Composite materials have a myriad of applications in complex engineering systems, and multiple structural health monitoring strategies have been developed. However, these methods are challenging due to signal attenuation and excessive noise interference in composite materials. Signal processing can capture a small difference between the input-output signals associated with the severity of the damage in composites. Thus, the research question is "can signal processing techniques reduce the required number of features and assess the randomness of fatigue damage classification in composite materials using machine learning algorithms?" To answer this question, piezoelectric signals for carbon fiber reinforced polymer test specimens were taken from NASA Ames prognostics data repository. A framework based on a comparative analysis of signals was developed. For the first specific aim, the effectiveness of features based on statistical condition indicators of the sensor signals were evaluated. For the second specific aim, actuator-sensor signal pair were analyzed using cross-correlation to extract two features. These features were used to train and test four supervised machine learning (ML) algorithms for damage classification and their performance was discussed. For the third specific aim, randomness in the dataset of fatigue damage of the specimens was assessed. Results showed that by signal processing, the requirement of features for training ML was reduced with the improvement in the performance of ML. The randomness was captured by the utilization of two specimens from the same material. This work contributes to the improvement of intelligent damage classification of composite materials, operating under complex working conditions.