Jonathan Li

Date of Award

Spring 2023

Document Type



Santa Clara : Santa Clara University, 2023.


Electrical and Computer Engineering

First Advisor

Dat Tran


Air Conditioners are essential to human life. In an age of sudden temperature changes, moving systems, particularly HVAC (Heat Ventilation and Air Conditioning) systems, are the primary source to physically and financially protect the health of all workers, employees, and students. Air Conditioners are prone to mechanical and electrical malfunction/breakdown due to excessive use. Regular maintenance and service intervals are helpful but do not guarantee free-malfunction systems. When the systems break down, especially commercial systems, the repair cost can be highly expensive and time-consuming. Can we detect early signs of potential problems in the systems to minimize the repair cost and time? Our project attempts to explore a probable solution.

Since our solution must be simple and affordable, we propose an Air Conditioner Detector to alert users of any malfunction within the system by detecting early signs of worn-out parts. Based on sounds generated by the conditioner’s blower motor, the detector will detect abnormal sounds from normal ones in a running HVAC system. Once an abnormal sound is detected, the Detector will generate an alarm signal through an alarm system (Fig. 2) connected to the detector (Fig. 1). For the project to be feasible, the performance accuracy of the Detector in identifying abnormal sounds generated by the system should be, at least, 85%. In our design, we build a machine-learning model for the SparkFun EOS S3 Board (Fig. 1) and train the model with our recorded dataset. The dataset is a collection of normal sounds of an HVAC system and artificially abnormal sounds. Since the abnormal sounds generated by malfunctioning HVAC systems are not available (due to restricted access to the utility rooms), we need to create artificially abnormal sounds for training and testing our design. Using the built-in microphone on the SparkFun EOS S3 board, we record raw sound data and save them in “wave” format files. Next, with Data Capture Lab, the available tool from SensiML, we segment and label raw data files to form training and testing datasets. Finally, we build a machine learning model with Analytic Studio, an machine learning generative tool from SensiML. Since raw sound data have many features, it is challenging to use all features in identifying abnormal from normal sounds. Our experiments have shown that the four main features of sound data are sufficient in characterizing sounds into categories: amplitude, signal-to-noise ratio (SNR), rate of change, and frequency. Our testing results show that the SparkFun EOS S3 Board with machine learning model can correctly identify 95.24% (Fig. 19, Table 5) of the abnormal sounds from the testing dataset. Our project demonstrates that an IoT board with a machine learning model is a simple and affordable solution to detecting early signs of worn-out parts in an HVAC system.