A Novel FULS‐TSVM‐Based Model for Non‐Destructive Detection of Heavy Metals in Mussels via Near‐Infrared Spectroscopy

ABSTRACT With the rising global demand for seafood, mussels have become an important commercial product. They are located in estuarine environments and are prone to heavy metal accumulation. This poses a threat to the safety of mussel products for human consumption. Traditional detection methods are accurate, but limited by complex sample pretreatment, high costs, and slow processing speed. Near‐infrared spectroscopy (NIRS) can achieve non‐destructive and rapid detection, but the detection results are easily disrupted by imbalanced sample datasets, individual outliers, and low‐concentration contamination in samples. To solve these issues, this study first applied a fuzzy universum least squares twin support vector machine (FULS‐TSVM) algorithm to the detection of heavy metals in mussels. By integrating universum data to incorporate prior distribution knowledge and assigning fuzzy memberships to handle outlier samples and class imbalance, this model effectively extracts weak spectral differences caused by heavy metals. This study focused on mussels contaminated by cadmium (Cd). We conducted comparative experiments in imbalanced datasets, with outlier samples, and under high and low contamination concentrations. We also expanded the test objects to lead (Pb) and zinc (Zn) contaminated mussels to verify the model's generalization ability. The results show that compared to LS‐TSVM, FLS‐TSVM, and ULS‐TSVM models, the FULS‐TSVM has overall better detection accuracy and robustness. It maintains reliable detection performance even in extreme imbalanced datasets, numerous outliers, and low‐concentration contamination scenarios. This research proves that combining NIRS technology with the FULS‐TSVM algorithm provides an efficient, low‐cost way to quickly screen heavy metal contamination in mussels. The method has great application value for daily safety monitoring of marine aquatic food.