Quality　assessment　of　electrocardiogram　(ECG)　signals　holds　significant　importance　in　reducing　the　high-frequency　false　alarms　in　automated　arrhythmia　diagnosis.　Existing　quality　assessment　methods　primarily　rely　on　empirical　threshold　setting　or　the　calculation　of　relevant　statistical　values　to　formulate　decision　rules.　However,　the　thresholds　set　in　this　manner　are　typically　based　on　experience　and　knowledge,　resulting　in　decision　rules　that　lack　flexibility　and　struggle　to　adapt　to　changes　in　the　dataset.　To　address　this　issue,　many　researchers　have　turned　to　deep　learning　algorithms　for　ECG　signal　quality　assessment.　Nevertheless,　most　existing　deep　learning　methods　usually　require　large-scale　and　labeled　ECG　data　for　training,　whose　acquiring　is　typically　challenging.　Therefore,　this　paper　proposes　an　efficient　method　for　ECG　signal　quality　assessment.　First,　we　use　a　contrast　learning　framework　to　construct　sample　pairs　through　multilevel　data　augmentation.　Second,　a　densely　connected　convolutional　network　(DenseNet)　is　used　to　improve　the　model＇s　ability　of　extracting　local　features　in　ECG　signals,　and　a　Squeeze-and-Excitation　attention　module　is　introduced　to　make　the　model　pay　more　attention　to　critical　features.　Finally,　the　feature　representation　is　learned　by　minimizing　the　contextual　contrast　loss　to　bring　similar　samples　closer　and　push　dissimilar　samples　farther　away.　In　this　way,　we　can　utilize　the　contrastive　learning　framework　to　extract　temporal　representations　from　unlabeled　data　and　transfer　them　to　downstream　ECG　signal　quality　assessment　tasks.　Extensive　experiments　on　the　Physionet/CinC　Challenge　2011　dataset　demonstrate　that　our　approach　outperforms　most　existing　quality　assessment　methods.　©　2023　IEEE.