This article proposes a hybrid unsupervised learning method as a combination of a novel two-level artificial neural network (TLANN) algorithm for data normalization regarding the removal of environmental variations and k-means clustering (KMC) for damage detection. In the proposed TLANN algorithm, feature samples are fed into the first neural network to generate its output and determine a residual matrix from the difference between the network input and output. In such a case, a new residual matrix from the difference between the input and output of the second network is extracted as the main feature for damage detection via the KMC, classical Silhouette value technique is applied to determine the number of clusters. The contribution of this article is to develop an innovative hybrid unsupervised learning method. The great advantages of this method consist in dealing with the negative effects of environmental variability and increasing the detectability of damage. The performance and reliability of the proposed method are validated by the well-known Z24 bridge along with several comparisons. Results show that the proposed hybrid method is highly capable of detecting damage and removing strong environmental variations. It is also observed that this method is superior to some classical and existing techniques.