Multi-Channel Time-Domain Boring-Vibration-Enhancement Method Using RNN Networks

Xiaolin Xu, Juhu Li, Huarong ZhangInsects 2023 · DOI

TL;DR

This paper introduces a novel multi-channel time-domain signal processing method, MultiSAMS, that uses recurrent neural networks (RNNs) and an attention mechanism to effectively denoise and separate faint boring vibrations from tree-trunk-boring insects, even in noisy environments. This advancement significantly improves the detection efficiency of these destructive pests, offering a more robust and less labor-intensive solution than traditional methods.

What Problem Does This Paper Address?

Wood-boring insects, like the emerald ash borer, are a major threat to forest ecosystems. Their larvae burrow into trees, causing significant damage that is often difficult to detect in its early stages. Traditional detection methods, such as manual inspection, are labor-intensive and have limited effectiveness. Acoustic monitoring, which detects the subtle vibrations produced by larvae feeding, offers a promising alternative. However, existing acoustic methods often rely on single sensors and are highly susceptible to environmental noise, leading to inaccurate detections. This paper aims to overcome this limitation by developing a multi-channel system that can effectively isolate and enhance these faint insect-generated vibrations from noisy real-world data.

Methods

The core of this research is the MultiSAMS (Multi-Channel Boring-Vibration Signal-Separation) model. It's an end-to-end, time-domain approach that leverages several key components:

  • Multi-Channel Input: Instead of a single sensor, the system utilizes multiple vibration sensors placed at different locations on the tree trunk to capture a richer set of signals.
  • Recurrent Neural Networks (RNNs): Specifically, the paper employs bidirectional GRU (Gated Recurrent Unit) networks. RNNs are well-suited for processing sequential data like audio vibrations, and the bidirectional nature allows them to capture dependencies in both forward and backward directions, improving context understanding.
  • Attention Mechanism (Multi-Head Self-Attention Module - MSAM): This is a crucial innovation. The attention mechanism allows the model to dynamically focus on the most relevant parts of the input signals and the most informative channels, effectively down-weighting noise and irrelevant information. This is achieved by learning to assign different weights to different features and channels.
  • Beamforming Techniques: The paper incorporates time-domain beamforming to combine signals from multiple sensors, aiming to enhance the target signal while suppressing noise.
  • Feature Extraction: The model uses one-dimensional convolutional layers to extract features from each channel and calculates normalized cross-correlation (NCC) to capture inter-channel relationships. These features are then fed into the attention mechanism.

The system is designed to be trained end-to-end, meaning the entire process from raw sensor data to denoised signal is optimized together.

Datasets

The study utilizes a combination of datasets:

  • Field Data: Collected from outdoor environments in Shunyi District and Changping District, Beijing. This data includes vibrations from white ash trees, categorized as infected but alive, dying from infection, and uninfected. Probes were inserted 3-5 cm below the bark.
  • Laboratory Data: Collected in a controlled environment using white ash trees. This data is used for more precise signal acquisition and also includes synthesized noise data to improve model generalization.
  • Synthetic Data: A portion of the dataset (20%) was generated through laboratory simulations, particularly to validate the impact of source location on signal acquisition.

The researchers also mention collecting data for other wood-boring insects, such as the wood-borer moth, for future research.

Key Insights and Ideas

  • The Power of Multi-Channel Sensing: Using multiple sensors significantly improves the ability to capture faint insect vibrations and reject environmental noise compared to single-sensor approaches.
  • Attention is Key for Noise Robustness: The attention mechanism is critical for allowing the model to selectively focus on the insect's boring vibrations and ignore distracting background noise, a common problem in outdoor acoustic monitoring.
  • Time-Domain Efficiency: The choice of time-domain processing, combined with GRU networks, offers a good balance between performance and computational efficiency, making it potentially suitable for deployment on resource-constrained devices.
  • GRU Outperforms LSTM in this Context: For this specific task, GRU networks showed superior performance and a smaller parameter count compared to LSTMs, suggesting they are a more efficient choice for this application.

Key Results

  • Improved Signal-to-Noise Ratio (SI-SNR) and Signal-to-Distortion Ratio (SDR): The MultiSAMS model demonstrated significant improvements in SI-SNR and SDR, with gains ranging from 5% to 15% compared to baseline models when using different numbers of sensors and network configurations.
  • Superior Performance with More Sensors: Using four vibration sensors consistently yielded better results than using two sensors, especially in terms of SDR, highlighting the benefit of increased spatial information.
  • Attention Mechanism Boosts Performance: Experiments showed a noticeable improvement in both SI-SNR and SDR when the attention mechanism module (MSAM) was included, regardless of the number of sensors used.
  • Effective Noise Reduction: Spectrogram analysis (Figures 7-10) visually confirms the model's ability to effectively denoise and separate the target boring vibration signals from noisy recordings.
  • High Recognition Rates: When combined with state-of-the-art recognition models like ResNetSE, the denoised signals from MultiSAMS led to higher recognition rates for identifying infested trees.

Limitations and Future Work

  • Computational Complexity: While the time-domain approach is efficient, the inclusion of multiple attention layers increases the model's overall complexity, posing a challenge for deployment on microcontrollers. Future work will focus on simplifying the model.
  • Limited Sensor Configurations Tested: The study primarily tested with two and four vibration sensors. Exploring the impact of a larger number of sensors is a future direction.
  • Overfitting Concerns: The paper notes that in some experiments, a four-layer model performed better than an eight-layer model, suggesting potential overfitting or loss of features with deeper networks. More field data and further optimization are planned.
  • Generalization to Other Pests: The current model is primarily tested for the emerald ash borer. Future work aims to extend its applicability to other wood-boring pest species by adapting and fine-tuning the model with diverse datasets.
  • Species and Population Recognition: The ultimate goal is to move beyond detecting a specific pest to recognizing different species and even estimating population levels, which will require further dataset expansion and model development.

Summary updated

Abstract

The larvae of certain wood-boring beetles typically inhabit the interior of trees and feed on the wood, leaving almost no external traces during the early stages of infestation. Acoustic techniques are commonly employed to detect the vibrations produced by these larvae while they feed on wood, significantly increasing detection efficiency compared to traditional methods. However, this method's accuracy is greatly affected by environmental noise interference. To address the impact of environmental noise, this paper introduces a signal separation system based on a multi-channel attention mechanism. The system utilizes multiple sensors to receive wood-boring vibration signals and employs the attention mechanism to adjust the weights of relevant channels. By utilizing beamforming techniques, the system successfully removes noise from the wood-boring vibration signals and separates the clean wood-boring vibration signals from the noisy ones. The data used in this study were collected from both field and laboratory environments, ensuring the authenticity of the dataset. Experimental results demonstrate that this system can efficiently separate the wood-boring vibration signals from the mixed noisy signals.
BibTeX
@inproceedings{xu2023multi,
  author = {Xiaolin Xu and Juhu Li and Huarong Zhang},
  title = {{Multi-Channel Time-Domain Boring-Vibration-Enhancement Method Using RNN Networks}},
  booktitle = {Insects},
  year = {2023},
  doi = {10.3390/insects14100817},
}
Figure