Wire Arc Directed Energy Deposition (WA-DED) is emerging as a cost-effective method for fabricating large metallic components. Despite its robustness and high deposition rates, the process remains prone to instabilities and defects, making reliable monitoring essential for industrial adoption. However, most advanced monitoring strategies rely on large, labelled datasets, whose generation is costly and impractical in real manufacturing environments. This work proposes an unsupervised monitoring framework that learns normal behaviour exclusively from defect-free depositions and detects online deviations using welding current and voltage signals. A set of time–frequency features, comprising Fast Fourier Transform (FFT) descriptors and multi-level Discrete Wavelet Transform (DWT) statistics, was extracted to represent the signals. An Isolation Forest model was trained on in-control data and benchmarked against Statistical Process Monitoring (SPM) based on Shewhart control charts applied to time- and frequency-domain energy features. Validation on two multi-layer wall specimens containing both good and anomalous regions showed that the proposed approach achieved an F1-score of 85.3%, a 57% improvement over the best control chart (54.3%). The method exhibited greater sensitivity to arc-related instabilities while maintaining high precision and avoiding excessive false alarms. The framework requires no defect labels and only normal data for training, aligning with industrial constraints where labelled anomaly datasets are scarce. Limitations remain, notably the absence of defect localisation and the qualitative nature of alerts. Future work will address these aspects through the development of a quantitative quality index, improved localisation, and extension to complex geometries such as overhanging structures, alongside multi-sensor and multimodal fusion.