Planing It by Ear: Convolutional Neural Networks for Acoustic Anomaly Detection in Industrial Wood Planers

Anthony Deschênes, Rémi Georges, Cem Subakan, Bruna Ugulino, Antoine Henry, Michael MorinICASSP 2025

TL;DR

This paper introduces a novel approach to detecting anomalies in industrial wood planers using deep convolutional neural networks. By developing a new dataset of real-world planing sounds and proposing advanced architectures like Skip-CAE and Skip-CAE-Transformer, the authors achieve state-of-the-art performance in identifying subtle machine malfunctions, paving the way for more reliable and automated industrial monitoring.

What Problem Does This Paper Address?

The wood product industry faces significant challenges due to a shortage of skilled labor, leading to more frequent equipment failures and increased operational costs. Industrial environments like sawmills are inherently noisy, making it difficult to detect subtle anomalies in machinery. Experienced operators can often diagnose issues by ear, but this human expertise is becoming scarce. The paper aims to bridge this gap by developing an automated acoustic monitoring system that can detect anomalies in wood planers, assisting both novice operators and improving overall machinery reliability.

Methods

The core of this research lies in developing and evaluating deep neural network architectures for acoustic anomaly detection. The authors propose two main models:

  • Skip-CAE (Convolutional Autoencoder with Skip Connections): This architecture builds upon standard convolutional autoencoders by incorporating skip connections. These connections allow the network to pass information directly from earlier layers to later layers, helping to preserve fine-grained details and improve the learning of complex patterns, which is crucial for detecting subtle anomalies. They also utilize leaky ReLU activation functions to prevent vanishing gradients and batch normalization for better convergence.
  • Skip-CAE-Transformer: This model further enhances the Skip-CAE by integrating Transformer encoder and decoder blocks. Transformers are known for their ability to capture long-range dependencies in sequential data. By combining convolutional layers with transformers, this architecture aims to leverage both local feature extraction and global context understanding for more robust anomaly detection.

Both models are trained as autoencoders, meaning they learn to reconstruct normal operating sounds. Anomalies are then detected as deviations from this learned reconstruction.

Datasets

The paper introduces a new, publicly available industrial acoustic dataset specifically for wood planers. This dataset comprises:

  • Source: Recordings from a real industrial planer (Gilbert High Speed Planer).
  • Format: 7,562 10-second single-channel recordings sampled at 20 kHz, totaling approximately 21 hours.
  • Content: Includes normal planing sounds, machine start/shutdown, and importantly, 105 labeled anomalies. These anomalies are categorized as broken boards, stuck boards, and uneven or thick boards.
  • Preprocessing: Audio signals are converted into mel spectrograms with a frame size of 50 ms and a hop length of 25 ms, resulting in 80 mel bins per spectrogram.

This dataset is split into a training set (containing only normal sounds) and an evaluation set (containing both normal and anomalous sounds).

Key Insights and Ideas

  • The Power of Skip Connections: The Skip-CAE architecture demonstrates that simply adding skip connections to a convolutional autoencoder significantly improves its ability to learn and reconstruct normal sounds, leading to better anomaly detection. This suggests that preserving spatial information across layers is vital for this task.
  • Transformers Enhance Contextual Understanding: The integration of transformer blocks in the Skip-CAE-Transformer model further boosts performance. This highlights the benefit of capturing longer-range dependencies and contextual information within the acoustic signals, which is likely important for distinguishing subtle anomalies from normal variations.
  • Real-World Data is Crucial: The creation and use of a new, real-life industrial dataset are critical. It allows for a more realistic evaluation of anomaly detection models compared to synthetic or laboratory-generated data, which may not capture the full complexity of noisy factory environments.
  • Anomaly Detection as Reconstruction Error: The paper reinforces the effectiveness of using reconstruction error from autoencoders as a proxy for anomaly detection. Deviations from the learned "normal" sound patterns result in higher reconstruction errors, signaling an anomaly.

Key Results

The proposed models were compared against several baselines, including a DCASE baseline autoencoder, a one-class SVM, an isolation forest, and a previously published convolutional autoencoder.

  • Overall Performance: The Skip-CAE-Transformer achieved the highest Area Under the ROC Curve (AUC) of 0.875 and a partial AUC (pAUC) of 0.785 on the evaluation set. The Skip-CAE model also performed strongly, with an AUC of 0.846 and pAUC of 0.787.
  • Outperforming Baselines: Both Skip-CAE and Skip-CAE-Transformer significantly outperformed the DCASE baseline (AUC 0.520), Isolation Forest (AUC 0.518), and One-Class SVM (AUC 0.683).
  • Specific Anomaly Detection: The Skip-CAE-Transformer showed particular strength in detecting "stuck boards" and "uneven or thick wood" anomalies, outperforming other models in AUC and pAUC for these specific types.
  • Qualitative Improvement: The reconstructed mel spectrograms from the proposed models showed a clearer representation of the original sounds, indicating a better understanding of normal operating patterns compared to baseline models.

Limitations and Future Work

  • Dataset Limitations: While the dataset is valuable, the authors note that "some anomalies might not be labeled in the evaluation set" due to the nature of real-life operations. This could potentially impact the evaluation of models on unseen anomalies.
  • Generalizability: The models were trained and evaluated on data from a single type of industrial wood planer. Future work could explore the generalizability of these models to other types of wood processing machinery or different industrial environments.
  • Operator Trust: The authors suggest that acoustic anomaly detection is close to how expert operators function and has the potential to gain their trust. Further research could focus on developing explainable AI methods to make the anomaly detection system more transparent and interpretable for human operators.
  • Real-time Detection: While not explicitly detailed, extending these models for real-time anomaly detection would be a logical next step for practical industrial deployment.

Summary updated

BibTeX
@inproceedings{deschenes2025planing,
  author = {Anthony Deschênes and Rémi Georges and Cem Subakan and Bruna Ugulino and Antoine Henry and Michael Morin},
  title = {{Planing It by Ear: Convolutional Neural Networks for Acoustic Anomaly Detection in Industrial Wood Planers}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2025},
}
Figure