Anomaly　detection　in　time　series　data　is　crucial　for　many　fields　such　as　healthcare,　meteorology,　and　industrial　fault　detection.　However,　traditional　unsupervised　time　series　anomaly　detection　methods　suffer　from　biased　anomaly　measurement　under　contaminated　training　data.　Most　of　existing　methods　employ　hard　strategies　for　contamination　calibration　by　assigning　pseudo-label　to　training　data.　These　hard　strategies　rely　on　threshold　selection　and　result　in　suboptimal　performance.　To　address　this　problem,　in　this　paper,　we　propose　a　novel　unsupervised　anomaly　detection　framework　for　contaminated　time　series　(NegCo),　which　builds　an　effective　soft　contamination　calibration　strategy　by　exploiting　the　observed　negative　correlation　between　semantic　representation　and　anomaly　detection　inherent　within　the　autoencoder　framework.　We　innovatively　redefine　anomaly　detection　in　data　contamination　scenarios　as　an　optimization　problem　rooted　in　this　negative　correlation.　To　model　this　negative　correlation,　we　introduce　a　dual　construct:　morphological　similarity　captures　semantic　distinctions　relevant　to　normality,　while　reconstruction　consistency　quantifies　deviations　indicative　of　anomalies.　Firstly,　the　morphological　similarity　is　effectively　measured　based　on　the　representative　normal　samples　generated　from　the　center　of　the　learned　Gaussian　distribution.　Then,　an　anomaly　measurement　calibration　loss　function　is　designed　based　on　negative　correlation　between　morphological　similarity　and　reconstruction　consistency,　to　calibrate　the　biased　anomaly　measurement　caused　by　contaminated　samples.　Extensive　experiments　on　various　time　series　datasets　show　that　the　proposed　NegCo　outperforms　state-of-the-art　baselines,　achieving　an　improvement　of　6.2%　to　26.8%　in　Area　Under　the　Receiver　Operating　Characteristics　(AUROC)　scores,　particularly　in　scenarios　with　heavily　contaminated　training　data.　©　2024　Elsevier　Ltd