The degradation of a system's performance poses a significant challenge to the effective application of fault diagnosis methods for dynamic systems. Consequently, the underlying feature distribution changes over time during the actual process, resulting in a decline in the effectiveness of existing diagnosis methods. In this article, we present a real-time adaptive fault diagnosis scheme to address this issue. A latent variable-guided broad learning system (LVGBLS) is proposed to construct the fundamental diagnosis model, which effectively extracts dynamic features from the monitored data. An incremental update procedure is then designed based on pseudolabel learning to adapt to dynamic process changes while minimizing labeling costs. We also introduce the condition detection mechanism (CDM) to detect dynamic changes under the degradation process based on statistical information. To demonstrate the effectiveness of our proposed method, we conduct several comparison experiments and ablation experiments on electrical drive systems and XJTU-SY bearing datasets. The results show that our proposed scheme exhibits superior performance with low labeling costs in most scenarios with performance degradation.