Industrial machine failures are rare and often difficult to label, making unsupervised anomaly detection a practical necessity. This paper studies a lightweight self-supervised contrastive representation learning approach for acoustic monitoring of centrifugal pumps using large unlabeled datasets. To this end, a compact convolutional audio encoder is trained with a SimCLR-inspired objective on proprietary pump audio and a subset of the publicly available MIMII-DUE pump recordings, without utilizing labels. The learned embeddings show meaningful structure, with recordings clustering both by pump station and by individual pump, indicating machine- and site-specific acoustic fingerprints, and some pumps exhibit temporal organization suggestive of gradual operational change. For evaluation, the frozen encoder is transferred to the held-out MIMII pump dataset containing real faults. A simple density model based on principal component analysis (PCA) and kernel density estimation (KDE) separates normal and abnormal sounds. Recall ranges from 63% to 85% across pump IDs at a fixed 5% nominal false-positive threshold. Overall, the results show that compact contrastive audio embeddings can provide interpretable structure in unlabeled industrial audio and support zero-shot anomaly detection, while comparison with the 51.6x larger contrastive encoder AADSCL shows that mean recall rises from 75% to 85% under the same PCA+KDE protocol.