Abstract: Coking at the fractionating tower bottom and the decant oil circulation system disrupts the heat balance, leading to unplanned shutdown and destroying the long period stable operation of the Fluid Catalytic Cracking Unit (FCCU). The FCCU operates through interconnected subsystems, generating high-dimensional, nonlinear, and non-stationary data characterized by spatiotemporally correlated. The decant oil solid content is the crucial indicator for monitoring catalyst loss from the reactor-regenerator system and coking risk tendency at the fractionating tower bottom that relies on sampling and laboratory testing, which is lagging responsiveness and labor-intensive. Developing the online decant oil solid content soft sensor using industrial data to support operators in conducting predictive maintenance is essential. Therefore, this paper proposes a hybrid deep learning framework for soft sensor development that combines spatiotemporal pattern extraction with interpretability, enabling accurate risk identification in dynamic operational conditions. This framework employs a Filter-Wrapper method for dimensionality reduction, followed by a 2D Convolutional Neural Network (2DCNN) for extracting spatial patterns, and a Bidirectional Gated Recurrent Unit (BiGRU) for capturing long-term temporal dependencies, with an Attention Mechanism (AM) to highlight critical features adaptively. The integration of SHapley Additive exPlanations (SHAP), Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), 2DCNN, and expert knowledge precisely quantifies feature contributions and decomposes signals, significantly enhancing the practicality of risk identification. Applied to a China refinery with processing capacity of 2.80 × 106 t/a, the soft sensor achieved the R2 value of 0.93 and five-level risk identification accuracy of 96.42%. These results demonstrate the framework's accuracy, robustness, and suitability for complex industrial scenarios, advancing risk visualization and management.