Melamine identification method of 1D-VDCNN-BiLSTM fused with attention characteristics

Abstract

The invention relates to a melamine identification method of 1D-VDCNN-BiLSTM fused with attention characteristics, and belongs to the field of identification of melamine in food. The method comprises the following four stages: 1) sample data division and preprocessing: completing preprocessing of sampling data: preprocessing a sample set by adopting a KS method; (2) constructing a one-dimensional ultra-deep convolutional neural network (1D-VD-CNN): improving a two-dimensional CNN model into the 1D-CNN for melamine one-dimensional data, and then adding an ultra-deep VD structure to construct the one-dimensional ultra-deep convolutional neural network (1D-VD-CNN); 3) structural design of a bidirectional long short-term memory network (BiLSTM); and 4) constructing a melamine identification model fused with the attention mechanism, and training, verifying and testing the model. Wherein the data classification is changed into layered sampling from traditional pure random sampling. By establishing a novel melamine identification method with high precision and strong interpretability, intelligent technical support is provided for food safety risk prevention and control.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of melamine identification in food, and is a practical application of deep learning for food melamine identification, which meets the research needs of actual food identification and has important application prospects. Specifically, it relates to a melamine identification method based on 1D-VDCNN-BiLSTM fusion attention features. BACKGROUND

[0002] Melamine is a nitrogen-containing organic compound widely used in industries such as plastics, coatings, and textiles. However, due to its high nitrogen content, the intake of melamine can cause serious harm to human health, especially for infants, which may lead to kidney stones, kidney failure, and even death. Therefore, developing a fast, accurate, and low-cost melamine detection method has important practical significance.

[0003] Illegal addition of melamine not only threatens public health, but also undermines the trust system of the food market. Traditional melamine detection methods mainly include high-performance liquid chromatography (HPLC), gas chromatography-mass spectrometry (GC-MS), and partial least squares discriminant analysis (PLS-DA). These methods have high accuracy and sensitivity, but also have obvious limitations:

[0004] (1) Expensive equipment: requires high-precision instruments, which are costly.

[0005] (2) Complex operation: requires professional personnel for sample pretreatment and instrument operation.

[0006] (3) Time-consuming: from sample processing to result analysis requires a long time, making it difficult to meet the needs of on-site rapid detection.

[0007] (4) Inaccessible: due to cost and technical limitations, it is difficult to promote and apply in grassroots or remote areas. Therefore, developing a fast, low-cost, and easy-to-operate melamine detection method has become an important research direction in the field of food safety.

[0008] Spectral analysis techniques such as near-infrared spectroscopy and Raman spectroscopy have been widely used in food safety detection due to their speed, non-destructiveness, and lack of complex sample pretreatment. Spectral data is usually presented in the form of one-dimensional signals, which can reflect the molecular structure and chemical composition information of substances. However, spectral data often has high dimensionality, nonlinearity, and noise interference, and traditional chemometric methods such as partial least squares (PLS) have certain limitations in processing complex spectral data.

[0009] In recent years, deep learning techniques have made breakthroughs in data feature extraction, image recognition, and object detection. Convolutional Neural Networks (CNN) as a representative model of deep learning, has strong feature extraction and classification ability. Compared with traditional machine learning methods, deep learning techniques can automatically learn features from raw data, avoiding the tedious process of manually designing features, while having higher generalization ability and robustness. One-dimensional Convolutional Neural Networks (1D-CNN) is a variant of CNN, specifically designed for processing one-dimensional signal data (such as spectra, time series, etc.). 1D-CNN extracts features from one-dimensional signals through convolutional layers and pooling layers, and realizes classification or regression tasks through fully connected layers. The melamine detection method based on 1D-CNN can directly process raw spectral data, avoiding complex preprocessing steps, while having high detection accuracy and efficiency.

[0010] Currently, some research has applied deep learning techniques to food safety detection. Convolutional Neural Networks (CNN) extract spatial features through local receptive fields. The literature "Improved 1D convolutional neural network adapted to near-infrared spectroscopy for rapid discrimination of Anoectochilus roxburghii and its counterfeits" designs an algorithm based on an improved one-dimensional convolutional neural network (1D-CNN) to identify gold orchid and its similar objects, with an accuracy of about 98%; the literature "Behavior recognition based on spatiotemporal attention LSTM" uses LSTM to analyze time-resolved spectra, improving 0.66% on UCF101 and 0.75% on HMDB51; the literature "A prediction approach to COVID-19 time series with LSTM integrated attention mechanism and transfer learning" applies an attention mechanism to COVID-19 time series prediction, and the results show that RNN and GRU have improved performance after integrating attention mechanism and transfer learning. However, spectral data of melamine (such as Raman or near-infrared spectra) presents typical one-dimensional sequence characteristics, containing local feature peaks and global band correlation. Traditional CNN models are good at extracting local features, but have difficulty in capturing long-range dependencies; BiLSTM has advantages in time series modeling, but lacks sensitivity to local details; most model decision-making processes lack physical interpretation, restricting practical application, and lack of explainability. SUMMARY

