Machine　learning　has　become　an　influential　and　effective　tool　in　numerous　civil　engineering　applications,　especially　in　the　field　of　structural　health　monitoring　(SHM).　Recently,　the　emergence　of　self-supervised　learning　has　led　to　the　development　of　many　industries,　and　its　accuracy　and　stability　are　superior　to　previous　methods.　Self-supervised　learning　learns　generalizable　information　representation　from　unlabeled　mixed　data　by　solving　pretext　tasks,　and　this　feature　is　exactly　in　line　with　the　mixed　and　unlabeled　data　in　the　SHM　field.　This　is　of　great　significance　to　the　improvement　of　detection　accuracy　in　SHM　practical　applications.　Therefore,　this　paper　proposes　a　new　self-supervised　method　for　structural　damage　detection.　The　key　to　this　method　is　that　we　use　two　self-supervised　pretext　tasks　to　learn　the　latent　feature　representation　of　the　data,　and　we　introduce　homoscedastic　uncertainty　for　automatically　assigning　weights　to　the　two　pretext　tasks.　The　relative　confidence　between　tasks　is　captured,　the　impact　of　noise　on　tasks　is　reduced,　the　pretext　tasks　can　better　learn　data　feature　representation,　and　the　purpose　of　improving　the　accuracy　of　damage　detection　is　achieved.　©　2023　SPIE.