The early detection of lung diseases in community healthcare settings is crucial, yet clinicians face challenges such as equipment limitations and heavy workloads. While AI-based automated screening using wearable devices offers a solution, developing robust models is difficult owing to the scarcity of abnormal data and the prevalence of environmental noise in real-world settings. We propose a noise-robust unsupervised anomaly detection framework for lung sounds acquired from wearable devices. This framework relies on continuous acoustic data acquired through a microphone sensor embedded in a prototype wearable device. The model is trained exclusively on normal lung sounds using a U-Net–based autoencoder with a composite loss function (mean squared error and structural similarity loss) to learn the manifold of normal respiration. The primary contribution of this work is a novel anomaly detection algorithm designed to distinguish between pathological anomalies and environmental noises. While pathological sounds (e.g., fine crackles) are stationary and synchronized with respiration, environmental noises (e.g., coughing and friction) are typically transient. The proposed algorithm evaluates the temporal persistence of reconstruction errors across sliding windows to calculate an anomaly detection rate. Experimental results demonstrate that while standard autoencoders struggle with noise, the proposed method successfully differentiates between persistent disease-related anomalies and transient noise artifacts. In this work, we demonstrate the feasibility of robust, unsupervised lung sound screening in noisy, real-world environments.