[0011] The object of the present application is to overcome the deficiencies of the existing conventional technology, improve the low efficiency and complex operation of the traditional detection method, and provide a melamine identification method based on 1D-VDCNN-BiLSTM fusion attention feature, which can quickly and accurately identify different formulations of melamine by establishing a multi-modal model based on near-infrared spectrum data and applying it to the identification of melamine in food ingredients.

[0012] To achieve the above object, the technical scheme of the present application is: a melamine identification method based on 1D-VDCNN-BiLSTM fusion attention feature, a 1D-VDCNN-BiLSTM hybrid architecture with fusion attention feature is proposed, which is applied to the identification of melamine with different formulations; first, the VD structure is introduced, and a one-dimensional ultra-deep convolutional neural network model 1D-VDCNN is designed based on near-infrared spectrum NIRS data, second, a feature dynamic change model under different concentrations is established through the memory gate mechanism to construct a bidirectional long short-term memory network BiLSTM; finally, the attention mechanism is fused, the SE module is used to quantify the importance of each channel, the weight of the key feature peak of melamine is improved, and the construction of the 1D-VDCNN-BiLSTM hybrid architecture with fusion attention feature, i.e. the melamine identification model, is completed.

[0013] Further, the one-dimensional ultra-deep convolutional neural network model 1D-VDCNN is improved from a two-dimensional CNN model to a 1D-CNN for one-dimensional data of melamine, and then a super-deep VD structure is added.

[0014] Further, the NIRS data is preprocessed by the standard deviation normalization algorithm.

[0015] Further, the NIRS data is preprocessed by the standard deviation normalization algorithm, and the specific implementation manner is:

[0016] The obtained NIRS data is processed by zero mean and variance normalization of each feature dimension, so that the data conforms to the standard normal distribution, the mean is 0, and the standard deviation is 1; the transformation function is represented as:

[0017]

[0018] In the formula, μ is the mean of all NIRS data, and σ is the standard deviation of all NIRS data.

[0019] Further, the sample set obtained after preprocessing the NIRS data is divided by the KS algorithm.

[0020] Further, the KS algorithm is adopted to realize the sample set division, and the specific implementation manner is as follows: the selection of the correction set is performed by calculating the average spectrum of the data in the sample set and the Euclidean distance between each spectrum, the data with large spectrum difference is selected into the correction set, so that the model established can have representativeness, and the uneven distribution of the modeling set data is avoided, and the correction set and the modeling set are divided, wherein the modeling set is used for training the model.

[0021] Further, the method, the specific implementation steps are as follows:

[0022] Step one, constructing a one-dimensional super-depth convolutional neural network model 1D-VDCNN, that is, improving the two-dimensional CNN model to 1D-CNN for melamine one-dimensional data, and then adding a super-depth VD structure;

[0023] Step two, designing a bidirectional long short-term memory network BiLSTM, the BiLSTM core includes two independent LSTM processing modules: a time sequence forward processing module and a time sequence reverse processing module; in the data processing process, the forward module analyzes the input sequence in the natural order of the time axis, realizes the progressive accumulation of historical features; the reverse module processes the input data in the reverse order of the time axis, and is used for extracting subsequent features of the sequence; the two complementary processing modules finally integrate their feature outputs to a unified classification layer, and at the same time, the two modules maintain complete independence in the feature extraction stage, and there is no cross transmission of internal state information in any form;

[0024] The forward and reverse calculation formulas of BiLSTM are as follows:

[0025]

[0026] h i is the result obtained at the current time of the hidden state in the LSTM network; h i-1 is the result obtained at the previous time step of the hidden state in the LSTM network; f1 and f2 represent the sigmoid activation function; A and B represent the input weight matrix of the LSTM network at the current and previous time steps respectively; c represents the bias value of the LSTM network.

