Source-linked AI summary
SpecRegMatch: Robust Semi-Supervised Regression for Vehicle Interior Noise Prediction
Sejin Sim, Jinsoo Bae, Seoung Bum Kim
TL;DR
Costly vehicle interior-noise measurements and the computational burden of multi-model SSR motivate a simpler approach. SpecRegMatch combines consistency regularization and information maximization in one model using spectrogram augmentations. It outperformed the compared SSR methods across labeled-data rates, including when labeled data were scarce.
Problem
Vehicle interior-noise measurements are costly, and prior SSR approaches can require multiple models or data samplings, increasing computational costs.
Source: §I · pp. 1–2Method
SpecRegMatch uses one model for consistency regularization and information maximization, applying weak, strong, and mixup augmentations to spectrograms.
Source: §I · p. 2, B. SPECTROGRAM AUGMENTATION · p. 4Results
At a 2% labeling rate, mean R2 was −0.555 for the baseline and 0.484 for SpecRegMatch, which performed best among the SSR methods across labeled-data rates.
Source: §IV · p. 8Takeaways & Limitations
The study applies spectrograms to semi-supervised regression and predicts interior noise across 11 frequency bands.
Source: §I · p. 2
Abstract
from arXiv · showhide
The rapid advancement of artificial intelligence has observed increased application in predicting vehicle interior noise levels within the automotive industry. However, the collection of labeled data for training models in this context involves significant costs. Previous studies in semi-supervised regression (SSR) have effectively mitigated the reliance on labeled data by incorporating unlabeled data. Nonetheless, these approaches often introduce a high computational cost due to the training of multiple models and data sampling. This study introduces SpecRegMatch, a novel SSR method aimed at addressing the computational cost associated with training by leveraging a single model, thus eliminating the need for multiple data samplings. SpecRegMatch integrates consistency regularization and information maximization to robustly train the model, achieved through various augmentations applied to both the embedding vectors and predicted values. Experimental results demonstrate that SpecRegMatch achieves state-of-the-art performance across various scenarios, even when using a single model. It attains a remarkable performance, as indicated by an R^2 score of 0.434. This is especially noteworthy in scenarios where labeled data is scarce. You can access the code for our proposed method at https://github.com/sejin-sim/SpecRegMatch.
I. INTRODUCTION
Vehicle interior-noise measurements are costly, while semi-supervised regression can use unlabeled data but often relies on computationally expensive multiple models or samplings. SpecRegMatch uses one model with consistency regularization and information maximization, and applies spectrograms to regression across multiple noise-frequency bands.
I. INTRODUCTION
- Vehicle interior-noise data collection requires specialized equipment and expertise for measuring acceleration and noise levels.The described process uses sensors on a BNA assembly and motor, plus a microphone near the driver’s seat.Source: §I · pp. 1–2
- Prior semi-supervised regression methods can require multiple models or repeated data samplings, creating high computational costs.The study identifies single-model SSR as a way to address this challenge.Source: §I · p. 2
- The study frames unlabeled acceleration data collected in settings similar to real environments as a more stable training option than GAN-generated samples.It links this approach to the cost of gathering labeled interior-noise and acceleration measurements.Source: §I · p. 2
- SpecRegMatch combines consistency regularization and information maximization in a single model without multiple data samplings.It applies augmentations to spectrograms and uses them for embedding vectors and predicted values.Source: §I · pp. 2–3
- SpecRegMatch predicts noise across ten 100-Hz-spaced bands from 150 Hz to 1,150 Hz and the full 0–2,000 Hz band.The inputs are spectrograms derived from six devices’ acceleration data.Source: §I · p. 2
- The study presents spectrograms as an input for semi-supervised regression, which the authors identify as a first for this setting.It also notes that prior work often focused on a single representative value for the entire frequency band, although perceived noise impact varies by band.Source: §I · pp. 1–2
A. SEMI-SUPERVISED LEARNING METHODS
Semi-supervised learning incorporates unlabeled data, and classification methods have combined augmentation, consistency regularization, and pseudo-labeling. The reviewed regression approaches often use multiple models or computationally complex methods, in contrast with SpecRegMatch’s single-model design.
