Interpretable flatness anomaly monitoring and potential transmission path tracing based on a dynamic graph temporal convolution-informer model

Cold-rolled steel strips are widely used in automotive manufacturing, energy equipment, and other high-end industrial fields, where flatness is a critical indicator of product quality and service reliability. However, the cold rolling process is characterised by high speed, strong nonlinearity, and complex multivariable coupling, which makes flatness anomaly monitoring and potential association-path analysis challenging. To address these issues, a dynamic graph temporal convolution-informer (DGTC-I) model is proposed for flatness anomaly monitoring in tandem cold rolling. The model consists of a graph temporal embedding (GTE) module and an Informer-based forecasting module. In the GTE module, temporal convolution is used to extract local temporal features from process variables, while a process-knowledge-guided dynamic graph generator constructs input-conditioned adjacency matrices for different observation windows. Dynamic graph convolution is then performed to aggregate information among related process variables. The resulting graph-temporal representations are fed into the Informer module to model broader temporal dependencies across the observation window and generate multistep predictions. Experiments on real cold rolling production data show that DGTC-I achieves a monitoring rate of 99.82% and an R-score of 97.32%, achieving superior overall performance compared with representative statistical, temporal, and graph-based anomaly monitoring approaches. Moreover, by analysing the dynamic adjacency matrices generated under different anomalous operating conditions, variables strongly associated with flatness anomalies can be identified, and model-inferred potential transmission paths can be visualised. The proposed method provides an effective and interpretable framework for intelligent flatness quality monitoring and analysis in cold rolling.