This paper presents a machine learning framework that uses the bag of visual words (BOVW) for structural health monitoring (SHM) of a composite sandwich structure (CSS) using ultrasonic guided wave (GW) signals. Towards this, experimental analysis of GW propagation in CSS has been carried out for the healthy-state and multiple skin-to-core disbond cases. The registered time-domain signals from the assigned piezoelectric transducer networks on the CSS are converted to time-frequency scalograms by performing a continuous wavelet transform. Eventually, a BOVW based machine learning framework is proposed that uses the speeded-up-robust features for the features extraction and support vector machine for classification of CSSs with and without skin-to-core disbond. The proposed machine learning framework shows its SHM potential to characterise the CSS for healthy and disbond conditions (different locations) with high validation and test accuracy for an unseen dataset. A series of parametric studies are also carried out to analyse the influence of different grid sizes and polynomial order for the proposed framework.