Session: 01-06-01: AI and Machine Learning in Acoustics and Vibrations

Paper Number: 172998

Machine-Learning Classification of Air Blower Speed Using Acoustic Signatures 

Compact acoustic monitoring can reveal in a non-invasive way the operating state of small appliances. Existing acoustic classification tools fail to infer both thermal and airflow settings of small blowers using ultra‐short audio snippets. A machine-learning pipeline framework that classifies an air blower’s speed and heat settings from three-second audio recordings. Air blower speed is defined at three levels: low, medium, and high, heat is defined at four levels: none, low, medium, and high. Together these yield twelve distinct operating states. A controlled factorial study produced a dataset from four consumer air-blower configurations. Two units were identical except one carried an outlet attachment, the other two units were unique models. Trials were conducted in a sound-treated laboratory at 76 degrees F. Each heat-speed combination ran and was recorded for 20 second to reach a steady state, after an initial 10 second warm-up period. Three replicates per combination yielded 144 audio–vibration trials. Three PCB ¼″ condenser microphones are located at 12″ from the outlet nozzle at 0°, 45°, and 270° around the blower outlet axis. An accelerometer was attached to the blower housing to measure device vibrations 3 inches from the outlet on the top of the blower housing. Sound and vibration signals are recorded using a National Instruments DAQ at sampling rate of 48 kHz and a zero-phase band-pass filter was applied in the range from 20 Hz to 20 kHz to remove airflow noise and electrical interference. From each three-second clip there were three categories of spectral features extracted from the three second clips. Octave-band energies in eight standard bands from 125 Hz to 8 kHz. Mel-frequency cepstral coefficients (MFCCs)—thirteen coefficients that capture the perceptual spectral envelope were computed using a 25 ms frame. To quantify noise-vibration coupling, coherence peaks were extracted between microphone and accelerometer signals via cross-spectral analysis. All features were concatenated into a 30-dimensional feature vector. A random-forest classifier with 100 decision trees was trained and evaluated by a stratified 10-fold cross validation over the 144 samples. Accuracy defined in this context as the fraction of correct predictions reached 95%. Precision is true positives over predicted positives, and recall the true positives over actual positives both exceeded 92% for every class. The confusion matrix shows that most errors (under 5%) occurred between adjacent speed levels at the same heat. Misclassifications across different heat levels were below 1%.

Generalizability was assessed on 27 additional clips from three new blower models sampled at three heat and three speed settings. Applying the trained classifier without retraining produced a drop in accuracy of only 3%, demonstrating robust transfer across device types and outlet configurations. Feature importance was ranked using the mean decrease in Gini impurity metric. Mid-frequency octave bands (500 Hz–2 kHz) accounted for 60% of total importance. Coherence features contributed 25% and MFCCs contributed the remaining 15%. This ranking shows the pivotal role of both spectral content and acoustic-vibration coupling for discriminating operating states. End to end processing of the clip in MATLAB per clip was under 1 s on a standard i7 laptop with 40 ms for feature extraction and 12 ms for inference showing promises for real time classification.

These results demonstrate that very short acoustic snapshots can reliably infer detailed operating states. The method requires no specialized hardware beyond a calibrated microphone and accelerometer and it runs in standard machine-learning frameworks. Potential applications include smart-home monitoring, preventive maintenance, and user feedback systems. Future work will implement continuous streaming, explore edge computing for on-device inference, test one-second audio snippets, and expand to multi-device environments.

Presenting Author: Trisha Campanaro The Pennsylvania State University - Harrisburg

Presenting Author Biography: Trisha Campanaro is a Ph.D. student in Engineering Systems at Penn State Harrisburg. She holds a B.S. in Mechanical Engineering Technology from Penn State University—where her undergraduate capstone focused on railroad-car connector design—and an M.S. in Engineering Management with a focus on Engineering Analytics and Project Management from Rowan University.
Her research spans acoustics, vibration analysis, and machine learning for electromechanical and/or mechanical systems. She leads experimental campaigns on blower noise and vibration, developing calibration protocols, signal-processing pipelines, and statistical analysis workflows. She has presented at multiple Penn State research poster competitions and won the Penn State Harrisburg competition in 2024.
In the classroom, she has taught Engineering Design I and served as a teaching assistant for Instrumentation and Advanced CAD courses. She mentors undergraduate STEM students in instrumentation and data analysis.
Outside academia, she works part-time as a special project coordinator at Rowan University, focusing on analytics, data science and project management. She is treasurer of the Graduate and Professional Student Council, member of ASME, and an active member of Women in STEM. She volunteers with the National FFA, judging Agriscience events, and supports AACSB accreditation efforts at Rowan University. Her broader research interests include biomedical engineering, medical devices, and materials science.

Authors:

Trisha Campanaro The Pennsylvania State University - Harrisburg
Issam Abu-Mahfouz The Pennsylvania State University - Harrisburg