This paper introduces a novel pipeline, called AquaSignal, tailored for processing and interpreting underwater acoustic data, with functionality spanning signal processing, noise suppression, classification, and anomaly detection. Built taking into account the complexities of turbulent oceanic conditions, the system combines cutting-edge deep learning models to boost both the fidelity and robustness of acoustic classification. Evaluation is carried out using a hybrid dataset that merges samples from the Deepship and Ocean Networks Canada (ONC) datasets, encompassing a broad range of authentic underwater acoustic environments. At its core, the pipeline incorporates a U-Net for denoising, a ResNet18-based classifier for identifying acoustic signatures, and an AutoEncoder architecture dedicated to flagging unexpected or previously unseen patterns. This marks the first in-depth exploration combining these specific neural architectures for analysing vessel-generated underwater sounds. Performance benchmarks show improvement in both signal clarity and detection, with classification reaching 71.3% and anomaly detection achieving 91.5%. Although some existing models outperform AquaSignal in classification under different validation splits, such discrepancies highlight the challenges of direct comparison. The results suggest that AquaSignal holds substantial promise for deployment in real-time acoustic surveillance applications across scientific, ecological, and naval sectors.