Investigating process monitoring techniques is required to reduce the loss of property and life caused by industrial processes accidents. Fault diagnosis, which attempts to determine the fault type, is a vital step in process monitoring because it can help operators respond to abnormal situations appropriately. Adequate data labels to train supervised fault diagnosis models are difficult to acquire in practice; however, semisupervised methods, which are attracting increasing attention, can use unlabeled data. Self-labeled algorithms are an effective paradigm of semisupervised methods, but their applications in industrial process fault diagnosis do not meet expectations, because they are prone to performance deterioration when handling industrial process data. To address this issue, in this article, a self-training algorithm with a modified confidence measure is proposed. The confidence measure is temporal-spatial with temporal identities of data introduced to its definition and calculation, which makes the algorithm adaptable to industrial processes. The proposed algorithm is also self-adaptive to avoid time-consuming hyperparameter tuning processes. The benchmark Tennessee Eastman process data were used to evaluate the proposed algorithm, and the experiment results demonstrate its superiority compared to competing semisupervised methods.