Flotation pulp phase bubble images are captured from the pulp solution within the flotation cell and exhibit characteristics such as quantity and size that are closely related to flotation operating conditions. In this study, two specialized devices were employed to obtain bubble images from the flotation pulp solution: an underwater camera (UC) for in-situ collection of pulp phase bubbles and a bubble diameter analyzer (BDA) for ex-situ collection. The advantages and disadvantages of traditional thresholding segmentation and various deep learning models for segmenting pulp phase bubbles were compared. A segmentation model specifically designed for segmenting flotation pulp phase bubbles, named KSMPB-Net, was proposed based on the Kolmogorov–Arnold Networks (KAN) structure. The training results of the model demonstrated that, compared to the U-Net model, the KSMPB model achieved a 10.7% improvement in segmentation accuracy (MIoU), reaching an MIoU of 0.9985, while reducing the training loss by 65% to a value of 0.011. Feature analysis results further revealed that the quantity and size of flotation pulp phase bubbles were closely associated with air flow rate, impeller speed, and pulp concentration. Moreover, under laboratory experimental conditions, the in-situ image collection method outperformed the ex-situ image collection method. Applying deep learning-based methods to flotation bubble analysis facilitates the optimization of flotation conditions based on bubble characteristics, thereby advancing the integration of pulp phase bubble analysis into the intelligent control processes of industrial flotation operations.

中文翻译：

---

基于 KSMPB-Net 的浮选矿浆相气泡特性智能提取及过程控制

浮选纸浆相气泡图像是从浮选池内的纸浆溶液中捕获的，并表现出与浮选作条件密切相关的数量和尺寸等特性。在这项研究中，采用了两种专用设备从浮选浆溶液中获取气泡图像：用于原位收集纸浆相气泡的水下相机 （UC） 和用于非原位收集的气泡直径分析仪 （BDA）。比较了传统阈值分割和各种用于分割纸浆相气泡的深度学习模型的优缺点。基于 Kolmogorov-Arnold Networks （KAN） 结构，提出了一种专门用于分割浮选纸浆相气泡的分割模型，名为 KSMPB-Net。该模型的训练结果表明，与 U-Net 模型相比，KSMPB 模型的分割精度 （MIoU） 提高了 10.7%，达到 0.9985 的 MIoU，同时将训练损失降低了 65% 至 0.011 的值。特征分析结果进一步表明，浮选矿浆相气泡的数量和大小与空气流速、叶轮转速和矿浆浓度密切相关。此外，在实验室实验条件下，原位图像采集方法优于非原位图像采集方法。将基于深度学习的方法应用于浮选气泡分析有助于根据气泡特性优化浮选条件，从而推动将纸浆相气泡分析集成到工业浮选作的智能控制过程中。