The increasing complexity of civil engineering structures makes continuous Structural Health Monitoring (SHM) essential [1]. This is achieved through embedded sensors that generate large volumes of data [2, 3]. We propose an automatic data analysis approach to detect anomalies and abnormal structural behavior, based on an unsupervised machine learning method: Principal Component Analysis (PCA)[4, 5, 6]. Our approach involves using training samples from sensor data acquired when the structure is expected to be in normal operating conditions. PCA is employed to identify the subspace within the sensor data that retains the minimal yet essential information required to characterize samples under normal conditions. Subsequently, we introduce a criterion for detecting abnormal behaviors, based on a statistical analysis of the reconstruction error of the samples using the principal components derived from the PCA. The method was tested using synthetic data generated from a numerical model of a reinforced concrete beam. This case study is relevant for the comprehension and validation of the approach, as its behavior can be verified using theoretical formulations. Training data represent the normal elastic behavior of the structure, while test data are obtained from simulations under higher loads leading to structural damage. Results show that, with a representative training set, the proposed approach can reliably detect damage shortly after its onset. Nevertheless, further improvements are envisaged and will be addressed in future work, including earlier damage detection at its onset and the incorporation of noise effects in the anomaly detection criterion within the training dataset.