Acoustic prediction of flowrate: varying liquid jet stream onto a free surface

Balamurali B. T., Edwin Jonathan Aslim, Yun Shu Lynn Ng, Tricia Li Chuen Kuo, Jacob Chen Shihang, Dorien Herremans, Lay Guat Ng, Jer-Ming Chen

TL;DR

This paper explores using the sounds generated by a liquid jet hitting a free surface to predict the jet's flowrate and trajectory. They propose two methods: a machine learning approach using audio features and a more direct acoustic parameter-based method, both showing promising results in reconstructing flow dynamics from sound alone.

What Problem Does This Paper Address?

Understanding the flowrate and dynamics of liquid streams is critical in numerous real-world applications, from natural phenomena like waterfalls to engineered systems like toilets and pumps. When these streams interact with free surfaces, they produce distinct sounds. This paper investigates whether these acoustic signatures can be leveraged to non-invasively predict the flowrate and its trajectory, offering a potentially simpler and more accessible alternative to traditional measurement techniques. The challenge lies in deciphering the complex acoustic signals generated by the fluid-fluid interaction and correlating them with precise flow characteristics.

Methods

The researchers employed two distinct methodologies:

  • Machine Learning (ML) Approach: This method involves training supervised ML algorithms (k-nearest neighbors and gradient boosting) on audio recordings of water jets. Mel-frequency Cepstral Coefficients (MFCCs), a standard feature set in audio recognition, were extracted from 0.1-second audio frames. These features, along with the corresponding flowrate data (obtained gravimetrically), were used to train the models. The final prediction for each audio frame was determined by soft-voting the outputs of the two classifiers. This approach aims to directly predict the flowrate and subsequently reconstruct the flowrate trajectory.
  • Acoustic Parameter Method: This approach bypasses explicit machine learning training. Instead, it focuses on analyzing specific acoustic parameters derived from the sound's spectral energy. After pre-processing the audio to remove environmental noise, the researchers identified two key frequency regions (<1 kHz and >1 kHz) where energy was concentrated. They hypothesized that these regions relate to different aspects of the liquid-liquid interaction (direct impact vs. secondary splashes). By analyzing the energy within these bands, they aimed to estimate the flowrate trajectory. This method provides a relative flowrate estimation.

Both methods involved normalizing the resulting flowrate trajectories to a 0-1 scale for comparison.

Datasets

The study utilized a custom dataset collected by recording water jets of varying flowrates falling into a pool. The flowrate was simultaneously measured using a gravimetric flowmeter (LABORIE). A total of 51 flow episodes were recorded, each consisting of audio data and corresponding flowrate measurements. For the ML approach, 70% of these recordings (35 episodes) were used for training, and the remaining 30% (16 episodes) were used for testing.

Key Insights and Ideas

  • Sound as a Flow Indicator: The core idea is that the complex acoustic interactions at a liquid-liquid interface (splashing, bubbling) contain discernible information about the flow dynamics, acting as unique acoustic signatures.
  • Dual Frequency Regions: The analysis revealed that distinct frequency bands in the audio spectrum (<1 kHz and >1 kHz) appear to correlate with different physical phenomena of the liquid jet impact, suggesting a nuanced relationship between sound and flow.
  • ML vs. Direct Acoustic Analysis: The study effectively compares a data-driven ML approach with a more physics-informed acoustic parameter method, highlighting their respective strengths and weaknesses.
  • Non-Invasive Measurement Potential: The success of both methods suggests a promising avenue for non-invasive, remote monitoring of liquid flow dynamics, which could be valuable in applications where traditional sensors are impractical.

Key Results

  • Both the ML and acoustic parameter methods demonstrated comparable performance in accurately predicting the flowrate trajectory, as measured by Euclidean and Fréchet distances.
  • The ML method was able to directly predict the absolute flowrate for each audio frame, resulting in a predicted flow volume that showed a high correlation (Pearson correlation coefficient of 0.83) with the actual volume.
  • The acoustic parameter method, while providing only relative flowrate information, also offered good estimation of flowrate trajectories.
  • The low-frequency band (<1 kHz) in the acoustic analysis tended to underestimate flowrate, while the high-frequency band (>1 kHz) tended to overestimate it. Boosting the low-frequency band showed some improvement.

Limitations and Future Work

  • ML Computational Overhead: The ML approach requires significant training data and computational resources for feature extraction and model training, making it less suitable for real-time, low-power applications.
  • Acoustic Parameter Limitations: The acoustic parameter method, while computationally efficient, only provides relative flowrate information and is unable to predict absolute volumes directly.
  • Normalization Effects: The authors note that the comparison of Fréchet distances might be influenced by the normalization applied to the trajectories, suggesting this warrants further investigation.
  • Future Directions: The paper suggests that the ML method's ability to predict absolute flowrate, coupled with its high correlation to actual volumes, could be further exploited for calibration to achieve even more accurate flowrate trajectories. The potential for using these methods in real-life applications is also highlighted.

Summary updated

Abstract

Information on liquid jet stream flow is crucial in many real world applications. In a large number of cases, these flows fall directly onto free surfaces (e.g. pools), creating a splash with accompanying splashing sounds. The sound produced is supplied by energy interactions between the liquid jet stream and the passive free surface. In this investigation, we collect the sound of a water jet of varying flowrate falling into a pool of water, and use this sound to predict the flowrate and flowrate trajectory involved. Two approaches are employed: one uses machinelearning models trained using audio features extracted from the collected sound to predict the flowrate (and subsequently the flowrate trajectory). In contrast, the second method directly uses acoustic parameters related to the spectral energy of the liquidliquid interaction to estimate the flowrate trajectory. The actual flowrate, however, is determined directly using a gravimetric method: tracking the change in mass of the pooling liquid over time. We show here that the two methods agree well with the actual flowrate and offer comparable performance in accurately predicting the flowrate trajectory, and accordingly offer insights for potential real-life applications using sound.
BibTeX
@inproceedings{t2020acoustic,
  author = {Balamurali B. T. and Edwin Jonathan Aslim and Yun Shu Lynn Ng and Tricia Li Chuen Kuo and Jacob Chen Shihang and Dorien Herremans and Lay Guat Ng and Jer-Ming Chen},
  title = {{Acoustic prediction of flowrate: varying liquid jet stream onto a free surface}},
  booktitle = {Spcom},
  year = {2020},
  doi = {10.1109/spcom50965.2020.9179604},
}
Figure