Convolutional Recurrent Neural Network-Based Event Detection in Tunnels Using Multiple Microphones.

Area of Science:

• Signal Processing
• Machine Learning
• Acoustics

Background:

• Tunnel environments present severe noise challenges for existing sound event detection (SED) methods.
• Accidents in tunnels, such as crashes and tire skids, produce abnormal sounds crucial for early detection.
• Reduced accuracy in current SED methods can lead to delayed responses and escalating accident severity.

Purpose of the Study:

• To develop an accurate and robust sound event detection (SED) method specifically for noisy tunnel environments.
• To address the limitations of existing SED techniques in handling severe background noise.
• To enhance tunnel safety by enabling real-time detection of abnormal acoustic events.

Main Methods:

• A novel SED method combining Non-negative Tensor Factorization (NTF) for source separation and an adapted Convolutional Recurrent Neural Network (CRNN) classifier.
• NTF incorporates online noise learning to adaptively separate event signals from complex tunnel noise.
• The CRNN utilizes parallel event and noise convolution layers, followed by recurrent layers, processing mel-filterbank features.

Main Results:

• The proposed CRNN-based SED method with online noise learning achieved a 91.07% recognition rate in severe noise conditions.
• It reduced the relative recognition error rate by 56.25% in low noise and by 47.40% and 28.56% compared to GMM-HMM and conventional CRNN in severe noise, respectively.
• Real-time detection capability was demonstrated with a processing time of 599 ms for a one-second audio signal.

Conclusions:

• The proposed SED method effectively handles severe noise in tunnels, significantly outperforming conventional approaches.
• Online noise learning within NTF and the adapted CRNN architecture are key to achieving superior detection accuracy.
• This method offers a viable solution for real-time monitoring and accident prevention in tunnel environments.