cdot \frac{n}{60},$

with $z$ the tooth count; surface undulations and micro-geometry deviations often produce sidebands at $f_m \pm k f_{\mathrm{rot}}$ with $k \in \mathbb{Z}$.

The reported observations corroborate that self-excited vibrations triggered during finishing can intensify or decay and that their quality impact is captured more effectively by tracking amplitude trends at individual spectral components than by global maxima (Capurso et al., 24 Sep 2025). This is an important distinction. Fixed amplitude thresholds on acceleration or power signals are described as brittle because of fluctuations in tool wear, dressing, input gear quality, temperature, and feed rates. A plausible implication is that in gear power honing the diagnostically relevant object is not a single scalar alarm variable but a structured time–frequency pattern whose meaning depends on the process phase and the physical origin of excitation.

3. In-process monitoring architecture and signal representation

The proposed monitoring framework is data-driven and uses continuous data acquisition via accelerometers, followed by time-frequency signal analysis and machine learning (Capurso et al., 24 Sep 2025). Two MEMS accelerometers (unidirectional) were mounted on the spindle and on the tailstock. The dual placement increases robustness to localized noise and structural resonances. Sampling rate was $f_s = 25$ kHz. The short-time Fourier transform used a Hamming window of length $N = 512$, yielding frequency-bin resolution

$$\Delta f \approx \frac{25{,}000}{512} \approx 48.83\ \text{Hz}.$$

The spectrogram has $b = 512$ frequency bins and $n$0 time frames per observation, and the total observation window covers approximately the last $n$1 of the cycle, with roughing removed by approximately $n$2.

The continuous acceleration $n$3 is converted into a time–frequency representation using the STFT:

$n$4

with a Hamming window $n$5 of length $n$6. The spectrogram is

$n$7

Each spectrogram is normalized to grayscale $n$8 to reduce sensitivity to absolute amplitude changes from tool wear and dressing state. This normalization is not presented as a cosmetic image-processing step but as a mechanism to suppress nuisance variability in absolute amplitude.

Sensor fusion is performed by the Hadamard product,

$n$9

where $f_m=z\cdot \frac{n}{60},$0 and $f_m=z\cdot \frac{n}{60},$1 are normalized spectrograms from spindle and tailstock, and $f_m=z\cdot \frac{n}{60},$2 denotes the Hadamard product (Capurso et al., 24 Sep 2025). The merged spectrogram is then rescaled to enhance pattern visibility. The stated purpose is to emphasize features common to both sensor channels and suppress sensor-specific noise. Each observation is represented either as a single merged $f_m=z\cdot \frac{n}{60},$3D map of size $f_m=z\cdot \frac{n}{60},$4 or as a third-order tensor of size $f_m=z\cdot \frac{n}{60},$5 when both spectrograms are retained.

The control unit performs STFT in-line; an on-board PC collects machine parameters, timestamps, and spectrograms and uploads them to a dedicated cloud system. This makes the monitoring pipeline simultaneously a signal-processing chain and an industrial data architecture, with sensing, feature generation, classification, storage, and later retraining integrated into one workflow.

4. Quality classes, labeling, and physical interpretation of defects

The dataset comprises 429 components labeled into four classes, with a random split into 80% training and 20% test and 5-fold cross-validation against class imbalance, particularly because NOK3 has fewer samples (Capurso et al., 24 Sep 2025). Labels were established by combining geometrical inspections, including micro-geometry and waviness per VDI 2612, with end-of-line test bench NVH of assembled gearboxes. The workpieces were case-hardened pinion gears for dual-clutch transmissions monitored in production.

The four quality classes are defined as follows.

Class	Definition
OK	Compliant micro-geometry and NVH
NOK 1	Non-compliant NVH due to surface waviness on gear flanks
NOK 2	Micro-geometry deviations on flanks but compliant NVH
NOK 3	Non-compliant NVH attributed to periodic pitch patterns on gear teeth

This labeling scheme is notable because it is not based on vibration signatures alone. It explicitly couples process signals to downstream quality evidence from both geometry and gearbox NVH. As a result, the learning problem is not merely anomaly detection but multiclass discrimination among physically distinct and production-relevant defect modes.

