Abstract: To address the problem that missing data in industrial processes disrupts variable correlation structures and temporal dependencies, thereby degrading early fault monitoring performance, this paper proposes a CSAT-ICVDA method for early fault monitoring of industrial processes under missing-data conditions. First, a Convolutional Sparse Attention Transformer （CSAT） data imputation model is constructed. Based on the traditional Transformer architecture, convolutional feature enhancement and Top-k sparse attention mechanisms are integrated to achieve collaborative modeling of local dynamic features and key temporal dependencies. Meanwhile, an improved masking mechanism and a two-stage imputation–reconstruction structure are introduced to enhance missing data recovery capability. Second, an improved canonical variate dissimilarity analysis （ICVDA） model is developed by incorporating ensemble learning. Multiple CVDA detectors are fused, and singular value features are used to construct a new monitoring index, thereby improving the detection sensitivity and stability for incipient faults. Finally, experiments are conducted on penicillin fermentation and Escherichia coli fermentation processes. The results demonstrate that the proposed CSAT-ICVDA method can effectively recover process features under missing-data conditions and accurately monitor early faults in industrial processes, showing a high fault detection rate and promising engineering applicability.