A. SEMI-SUPERVISED LEARNING METHODS
- Semi-supervised learning uses unlabeled data to obtain additional information for deep-learning predictions.Hybrid approaches combine techniques such as consistency regularization, pseudo-labeling, and data augmentation.Source: A. SEMI-SUPERVISED LEARNING METHODS · p. 3
- FixMatch combines consistency regularization with pseudo-labeling, while MixMatch, ReMixMatch, FlexMatch, and SimMatch add other techniques for using unlabeled data.The section notes that MixMatch and ReMixMatch entail high computational costs from integrating various techniques.Source: A. SEMI-SUPERVISED LEARNING METHODS · pp. 3–4
- COREG and TNNR use two or more models, whereas SSDKL uses deep kernel learning and UCVME uses a Bayesian neural network and ensembles.The paper contrasts these approaches’ computational demands with its single-model method, which also avoids multiple samplings.Source: A. SEMI-SUPERVISED LEARNING METHODS · p. 4
B. SPECTROGRAM AUGMENTATION
Spectrogram augmentation generates diverse instances from unlabeled data to enhance model robustness. SpecRegMatch uses multiplicative noise, SpecAugment, and mixup, and applies consistency regularization between predictions from mixup- and weakly augmented spectrograms.
B. SPECTROGRAM AUGMENTATION
- Spectrogram augmentation generates diverse instances from unlabeled data and is used to enhance model robustness.The paper situates its approach alongside prior spectrogram augmentation work in audio classification.Source: B. SPECTROGRAM AUGMENTATION · p. 4
- SpecRegMatch uses multiplicative noise for weak augmentation, SpecAugment for strong augmentation, and mixup to combine the two.Its consistency regularization minimizes the disparity between predictions from mixup-augmented and weakly augmented spectrograms.Source: B. SPECTROGRAM AUGMENTATION · p. 4
III. PROPOSED METHOD
SpecRegMatch transforms labeled and unlabeled signal data into spectrograms, then augments unlabeled spectrograms in three ways. A single model produces predictions and embeddings used to calculate supervised and unsupervised losses.
III. PROPOSED METHOD
- The method applies STFT to labeled and unlabeled signal data to obtain their spectrograms.The STFT divides a signal into short overlapping windows and computes a Fourier transform for each window.Source: §III · p. 5
- Unlabeled spectrograms receive weak multiplicative-noise augmentation, strong time- and frequency-masking augmentation, and mixup between the two.The mixup ratio α controls their linear combination and is sampled between zero and one using a beta distribution.Source: §III · p. 5
- The model forwards labeled and augmented unlabeled spectrograms through a backbone and MLP, then computes supervised and unsupervised losses.The unsupervised loss combines consistency-regularization and information-maximization losses; the total loss adds this to the labeled-data loss.Source: §III · p. 6
B. SPECREGMATCH
SpecRegMatch predicts 11 vehicle interior-noise frequency bands from labeled and unlabeled spectrograms, combining supervised training with consistency regularization and information maximization. Its unlabeled-data losses use augmented views to align predictions and embedding features.
B. SPECREGMATCH
- SpecRegMatch predicts vehicle interior noise across 11 distinct frequency bands.The method transforms labeled and unlabeled signal batches into spectrograms using STFT.Source: B. SPECREGMATCH · p. 6
- The model extracts embeddings from augmented spectrograms and sends them through an MLP to produce predictions, with separate losses on embeddings and predictions.Information maximization uses embeddings, while consistency regularization uses MLP predictions.Source: B. SPECREGMATCH · p. 6
- Consistency regularization trains strong- and mixup-view predictions to align with the weak-view prediction used as a pseudo-label.The consistency loss uses separate weights for strong-view and mixup-view alignment.Source: B. SPECREGMATCH · p. 6
- Information maximization combines invariance and redundancy-reduction losses based on a cross-correlation matrix between augmented-sample embeddings.The loss uses hyperparameters to control the overall information-maximization influence and redundancy.Source: B. SPECREGMATCH · p. 6
- The cross-correlation matrix targets an identity matrix, with diagonal elements capturing invariance and off-diagonal elements reducing redundancy among embedding features.The matrix compares embeddings from weakly and strongly augmented spectrograms.Source: B. SPECREGMATCH · p. 7
IV. EXPERIMENTS
Across five labeled-data rates, SpecRegMatch achieves the best mean R2 among the SSR methods, including when labels are scarce. It also reports the lowest measured FLOPs among SSR methods, while ablations show component contributions vary with the labeling rate.
