In response to the characteristics of process data such as nonlinear coupling between variables and diverse process scales, this paper proposes an adaptive industrial process fault detection method based on multi-scale moving window kernel entropy component analysis (MS-MWKECA). First, the wavelet transform is used to decompose process variables into different detail scales and approximate scales in real-time. Then, KECA-based local scale models are established, and scales with salient features are selected for data reconstruction. Subsequently, an online detection model based on the moving window KECA is developed for the reconstructed data, and an adaptive weight distribution moving window fusion strategy is designed to update data within the moving window process. Finally, the simulation tests are conducted in the Tennessee-Eastman (TE) process and the hot rolling production process (HSMP), and the effectiveness and superiority of the method proposed are verified through experimental comparison.