The paper presents a novel data-driven baseline-free method for guided wave structural health monitoring (GWSHM). The proposed method utilizes unsupervised shapelets and shift-invariant dictionary learning to achieve baseline-free monitoring. Unsupervised shapelets are time series patterns that can be used to represent guided wave measurements and detect anomalies resulting from damage. Shift-invariant dictionary learning is the technique used to learn a set of shift-invariant patterns from the data, for detecting and characterizing structural changes. Furthermore, the proposed method supports the transfer of guided wave information across structures with different sizes, material properties, sensor configurations and structural complexities. The experiments conducted on composite coupon and complex stiffened panel validate the effectiveness and robustness of the proposed method for damage detection and localization. It demonstrates comparable performance to several baseline-free techniques when localizing single damages on composite coupons under 20°C variations, indicating its potential for practical applications in real-world scenarios.