A Method for Melamine Identification Based on A One-Dimensional Very Deep Convolutional Neural Network-Bidirectional Long Short-Term Memory Network Fused with An Attention Mechanism

To address the critical safety issue of melamine contamination in dairy products, a novel classification model that integrated one-dimensional very deep convolutional neural network (1D-VDCNN) with bidirectional long short-term memory and attention mechanism (BiLSTM-Attention) was developed for nondestructive and efficient identification of melamine. Traditional detection methods suffered from low efficiency and operational complexity. In the proposed 1D-VDCNN module, two constraints were introduced to optimize the feature layers, while downsampling and large convolutional kernels replaced fully connected layers to reduce computational complexity and enhance vertical feature extraction. Subsequently, the BiLSTM network was employed to capture bidirectional long-range dependencies within the feature sequences, thereby strengthening the relationships of horizontal feature. Finally, an attention mechanism was integrated to assign higher weights to critical spectral features. Experiments were conducted using an open-source near-infrared spectral dataset of melamine, comprising 1972 samples with a maximum of 500 samples per class, covering a spectral range of 5546‒6254 cm–1. The results demonstrated that the proposed model achieved a classification accuracy of 99.75%, with a parameter reduction of approximately 43% compared to the baseline model. Also, the model exhibited significantly improved convergence speed and feature extraction capability, along with robust stability and generalization across different sample sets. Compared to a standalone BiLSTM model and three traditional chemometric methods, the accuracy improvement could reach up to approximately 10%. This model was well-suited for small-sample spectral data and offered a high-accuracy, lightweight chemometric approach for food safety detection.