TLDR: This research introduces a novel, domain-knowledge-informed method for selecting optimal anomaly detection models for electric vehicle interior sounds. Addressing the common issue of scarce labeled faulty data, the approach generates ‘proxy-anomalies’ by applying structured perturbations to healthy spectrograms. These synthetic faults, mimicking real-world conditions, are used in the validation process to guide model selection. The study also presents a new, publicly available dataset of healthy and five types of faulty EV cabin sounds. Experimental results show that models selected using this proxy-anomaly strategy significantly outperform conventional methods, achieving performance comparable to scenarios with ideal access to real fault data, thus enabling robust anomaly detection in data-limited environments.
Ensuring the highest quality and comfort in electric vehicles (EVs) is paramount, and a crucial aspect of this is maintaining impeccable cabin sound quality. Even minor acoustic imperfections can significantly detract from the driving experience, making the early and accurate detection of sound anomalies a top priority for automotive manufacturers.
Traditionally, identifying these faults has relied on manual inspections or mechanical diagnostics. However, recent advancements have paved the way for sound-based and data-driven methods, offering more objective, scalable, and non-invasive alternatives. The challenge, however, lies in the nature of faults themselves: they are rare. This scarcity of labeled faulty data makes traditional supervised learning approaches, which require examples of both healthy and faulty conditions, impractical.
This is where unsupervised learning comes into play. In an unsupervised setting, a model is trained exclusively on healthy sound samples and then identifies anomalies as any significant deviations from this learned ‘normal’ behavior. While powerful, this approach faces a fundamental hurdle: how do you select the best model or fine-tune its parameters without any labeled faulty examples for validation?
A recent research paper, titled A Domain Knowledge Informed Approach for Anomaly Detection of Electric Vehicle Interior Sounds, proposes an innovative solution to this problem. Authored by Deepti Kunte, Bram Cornelis, Claudio Colangeli, Karl Janssens, Brecht Van Baelen, and Konstantinos Gryllias, the study introduces a domain-knowledge-informed approach for model selection.
Engineering Proxy-Anomalies for Better Model Selection
The core of the proposed methodology involves creating ‘proxy-anomalies.’ These are not real faults but rather engineered, structured perturbations applied to healthy spectrograms (visual representations of sound) in the validation set. These engineered anomalies are designed to mimic the diverse alterations observed in real fault scenarios, leveraging deep domain knowledge of automotive acoustics.
The researchers developed three types of structured perturbations: adding RPM-lines, frequency-lines, and order-lines to spectrograms. Each perturbation targets specific audio characteristics relevant to vehicle cabin noise, such as impacts, tonal noise, and order fluctuations. This technique is both fast and computationally efficient, allowing for the rapid generation of tailored anomalies that support effective model selection.
A New Public Dataset for EV Cabin Sounds
To evaluate their methodology, the team curated a comprehensive, high-fidelity electric vehicle cabin noise dataset. This dataset includes both healthy and faulty sound samples across five distinct fault types: Imbalance, Modulation, Whine, Wind, and Pulse Width Modulation (PWM). These sounds were generated using advanced sound synthesis techniques based on a sound quality equivalent model, built on real experimental recordings. The realism and appropriate fault levels of the dataset were validated through expert jury assessments, and importantly, this dataset has been made publicly available to foster further research in the field.
Outperforming Conventional Methods
The experimental evaluations demonstrated that models selected using these proxy-anomalies significantly outperformed conventional model selection strategies, which typically rely solely on reconstruction error from healthy validation data. The proposed approach achieved performance trends closely aligned with an ‘ideal’ scenario where labeled faulty data would be available for model selection, proving its robustness and reliability in data-scarce environments.
While the method showed strong performance across most fault types, the ‘Modulation’ fault proved more challenging to detect. This was attributed to the small spectrogram region it affects and its audibility at extremely low amplitudes, which can be masked by the inherent randomization in healthy orders.
Also Read:
- Advanced Sound Analysis: A New Method for Detecting Machine Anomalies Without Explicit Labels
- Enhancing Bearing Fault Detection with Limited Data: A New AI Framework
Visualizing Model Performance
To provide deeper insights, the researchers utilized various visualization techniques. Boxplots of reconstruction errors clearly showed the model’s ability to distinguish healthy from faulty samples. t-SNE visualizations of the latent space revealed how the model internally separates different fault classes, with distinct clustering for PWM and Wind noise, and partial overlap for the more challenging Modulation fault. Pixel-level reconstruction error maps and saliency maps further highlighted the specific spectral regions where the model struggled or focused its anomaly detection decisions, confirming its capacity to localize anomalies effectively for most fault types.
In conclusion, this study offers a scalable and generalizable framework for anomaly detection in automotive interior sounds. By leveraging domain-informed proxy-anomalies, it effectively addresses the critical challenge of limited faulty data during model development, paving the way for more robust and data-efficient fault detection in real-world applications.