IV. EXPERIMENTS
- The experiments use five labeled-data proportions—2%, 7%, 28%, 52%, and 100%—and evaluate MAE, R2, and one-epoch FLOPs.The experimental dataset comes from an electric vehicle; its acceleration spectrogram inputs are formed from six devices, and the model predicts 11 noise targets.Source: §IV · pp. 7–8
- SpecRegMatch achieved R2 0.458 versus second-best TNNR and the most individual best target R2 results: 5/11 at 2%, 6 at 7%, 9 at 28%, 10 at 52%, and 9 at 100% labeling.It also achieved superior results using 28% of training labels compared with fully supervised results at 100%.Source: §IV · p. 8
- At the 2% labeled-data rate (K), SpecRegMatch achieved mean R2 of 0.484 versus −0.555 for the supervised baseline, and was best among the five SSR methods across all labeled-data rates.Source: §IV · p. 8
- SpecRegMatch records 26.36 × 10^13 FLOPs, the lowest among SSR methods; UCVME records the highest, 109.19 × 10^13 FLOPs.The paper notes that SSR methods share the same stated big-O complexity except UCVME, while measured FLOPs distinguish their real-computer costs.Source: §IV · pp. 8, 10, Figure 4 · p. 10
- Ablations indicate that information maximization and mixup-augmented consistency regularization contribute at 2%, 7%, and 28% labeled data, while consistency regularization remains significant at 52% and 100%.The contribution of the other components is lower at the two higher labeling rates.Source: §IV · p. 10
V. CONCLUSION
SpecRegMatch uses one model for semi-supervised vehicle-noise prediction and reports state-of-the-art results; its frequency predictions remain less granular than industry practice. The paper identifies scaling multi-output training and handling label-distribution shifts as future work.
V. CONCLUSION
- SpecRegMatch achieves state-of-the-art results against other semi-supervised regression methods, especially with limited labeled data.Ablations identify consistency regularization, information maximization, and mixup augmentation as contributors to improved performance.Source: §V · p. 10
- The study predicts noise at 100 Hz intervals, while a Korean automotive vendor's noise-management program operates at 10 Hz intervals.The authors contrast their finer-grained frequency predictions with prior work focused predominantly on overall interior noise.Source: §V · p. 10
- Table 5 reports MAE means and standard deviations across five labeled-data rates, with Wilcoxon tests comparing the best and second-best mean results.Source: Table 5 · p. 11
- Future work will investigate improving multi-output performance as the number of predicted frequency bands increases and selecting reliable unlabeled predictions under distribution shifts.The authors note that shifts between labeled and unlabeled data in real-world scenarios may lead to unreliable unlabeled predictions.Source: §V · p. 12
APPENDIX EXPERIMENT RESULTS OF MAE
SpecRegMatch outperforms the other semi-supervised regression methods on mean MAE across 11 labeled-data targets at five labeling rates, and records the most per-band top performances.
APPENDIX EXPERIMENT RESULTS OF MAE
- 5 of 11 targets had the best MAE at 2% labeling, 5 at 7%, 8 at 28%, 8 at 52%, and 7 at 100%.SpecRegMatch also outperformed the other methods on mean MAE across the 11 labeled-data targets at all five rates.Source: APPENDIX EXPERIMENT RESULTS OF MAE · p. 12