In Industry 4.0, effective CNC tool wear monitoring and detection in a timely manner is very crucial. Traditional predictive maintenance often relies on supervised learning models and requires an extensive amount of failure labeled data. Raw sensor signals are often high-dimensional and contaminated by non-cutting noise, leading complex models to overfit on limited datasets. This study proposed a robust unsupervised anomaly detection framework that identifies tool wear eliminating signal noise using only healthy operational data. The approach isolates active cutting stages by categoric filtering. The raw sensor data is processed by a Fast Fourier Transform to yield spectral attributes, such as peak amplitude and frequency, which define the tool's vibrating pattern, suppressing the effect of transient anomalies. One-Class Support Vector Machine learning is employed to learn a decision boundary from normal operating conditions only. The proposed hybrid approach achieved an F1-score of 0.82 and a recall of 0.90, demonstrating that substantial labeling, typically required in supervised learning, is not necessary for effective Predictive Maintenance in Manufacturing.