Pressure vessel gas leakage intelligent detection and identification techniques are susceptible to interference from a variety of factors, and intelligent detection models require a large amount of monitoring data training. In the actual industrial environment, available data, especially data labels, are very scarce. To address the problems such as interference from multiple working conditions and lack of labeling information of data, this article proposes an unsupervised variable working condition intelligent detection technique by using transfer learning. Firstly, samples of multiple leaks are collected in laboratory environment and select three different pressure working conditions to divide the data into labeled source domain and unlabeled target domain. Secondly, a convolutional feature extractor is designed to propose an improved joint distribution adaptation mechanism for the edge distribution and conditional distribution of the two domains, and further improve the distribution difference metric to enhance the neighborhood confusion. Experimental results on six transfer learning tasks validate the effectiveness of the method, with higher accuracy than the classical domain adaptive algorithm.