Working state detection of microporous aerators in aquaculture based on improved YOLOv11

Microporous aeration constitutes a fundamental oxygenation technology in aquaculture, playing a critical role in determining dissolved oxygen concentrations in culture water and thereby influencing the survival rates of cultivated species. Existing monitoring techniques in aquaculture, notably manual observation and indirect sensor measurements, suffer from inherent drawbacks including delayed response, frequent inaccuracies, and the failure to provide direct insight into aeration performance. Although computer vision offers a viable modern alternative, the accuracy of bubble detection is greatly compromised by challenging environmental conditions such as surface ripples, glare, and non-ideal camera perspectives. To overcome the limitations of existing monitoring techniques, this paper focuses on creating a robust vision-based detection system capable of real-time, high-precision assessment of microporous aeration in complex settings. This capability is critical for achieving data-driven, intelligent operation and maintenance of aquaculture oxygenation infrastructure. Based on the YOLOv11 framework, an enhanced detection model designated as YOLOv11-SPPF-LSKA-LSCSBD was developed. The improvement involved two key structural modifications: first, the introduction of an SPPF-LSKA composite module that integrates multi-scale feature aggregation with lightweight attention guidance to enhance the capture of dynamic, low-contrast bubble features and suppress background interference; second, the incorporation of a Lightweight Shared Convolution–Separate Batch Normalization Detection Head (LSCSBD), which employs a “shared convolution, independent normalization” design to improve multi-scale feature fusion efficiency while reducing parameter complexity. Additionally, a dedicated dataset comprising 786 images was constructed, capturing bubble targets under varied lighting conditions, viewing angles, and distances. The dataset was partitioned into training, validation, and test subsets in a 7:2:1 ratio. Ablation studies demonstrated that the proposed YOLOv11-SPPF-LSKA-LSCSBD model outperformed the baseline YOLOv11, achieving a precision of 96.8% (+4.7 pp), a recall of 98.6% (+10.0 pp), and an mAP@0.5 of 98.4% (+7.2 pp), with a computational cost of 6.4 GFLOPs and 2.73 M parameters. In comparative experiments against mainstream models including YOLOv6, YOLOv8, and YOLOv10n, the proposed model exhibited superior overall detection performance. Further evaluation under challenging aquaculture-specific conditions confirmed its robustness. Under water ripple interference, the model attained an F1-score of 97.7%, surpassing the baseline YOLOv11 and the YOLOv11-LSCSBD variant by 7.35 and 1.35 percentage points, respectively. In scenarios with strong light reflection, it achieved a precision of 92.9%, representing a 4.2 percentage-point improvement over YOLOv11. When detecting bubbles from oblique viewing angles (30°–60°), the model maintained high precision (91.5%) and recall (89.2%), exceeding the baseline by 9.1 and 8.7 percentage points, respectively. The proposed YOLOv11-SPPF-LSKA-LSCSBD model effectively addresses the limitations imposed by complex aquaculture environments on aeration state detection, enabling accurate and real-time monitoring of microporous aeration systems. This approach provides a direct and reliable technical solution for the intelligent monitoring of aquaculture oxygenation, serving as a significant advancement over conventional indirect methods. Future work should focus on expanding the dataset to include a wider variety of water types and bubble sizes, optimizing the model for deployment on embedded devices, and extending its application to other aeration systems such as impeller and paddlewheel aerators.