In the nuclear industry, the relentless pursuit of operational excellence and reliability is paramount, especially considering the critical importance of safety and efficiency. This sector, a key component of global energy supply, is constantly driven to innovate and optimize its equipments. Early fault detection plays a central role in this endeavor, as it not only helps to prevent potential incidents, but also maximizes production. Traditional anomaly detection methods, primarily based on clustering algorithms, often overlook minor yet warning faults. Addressing this, our paper introduces a multiscale clustering approach, advancing beyond classical methods. This methods achieves a satisfactory classification of anomalies and culminates in a silhouette score of 90%. Additionally, it facilitates the computation of a preventive maintenance indicator.