This thesis presents a complete, low-cost vibration-based condition monitoring system designed for the early detection of faults in induction motors using modern machine learning techniques. Driven by the growing need for predictive maintenance in industrial settings, the work integrates all steps of the measurement and classification process, starting with vibration acquisition. This is followed by analog signal conditioning, digital sampling, wireless data transmission, and finally, machine learning-based fault classification. A low-noise ADXL335 accelerometer senses mechanical vibrations, while a custom-designed analog front end handles AC coupling, bias shifting, amplification, filtering, and protection to ensure signal compatibility with the ESP32 12-bit SAR ADC. The embedded system conducts timer-driven sampling at 6 to 10 kS/s, streams the digitized signal via UDP over Wi-Fi, and enables real-time data acquisition without significant packet loss. A digital twin of the motor, created in MATLAB/Simulink, generates high-quality synthetic datasets that represent various fault conditions, including rotor imbalance and shaft misalignment. These synthetic signals are combined with real measurements to train multiple machine learning models, such as Support Vector Machines, Random Forests, and a lightweight Convolutional Neural Network. The results show that the proposed approach can reliably identify subtle changes in vibration signatures and accurately predict early-stage motor faults. The combination of embedded data acquisition, analog conditioning, and data-driven classification offers a flexible and scalable base for predictive maintenance. Overall, this work highlights that modern machine learning methods, when used with inexpensive hardware platforms, can significantly improve the reliability and efficiency of industrial motor systems, providing a practical way toward intelligent maintenance solutions.