[0027] The final output of each time step i is:

[0028]

[0029] represents a splicing operation;

[0030] Step three, constructing a melamine identification model fused with an attention mechanism, and the implementation steps of the attention mechanism are as follows:

[0031] (1) Calculate weights: for a given query q and a set of keys k, calculate the similarity or relevance score d between them;

[0032] (2) Normalization: normalize the score d using the softmax function to get the weight a corresponding to each key;

[0033] (3) Weighted sum: according to the weight, the corresponding value a is weighted and summed to get the final attention output;

[0034] In combination with the 1D-VDCNN constructed in step one, the BiLSTM constructed in step two and the attention mechanism, a 1D-VDCNN-BiLSTM hybrid architecture is established to realize melamine identification.

[0035] Further, the core technology of the VD structure design includes feature layer optimization and classification layer optimization, the hidden layer of the 1D-CNN is divided into two parts: the feature layer and the classification layer, and the feature layer optimization and the classification layer optimization are respectively optimizing the feature layer and the classification layer.

[0036] The application also provides a melamine identification system of the 1D-VDCNN-BiLSTM with attention feature fusion, which comprises a memory, a processor and computer program instructions stored in the memory and capable of being executed by the processor, and when the processor executes the computer program instructions, the steps of the method described in any of the above embodiments can be realized.

[0037] The application also provides a computer readable storage medium having computer program instructions capable of being executed by a processor stored thereon, and when the processor executes the computer program instructions, the steps of the method described in any of the above embodiments can be realized.

[0038] Compared with the prior art, the application has the following beneficial effects: the application deeply fuses the representation learning ability of deep learning and the physical and chemical principles of spectral analysis, extracts multi-modal features in cooperation and selects features guided by attention, establishes a new melamine identification method with high precision and strong explanation, and provides intelligent technical support for food safety risk prevention and control. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 It is the 1D-VDCNN model of the application.

[0040] Figure 2 It is the bidirectional long short-term memory network (BiLSTM) network structure diagram of the application.

[0041] Figure 3 It is the attention mechanism structure diagram of the application. DETAILED DESCRIPTION

[0042] The technical solutions of the present application will be specifically described below with reference to the drawings.

[0043] The present application proposes a melamine identification method based on 1D-VDCNN-BiLSTM fusion attention features, proposes a 1D-VDCNN-BiLSTM hybrid architecture that fuses attention features, and applies it to melamine identification of different formulations. First, the VD structure is introduced, and a one-dimensional ultra-deep convolutional neural network model 1D-VDCNN is designed based on near-infrared spectrum NIRS data. Second, a feature dynamic change model under different concentrations is established through a memory gate mechanism to construct a bidirectional long short-term memory network BiLSTM. Finally, the attention mechanism is fused, and the SE module is used to quantify the importance of each channel, so that the weight of the key feature peak of melamine is improved, and the construction of the 1D-VDCNN-BiLSTM hybrid architecture that fuses attention features, that is, the melamine identification model, is completed.

[0044] The following is the specific implementation process of the present application.

[0045] The present application is to realize the rapid identification of melamine without formula, to take melamine NIRS data as the research object, to design a multi-modal network model, and to study a new method for rapid and efficient identification of melamine of different formulations.

[0046] The specific implementation steps are as follows:

[0047] Step 1, acquisition and division of sample data

[0048] The open source melamine sample data set (https: / / github.com / RNL1 / Melamine-Dataset) is used in this experiment. The melamine data set of this experiment contains Fourier transform near-infrared (FT-NIR) absorbance spectra, and the sample data is obtained by recording the formaldehyde melamine cloud point reading during the polycondensation reaction (i.e. polymerization process), and the data is collected from Metadynea GmbH company in Krems, Austria. Among them, the near-infrared spectrum is collected online through an optical fiber probe, and the cloud point data is obtained by offline determination of the cloud point temperature (i.e. the temperature when the condensate begins to become turbid). The cloud point is an important indicator for measuring the length of polymer chain and the degree of cross-linking (i.e. degree of polymerization), which determines the physical and chemical properties of the final resin. The data set information contains multiple batch records of 4 different formulations (raw material ratio and different additives), as shown in Table 1.

[0049] Table 1 Sample data information

[0050] Number Sample Distinctive Points Sample Number Pure Melamine R562 Melamine + Milk (Low Concentration) 500 1～500 R568 Melamine + Milk Powder (High Concentration) 500 501～1000 R861 Melamine + Other Additives 500 1001～1500 R862 Figure 1 472 15011～1972

[0051] Step 2, sample data preprocessing

[0052] In the process of neural network training, the weight parameter convergence can be accelerated by data standardization. The standard deviation standardization method (standard scaler) is used for data preprocessing, and the data is normalized to zero mean and variance, so that the data conforms to the standard normal distribution, the mean is 0, and the standard deviation is 1. The transformation function can be expressed as:

[0053]

[0054] In the formula, μ is the mean of all sample data, and σ is the standard deviation of all sample data.

