Comparative Evaluation of Machine Learning Algorithms for Multiclass Fault Diagnosis of Rotary Compressors Using Triaxial Vibration Signals

Compressors are essential in industrial operations, managing power generation, heating, ventilation, and airconditioning (HVAC) systems and refrigeration. The operation of compressor systems produces different kinds of hazards that create production stoppages along with financial setbacks and dangerous risks for personnel safety. Vibration-based monitoring is an effective method for early fault detection. The research analyzes multiple machine learning algorithms for compressor fault classification through vibration data analysis by implementing Random Forest, XGBoost, LightGBM, Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP). The ADXL345 accelerometer recorded vibration data under four operational conditions: healthy, tilted, vibration-added, and with a loosely mounted sensor. Every 2 -second window produced 20 timedomain features (mean, standard deviation, skewness, kurtosis per axis, plus RMS and its mean, standard deviation, skewness, kurtosis) and subsequently $z$-score normalization. Among the analyzed models, XGBoost showed the highest classification accuracy at $\mathbf{8 9. 1 0 \%}$, surpassing all other algorithms across key evaluation metrics. Random Forest followed with an accuracy of 86.32 %. Model performance was assessed using accuracy, precision, recall, and F1-score. Owing to its advanced diagnostic proficiency and relatively moderate computational requirements, XGBoost stands out as the most suitable candidate for on-device compressor fault monitoring.