Photoplethysmography (PPG) is gradually becoming popular tool for cardiovascular and respiratory function monitoring under ambulatory condition. However, these measurements are prone to motion artifact (MA) corruption, and hence, signal quality assessment (SQA) is essential before computerized analysis. The published research on PPG SQA, mostly utilizing supervised learning approaches, suffer from the universality of feature selection against PPG morphology variability. Secondly, beat detection from the MA corrupted is a challenging task and partly limits the success of SQA utilizing beat segmenting approaches. The present research describes an unsupervised learning approach for identification of ‘clean’, ‘partly clean’ and ‘corrupted’ segments in the MA contaminated PPG data. Few entropy features and some signal complexity related features calculated by statistical methods in a 5 s window were fed to a self-organizing map (SOM) for direct quality assessment of PPG data. The number of input node to the SOM was 7 and the output was connected to a square matrix consisting of 25 nodes. The multiclass classification model achieved 94.10%, 89.27%, 92.67% accuracy score for the three classes respectively on 200 min of PPG data collected from 30 healthy and CVD human volunteers under mild to high level of hand movement. The model achieved better result than recently published work utilizing non-segmenting approach based PPG SQA.