[0055] Step 3, construction of one-dimensional super-depth convolutional neural network (1D-VDCNN)

[0056] The application improves the traditional one-dimensional convolutional neural network into a one-dimensional super-depth convolutional neural network (1D-VD-CNN) model. The specific design idea is as follows: the traditional two-dimensional CNN model is adaptively improved for melamine one-dimensional data into 1D-CNN, and then the efficient super-depth (VD) structure is added. The core technology of the super-depth structure design includes feature layer optimization and classification layer optimization.

[0057] Step 1: The hidden layer of the traditional 1D-CNN is divided into two parts--feature layer and classification layer.

[0058] Step 2: Feature layer optimization design. The feature layer optimization design is converted into a constraint optimization design. It mainly includes two constraint condition designs: learning ability and learning necessity.

[0059] (1) Learning ability design (first constraint condition)

[0060] In order to quantify the learning ability, the parameter C is introduced:

[0061]

[0062] Where, n conc represents the size of the convolution kernel, n field represents the size of the receptive field. If C is too small, the feature extraction is insufficient. In the 1D-VDCNN model, the receptive field is gradually expanded, and the down-sampling interval step is k. After n times of down-sampling, the receptive field is expanded to 2 n ×k. The definite constraint condition of the 1D-VDCNN of the application is that the C value of each convolution layer should be greater than or equal to 1 / 6. Taking 1 / 6 is a best lower limit of C value.

[0063] (2) Learning necessity design (second constraint condition)

[0064] As the network deepens, the receptive field may exceed the length of the input data. At this point, neurons can cover the entire dataset, and further increasing the depth not only fails to improve performance but may also lead to overfitting. Therefore, the second constraint is: the size of the top-level receptive field should equal the length of the input data to avoid redundant computation.

[0065] Step 3: Classification Layer Optimization Design. This invention optimizes the classification layer in the following three aspects:

[0066] (1) Downsampling and compressing features: reducing the feature map size to a smaller scale (e.g., 8×1 or 16×1).

[0067] (2) Replacing fully connected layers used for feature integration with large convolutional kernels: Using two 1×5 convolutional layers and pooling layers with Dropout, the feature maps are downsampled to 1×1.

[0068] (3) Lightweight classification layer: Finally, a fully connected layer is used to output the classification result.

[0069] This scheme utilizes the weighted summation function of large convolutional kernels to replace traditional fully connected layers, significantly reducing the number of parameters and mitigating the risk of overfitting. The design results are as follows: Figure 2 As shown.

[0070] Step 4: Design the Bidirectional Long Short-Term Memory (BiLSTM) network structure

[0071] The specific structure of this invention is as follows: Figure 3 As shown, the core architecture of the BiLSTM network comprises two independently operating LSTM processing modules: a temporal forward processing module and a temporal backward processing module. In the data processing flow, the forward module parses the input sequence following the natural order of the timeline, achieving a gradual accumulation of historical features; the backward module processes the input data in reverse chronological order, specifically for extracting subsequent features of the sequence. These two complementary processing modules ultimately integrate their feature outputs into a unified classification layer. Furthermore, the two modules maintain complete independence during the feature extraction stage, with no cross-transmission of their internal state information.

[0072] The forward and backward calculation formulas for this network model are as follows:

[0073]

[0074] h i This is the result obtained from the hidden state at the current time in the LSTM network; h i-1 f1 and f2 represent the results obtained from the previous time step of the hidden state in the LSTM network; f1 and f2 represent the sigmoid activation functions; A and B represent the input weight matrices of the LSTM network at the current and previous time steps, respectively; c represents the bias value of the LSTM network.

[0075] The final output for each time step i is:

[0076]

[0077] This indicates a concatenation operation, meaning that if the hidden layer dimension of the LSTM is 50, then the dimension of the BiLSTM is 100.

[0078] Step 5: Construction of a melamine identification model incorporating attention mechanisms

[0079] The structure of the fusion attention mechanism of the present invention is as follows: ​ As shown. The implementation steps are as follows:

[0080] (1) Calculate weights: For a given query (q) and a set of keys (k), calculate the similarity or relevance score d between them.

[0081] (2) Normalization: The score d is normalized using the softmax function to obtain the weight a corresponding to each key.