The interpretability analysis in the study links these classes to time–frequency behavior. Healthy honing (OK) corresponds to stable energy distribution with subdued harmonics and consistent patterns across spindle and tailstock merged maps. Surface waviness (NOK1) is associated with enhanced energy at specific bands, often aligned with multiples of $f_m=z\cdot \frac{n}{60},$6 and structural resonances of the clamping system, producing spectral “ghost components.” Micro-geometry deviations (NOK2) show distinct time–frequency behavior from NOK1 and often manifest as changes in energy distribution during finishing or spark-out without pronounced harmonic sidebands. Periodic pitch patterns (NOK3) are described as subtle yet consistent features, often sideband-like structures and narrowband peaks (Capurso et al., 24 Sep 2025). This suggests that physically different mechanisms can produce separable signatures even when one class remains NVH-compliant, as in NOK2.

5. Subspace learning, tensor methods, and multiclass SVM inference

Three subspace-learning methods were compared for feature extraction: Principal Component Analysis (PCA), a two-stage PCA + Linear Discriminant Analysis (LDA) framework, and Uncorrelated Multilinear Discriminant Analysis with Regularization (R-UMLDA) (Capurso et al., 24 Sep 2025). All extracted features are then fed into a Support Vector Machine (SVM) classifier to predict the four distinct gear quality categories.

For PCA, let $f_m=z\cdot \frac{n}{60},$7 be $f_m=z\cdot \frac{n}{60},$8-dimensional flattened maps, with mean $f_m=z\cdot \frac{n}{60},$9 and total scatter matrix

$z$0

The objective is

$z$1

which yields eigenvectors $z$2 of $z$3. The projection of a centered sample $z$4 is

$z$5

In the implementation, maps are flattened to $z$6, and the optimal feature count was $z$7 by 5-fold cross-validation.

For PCA + LDA, the between-class and within-class scatter matrices are

$z$8

The Fisher criterion is

$z$9

equivalently solving

$f_m \pm k f_{\mathrm{rot}}$0

To avoid $f_m \pm k f_{\mathrm{rot}}$1 singularity, the method first applies PCA to reduce $f_m \pm k f_{\mathrm{rot}}$2 and then LDA to project $f_m \pm k f_{\mathrm{rot}}$3. In the reported implementation, PCA reduces to $f_m \pm k f_{\mathrm{rot}}$4 and LDA then reduces to $f_m \pm k f_{\mathrm{rot}}$5.

R-UMLDA operates directly on high-order tensors, performing mode-wise projections to preserve structure. Alternating projections solve generalized eigenproblems along each mode $f_m \pm k f_{\mathrm{rot}}$6,

$f_m \pm k f_{\mathrm{rot}}$7

while enforcing orthonormality or uncorrelated features across modes, for example $f_m \pm k f_{\mathrm{rot}}$8. Regularization mitigates small-sample singularity through $f_m \pm k f_{\mathrm{rot}}$9 with $k \in \mathbb{Z}$0. In the study, the regularization parameter was tuned in $k \in \mathbb{Z}$1, with optimal $k \in \mathbb{Z}$2. Feature tensor projection is expressed as

$k \in \mathbb{Z}$3

R-UMLDA was tested on merged 2D maps and on raw 3D tensors, with similar performance; results are reported for merged maps, with optimal feature count $k \in \mathbb{Z}$4.

Classification uses ECOC (Error-Correcting Output Codes) with binary SVM base learners. The SVM decision function is

$k \in \mathbb{Z}$5

with Gaussian/RBF kernel

$k \in \mathbb{Z}$6

Hyperparameters $k \in \mathbb{Z}$7 and $k \in \mathbb{Z}$8 were tuned via 5-fold cross-validation on training data, although numerical values were not disclosed (Capurso et al., 24 Sep 2025). The dimensionality of the SVM input equals the number of selected features: $k \in \mathbb{Z}$9 for PCA, $f_s = 25$0 for PCA + LDA, and $f_s = 25$1 for R-UMLDA.

