Multivariate time series classification plays a key role in industrial state identification and decision support, but irregular data formed by asynchronous sampling and random missingness severely degrades its classification accuracy. Imputation-based methods are prone to introducing noise and artifacts, whereas direct modeling tends to dilute or distort critical information. This paper proposes a unified “sequence-image-graph” modeling method based on an enhanced multi-scale graph Transformer. First, enhanced multi-channel image Transformation encodes irregular sequences into imputation-free RGB images, directly leveraging value dynamics, missingness patterns, and sampling information. Second, dynamic sector graph construction maps image patches to graph nodes and adaptively builds adjacency edges within local sector-based neighborhoods to cover cross-time and cross-variable dependencies while reducing redundancy. Finally, adaptive multi-scale relative graph convolution introduces multi-scale difference features into neighbor aggregation and combines attention weighting to emphasize critical neighbors, thereby alleviating over-smoothing in graph convolution. Experiments on a blast furnace gas dataset from a steel enterprise and four public irregular benchmarks validate the effectiveness of the proposed method under diverse irregular patterns.