Intelligent Industrial Scene Interpretation via Synergy of Deep Learning and Knowledge Embedding

Abstract

Equipment such as eagle-eye cameras, low-altitude unmanned aerial vehicles, and high-resolution satellites can capture high-precision industrial scene images in real time, providing a data foundation for production, ecological, and safety supervision. However, the massive volume of data imposes a significant burden on manual interpretation, creating an urgent need to develop intelligent interpretation methods. Intelligent interpretation of industrial scenes requires algorithms to not only accomplish pixel-level entity recognition but also achieve scene-level compliance reasoning and business-level cause-oriented attribution. This paper proposes the Knowledge-Embedding-Driven Intelligent Interpretation Framework (KEDIF) for industrial scenes, with core innovations at two levels. At the architectural level, we propose a three-tier interpretation framework grounded in the Mining-Damaged Land Industrial Domain Knowledge Graph (MDKG), threading unified knowledge embedding vectors through the complete perception-reasoning-attribution pipeline. The first tier fuses knowledge graph embeddings to construct a comprehensive entity-type interpretation score; the second tier introduces the Knowledge-Embedding Compliance Score (KECS), a multi-regulation aggregation mechanism that provides graded risk assessments and generalizes to unseen entity-pair configurations through continuous embedding space interpolation; and the third tier employs a two-component cause-oriented attribution model to drive structured business-level output with regulatory traceability. At the model level, we propose ACDR-Former, a dynamic regulation segmentation model that establishes a dynamic regulation paradigm for query-based segmentation through three synergistic modules: Adaptive Query Evolution (AQE), Confusion-Aware Assignment (CAA), and Targeted Representation Disentanglement (TRD). Experiments on the self-constructed MDID benchmark dataset demonstrate that the complete system achieves an mIoU of 50.83%±0.37%, a cumulative improvement of 4.35 percentage points over the Mask2Former baseline (paired t-test, p= 0.006, Cohen’s d= 7.41); the KECS compliance alert F1 reaches 85.6%±1.3% (+ 18.3 pp over the hard-rule baseline); and the cause-oriented attribution accuracy reaches 91.7%±1.0% (+ 13.5 pp). Generalization experiments on LoveDA and UAVid further validate the transferability of the dynamic regulation paradigm.