6. Reported performance, deployment logic, and limitations

The reported test accuracy is 97.65% for PCA, 98.83% for PCA + LDA, and 100.00% for R-UMLDA; the comparative summary lists PCA at 97.65%, PCA–LDA at 98.82%, and R-UMLDA at 100.00% (Capurso et al., 24 Sep 2025). Class-wise F1 scores show that the main weakness of PCA is the minority and subtle defect class NOK3, where F1 is 85.71%, whereas PCA + LDA and R-UMLDA both reach 100.00% F1 for NOK3 under the reported test split.

Method	Feature dimension	Test accuracy
PCA	$f_s = 25$2	97.65%
PCA + LDA	$f_s = 25$3	98.83%
R-UMLDA	$f_s = 25$4, $f_s = 25$5	100.00%

The paper also reports physically interpretable feature structure. For PCA, PC1 captures the “healthy honing” state, separating OK from NOK; PC2 distinguishes NVH waviness (NOK1) from micro-geometry deviations (NOK2); PC5 isolates periodic pitch (NOK3). For PCA + LDA, DC1 separates OK versus NOK, DC2 distinguishes NOK1 versus NOK2, and DC3 captures NOK3. For R-UMLDA, Feature 1 separates OK, Feature 2 distinguishes waviness (NOK1) versus micro-geometry deviations (NOK2), and Feature 3 isolates periodic pitch (NOK3). The reported interpretability is therefore not incidental: the low-dimensional features align with defect semantics rather than only with abstract discriminant directions.

The real-time integration architecture consists of dual MEMS accelerometers feeding the control unit, STFT generation with a 512-point Hamming window at $f_s = 25$6 kHz, Hadamard merging of spindle and tailstock spectrograms, projection into low-dimensional subspaces, and ECOC-SVM classification. Alarms are raised for NOK categories, and results are stored alongside process data in the cloud. The pipeline uses small feature counts, 3 to 5, and fixed-size STFT windows, supporting real-time deployment on standard industrial PCs. Model updates follow periodic retraining using 5-fold cross-validation on recent labeled samples; thresholds can trigger SPC interventions or predictive maintenance actions such as tool dressing or parameter adjustment (Capurso et al., 24 Sep 2025).

Deployment guidance in the study recommends mounting one accelerometer on the spindle and another on the tailstock, using rigid mounting and consistent orientation; using $f_s = 25$7 kHz and STFT with $f_s = 25$8; removing roughing if input variability dominates; normalizing spectrograms to $f_s = 25$9; merging channels via Hadamard product; and favoring supervised subspace methods, specifically PCA–LDA or R-UMLDA, with 3 features for performance and efficiency. The implementation outline further ties specific defect classes to action rules: NOK1 triggers immediate analysis of process parameters and dressing; NOK2 triggers upstream dimensional checks; NOK3 prompts clamping or centering and kinematics inspection.

The limitations are explicit. Validation was performed on case-hardened pinions under one honing setup. Changes in module, tooth count, stone characteristics, cross-axis angle, feed, and coolant could shift spectral patterns. Sensor noise, structural resonances, and external disturbances can still affect spectra despite dual sensing and merging. Small-sample issues are addressed by the PCA–LDA two-stage approach and by R-UMLDA regularization with $N = 512$0. A plausible implication is that the perfect test accuracy reported for R-UMLDA should be interpreted as a result obtained under the studied conditions rather than as a process-invariant guarantee. The future directions stated in the paper are to enrich sensing with honing wheel signals, electrical consumption, temperature, and acoustic emissions for sensor fusion, and to extend validation across gear families, stone specifications, and kinematic variations in order to quantify domain transfer and robustness (Capurso et al., 24 Sep 2025).