Classification of damages in honeycomb composite sandwich structure (HCSS) is important to decide remedial actions. However, previous studies have only detected damages using deviations of monitoring signal from healthy (baseline) using a guided wave (GW) based structural health monitoring system. Classification between various types of damages has not been reported for challenging cases. We show that using careful feature engineering and machine learning it is possible to classify between various types of damages such as core crush (CC), high density core (HDC), lost film adhesive (LFA) and teflon release film (TRF). We believe that we are the first to report numerical models for four types of damages in HCSS, which is followed up with experimental validation. We found that two out of four damages affect the GW signal in a particularly similar manner. We extracted and evaluated multiple features from time as well as frequency domains, and also experimented with features relative to as baseline as well as those that were baseline-free. Using Pearson's correlation coefficient based filtering, redundant features were eliminated. Finally, using an optimal feature set determined using feature elimination, high accuracy was achieved with a random forest classifier on held-out signals. For evaluating performance of the proposed method for different damage sizes, we used simulated data obtained from extensive parametric studies and got an accuracy of 77.89%. Interpretability studies to determine importance of various features showed that features computed using the baseline signal prove more effective as compared to baseline-free features.