Anomaly detection is a widely utilized technique in the field of wireless sensor networks (WSNs) data stream analysis, aimed at identifying unusual events or anomalies in an early stage. However, the constraints of WSN applications pose a significant challenge in achieving effective and efficient anomaly detection. In this work, we proposes a new multi-head attention-based Bi-LSTM approach for anomaly detection in multivariate time-series. Rather than modeling the time series of individual sensor independently, the proposed approach models the time series of multiple sensors concurrently, taking into account potential latent interactions among them, thereby enhancing the accuracy of anomaly detection. The proposed approach does not require labeled data and can be directly applied in real-world scenarios where labeling a large stream of data from heterogeneous sensors is both difficult and time-consuming. Finally, empirical evaluations using a real-world WSN demonstrate effectiveness and robustness of the proposed approach, outperforming traditional deep learning approaches.