The challenge they set out to solve is one that plagues machine learning across manufacturing: class imbalance. In a typical transmission plant, the overwhelming majority of units are good, while defective transmissions are rare. A classifier trained on such data tends to favor the majority class, quietly waving defective units through the inspection process. Worse still, real production data contains outliers, mislabeled samples, and measurement noise from nearby machinery and inconsistent sensors, unlike the clean, balanced laboratory datasets that dominate the academic literature on fault diagnosis. Existing over-sampling techniques such as SMOTE and its many variants attempt to balance the classes by generating synthetic minority samples, but they are notoriously fragile in the presence of noise, often amplifying the very errors they should ignore.

Writing in the journal Knowledge and Information Systems, Emeka Ndupuechi and Christian Beecks introduce a new over-sampling framework called Modified Z-Score Cyclic Exponential Averaging of Minority Data, or mZ-cEAMD. The method operates in two stages. First, it cleans the dataset using the Modified Z-Score technique, a robust outlier detection approach based on the median and the Median Absolute Deviation rather than the mean and standard deviation. Because medians resist being dragged around by extreme values, the method can flag suspicious samples without the detection process itself being corrupted by them. Class-specific thresholds, tuned through an ablation study using master reference transmissions ranging from worst-of-worst to best-of-best units, determine which samples are removed before any synthetic data is created.

The second stage is where the method departs most sharply from its predecessors. Rather than relying on the fixed linear interpolation used by SMOTE, mZ-cEAMD generates synthetic minority samples through cycles of two-point exponential averaging between consecutive minority class feature vectors. Adjustable time constants control the interpolation weights: larger time constants produce synthetic samples that hug the original data closely, preserving precision, while smaller time constants spread samples across a broader range, boosting diversity. By cycling through a series of time constants, the algorithm builds up a balanced minority class whose synthetic members stay within the natural boundaries of the real data, minimizing overlap with the majority class and reducing the boundary ambiguity that degrades classifier performance.

The researchers paired their over-sampling method with a carefully engineered feature extraction pipeline tailored to transmission acoustics. From the raw vibration signal captured by an axial accelerometer and the angular position recorded by a 2048-pulse-per-revolution encoder, they extracted two complementary feature types. Spectral order tracking resamples the vibration data into the rotation domain, so that gear meshing frequencies remain fixed in the order spectrum regardless of shaft speed. The amplitude of the fundamental gear meshing order, tracked across every revolution and averaged in decibels, serves as a sensitive indicator of gear wear, pitting, scuffing, and misalignment. In parallel, the Short-Time Fourier Transform captures the time-frequency behavior of the signal, revealing broadband, non-synchronous defect signatures such as nicks or dents that order-based analysis would miss.

The study focused on the sixth gear of two seven-speed transmission variants produced on a real line: a hybrid dual-clutch transmission with an added e-Motor, and a conventional dual-clutch transmission. Each unit was tested at the end-of-line station by ramping the input shaft from 1000 to 4000 rpm and back under controlled loads, while the data acquisition system sampled vibration and encoder signals at 50 kilohertz for 4.5 seconds. Combining the order-track and STFT features produced two imbalanced datasets drawn directly from production: one with 3,947 transmissions, 113 attributes, and an imbalance ratio of 22.5, and another with 5,038 transmissions, 15 attributes, and an imbalance ratio of 20. Both contained mislabeled samples arising from human error in the predefined NVH thresholds, and the researchers deliberately injected additional label noise to stress-test their method.

For classification, the team selected a compact multilayer perceptron after an ablation study showed that deeper and wider networks offered only marginal gains while inflating training time and over-fitting risk, a crucial consideration on a production line where the interval between tests is measured in seconds. They benchmarked mZ-cEAMD against three established techniques, SMOTE, Borderline-SMOTE, and SVM-SMOTE, using the Matthews Correlation Coefficient and specificity as primary metrics, since accuracy is misleading under severe imbalance and minimizing false positives, meaning defective units wrongly passed as good, is the paramount operational concern.

The results were decisive. On the first dataset, mZ-cEAMD achieved a best MCC of 0.90 and a best specificity of 0.92, far ahead of SMOTE at 0.64, Borderline-SMOTE at 0.59, and SVM-SMOTE at 0.49. On the second, noisier dataset, the proposed method again led with an MCC of 0.77 and specificity of 0.86, compared with 0.58 and 0.83 for Borderline-SMOTE, 0.43 and 0.74 for SMOTE, and 0.47 and 0.69 for SVM-SMOTE. Confusion matrices confirmed that the new method produced the fewest defective transmissions misclassified as good and the fewest good units wrongly rejected, on both datasets.

The authors attribute this consistent advantage to the framework’s integrated noise detection, which prevents outliers and mislabeled samples from being amplified during over-sampling, together with the adaptive structure of exponential averaging, which preserves the intrinsic distribution of the minority class. Because the method was validated on genuinely imbalanced, noisy production data rather than curated laboratory datasets, the researchers argue it offers a realistic pathway toward machine learning based NVH testing that could complement or eventually replace manual threshold evaluation. Future work will focus on automating the selection of time constants, validating the framework across other domains and under injected noise scenarios such as Gaussian feature noise, and exploring the synchrosqueezing transform as an alternative to the STFT for feature extraction. If those efforts succeed, the humble end-of-line test stand could become one of the clearest examples of artificial intelligence quietly transforming quality assurance in automotive manufacturing.

Subject of Research: Noise-robust over-sampling for imbalanced transmission production NVH data classification

Article Title: A noise-robust over-sampling method for imbalanced transmission production data classification

Article References: Ndupuechi, E., & Beecks, C. (2026). A noise-robust over-sampling method for imbalanced transmission production data classification. Knowledge and Information Systems, 68(1), Article 257. https://doi.org/10.1007/s10115-026-02852-9

Image Credits: AI Generated

DOI: 10.1007/s10115-026-02852-9

Keywords: machine learning, over-sampling, class imbalance, NVH testing, transmission manufacturing, SMOTE, outlier detection, spectral order tracking, Short-Time Fourier Transform, automotive quality assurance, fault detection, mZ-cEAMD