Identification of sloshing noises using convolutional neural network

Area of Science:

• Automotive Engineering
• Acoustics
• Machine Learning

Background:

• Vehicle noise reduction has intensified focus on secondary sources like fuel tank sloshing.
• Sloshing generates distinct 'hit' and 'splash' noises due to fluid-wall and fluid-fluid interactions, respectively.
• Different noise types require tailored control strategies, necessitating accurate identification during design.

Purpose of the Study:

• To develop a convolutional neural network (CNN) based method for identifying fuel tank sloshing noises.
• To evaluate the CNN's performance against traditional feature-based methods.
• To assess the model's applicability in real-world driving scenarios.

Main Methods:

• A reciprocating test setup generated controlled sloshing noises (hit and splash) in a fuel tank.
• A CNN was trained and tested on collected acoustic data.
• CNN-identified features were compared with hand-crafted features using Support Vector Machines (SVM).

Main Results:

• The CNN demonstrated effective identification of sloshing noises under varying conditions (fill level, excitation, baffles).
• CNN-based feature identification showed comparable or superior accuracy to SVM with hand-crafted features.
• The model proved applicable in simulated practical scenarios, such as vehicle braking.

Conclusions:

• CNNs offer a robust methodology for identifying distinct fuel tank sloshing noise types.
• Accurate noise identification facilitates the design of quieter vehicles by enabling targeted noise control.
• This approach supports the development of advanced driver-assistance systems by characterizing acoustic events.