To address the complexity of multivariate time series data in the chemical industry and privacy protection requirements, TAGSFA (temporal attention gated source-free adaptation algorithm was proposed, significantly improving the performance of crossdomain fault diagnosis. The temporal correlations between source and target domains were effectively captured and negative transfer effects were avoided through the integration of gated Transformer networks, embedded discretization techniques, and neighborhood clustering strategies. The experimental results on both TEP (Tennessee Eastman process) dataset and CSTR (continuous stirred tank reactor) dataset have demonstrated that TAGSFA outperforms existing methods in diagnostic accuracy and robustness without source domain data access. The critical roles of time-domain attention and neighborhood clustering have been validated through ablation studies.