The model based on autoencoder performs poorly in identifying anomalies in concrete dam measured data, and existing methods struggle to accurately detect minor abnormal values in time series. This paper constructs a novel unsupervised anomaly detection model that integrates long short-term memory network (LSTM), multi-head attention mechanism (MA), and generative adversarial network (GAN) into memory-augmented deep autoencoder (MemAE) model, namely the multi-head attention LSTM- memory-augmented deep autoencoder- generative adversarial network (MALSTM-MemAE-GAN). This model takes MemAE as the core, replaces the encoder part with LSTM to extract hidden features dependent on time series, and then uses MA to enhance the model’s focusing ability on minor abnormal values after the memory-augmented module. GAN is introduced to make the reconstructed data closer to the real samples. Taking the displacement and crack data of a concrete dam as an example, two experimental datasets are established (The abnormal intensity of the second experimental datasets is halved on the basis of the first experimental datasets). Using a multi-metrics evaluation system centered on the F1 score, ablation and comparative experiments are conducted on the proposed MALSTM-MemAE-GAN model. Dynamic detection is achieved through a sliding window approach. Results demonstrate average F1 scores of 0.9756 and 0.9435 for static detection across two experimental datasets, validating the model’s robust performance and accurate detection capability for minor abnormal values. Furthermore, the model enables dynamic detection of dam measured data with excellent results. This paper provides scientific basis and technical support for the dam structural health monitoring.