Learning Granger Structure over Time and System State from Non-stationary Event Sequences

Event logs from systems such as semiconductor fabrication lines exhibit non-stationarity in both background activity and causal dependence, including changes in event dependencies and their influence strength. We show that segmentation- and regime-based approaches can absorb within-segment background drift into influence estimates, leading to spurious edges. We introduce CausalField, a point-process framework that represents causal influence as a continuous field $C(t,z)$ over time and system state. The framework parameterizes this field with a coordinate-based neural network and learns it directly from the point-process likelihood, allowing excitation to vary with system state while modeling background drift separately. Under the stated conditions, the population field identifies a direct state-conditional Granger graph on the visited $(t,z)$ region, together with diagnostics for assessing whether recovery is supported by the data. On synthetic data, CausalField recovers the influence field and reproduces the predicted failure mode of discretized alternatives. On a real alarm log, it remains competitive for thresholded static-graph recovery, while state conditioning improves held-out fit beyond static and time-only variants.