Abstract: To address the problems of fixed neighborhood parameters in existing visual models that cannot adapt to the drastic changes in flotation froth density, and the decline in recognition accuracy caused by strong mechanical vibration in industrial fields, this study proposes and validates a three-dimensional point cloud multi-scale adaptive perception system for complex working conditions (DGCNN-IO). First, an industrial-grade dataset containing 3,577 high-precision samples was constructed based on an automatic slide-rail inspection platform, and 10-dimensional enhanced geometric features including curvature and roughness were introduced to enrich the physical semantics of point clouds. Second, a neighborhood scale adaptive mechanism based on reinforcement learning was proposed to dynamically adjust the graph convolution neighborhood scale (k value) according to the local density of point clouds, effectively solving the perceptual scale mismatch problem when large-scale merged bubbles and densely stacked froth coexist. Finally, a joint training strategy of knowledge distillation and consistency regularization was introduced to transfer the discriminative experience of the teacher model to the inference network, improving the robustness of the model in harsh environments. Experimental results show that the system achieves a comprehensive classification accuracy of 97.0% and an F1 score of 96.9% on a field-collected dataset containing typical working conditions such as excess and deficient frother; under a Gaussian noise environment simulating strong mechanical vibration, the recognition performance drops by only 1.1%, significantly outperforming the baseline model, verifying the effectiveness and engineering practical value of the proposed method.