[0082] (3) Weighted summation: The corresponding values ​​(a) are weighted and summed according to the weights to obtain the final attention output.

[0083] Based on the above design ideas, this invention combines the advantages of Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory Networks (BiLSTM), and attention mechanisms to establish a 1D-VDCNN-BiLSTM hybrid architecture that integrates attention features. This model can capture local features and long-range dependencies in sequence data.

[0084] The present invention also provides a melamine identification system based on 1D-VDCNN-BiLSTM with fused attention features, including a memory, a processor, and computer program instructions stored in the memory and executable by the processor. When the processor executes the computer program instructions, it can implement the steps of any of the methods described above.

[0085] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions executable by a processor, wherein when the processor executes the computer program instructions, it can implement the steps of any of the methods described above.

[0086] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A melamine identification method of fusion attention feature 1D-VDCNN-BiLSTM, characterized in that, A 1D-VDCNN-BiLSTM hybrid architecture is proposed to identify different formulations of melamine. First, a one-dimensional ultra-deep convolutional neural network model 1D-VDCNN is designed based on near-infrared spectroscopy (NIRS) data. Second, a bidirectional long short-term memory network (BiLSTM) is constructed by establishing a dynamic feature change model under different concentrations through a memory gate mechanism. Finally, the 1D-VDCNN-BiLSTM hybrid architecture is constructed by fusing the attention mechanism and using the SE module to quantify the importance of each channel, so as to improve the weight of the key feature peaks of melamine. 2.The melamine identification method of fusion attention feature 1D-VDCNN-BiLSTM according to claim 1, wherein, The one-dimensional ultra-deep convolutional neural network model 1D-VDCNN is improved from a two-dimensional CNN model for one-dimensional data of melamine, and then a VD structure is added. 3.The melamine identification method of fusion attention feature 1D-VDCNN-BiLSTM according to claim 1, wherein, The NIRS data is preprocessed by a standard deviation normalization algorithm. 4.The melamine identification method of fusion attention feature 1D-VDCNN-BiLSTM according to claim 1 or 3, characterized in that, The NIRS data is preprocessed by a standard deviation normalization algorithm, and the specific implementation is as follows: The NIRS data is preprocessed by a standard deviation normalization algorithm, and the specific implementation is as follows: The transformation function is represented as: In the formula, μ is the mean of all NIRS data, and σ is the standard deviation of all NIRS data. 5.The melamine identification method of fusion attention feature 1D-VDCNN-BiLSTM according to claim 1, wherein, The sample set obtained after preprocessing the NIRS data is divided into calibration set and modeling set by KS algorithm. 6.The melamine identification method of fusion attention feature 1D-VDCNN-BiLSTM according to claim 5, wherein, The sample set is divided into calibration set and modeling set by KS algorithm, and the specific implementation is as follows: the data with large spectral difference is selected into the calibration set by calculating the average spectrum of the data in the sample set and the Euclidean distance between each spectrum, so as to ensure that the established model has representativeness and avoid uneven distribution of data in the modeling set. The modeling set is used to train the model. 7.The melamine identification method of fusion attention feature 1D-VDCNN-BiLSTM according to claim 1, wherein, The method comprises the following steps: Step one, constructing a one-dimensional ultra-deep convolutional neural network model 1D-VDCNN, i.e. improving a two-dimensional CNN model for one-dimensional data of melamine into a 1D-CNN, and then adding a VD structure; Step two, designing a bidirectional long short-term memory network (BiLSTM), which includes two independent LSTM processing modules: a time sequence forward processing module and a time sequence reverse processing module. In the data processing process, the forward module analyzes the input sequence in the natural order of the time axis to realize the progressive accumulation of historical features. The reverse module processes the input data in the reverse order of the time axis to extract the subsequent features of the sequence. The two complementary processing modules finally integrate their feature outputs into a unified classification layer. Meanwhile, the two modules maintain complete independence in the feature extraction stage, and there is no cross transmission of internal state information in any form; The forward and backward calculation formulas of BiLSTM are as follows: h i is the result of the current time of the hidden state in the LSTM network;h i-1 is the result of the previous time step of the hidden state in the LSTM network; f1, f2 represent the sigmoid activation function; A, B represent the input weight matrix of the current and previous time step of the LSTM network respectively; c represents the bias value of the LSTM network; The final output of each time step i is as follows: represents a concatenation operation; Step three, constructing a melamine identification model by fusing the attention mechanism, and the steps of implementing the attention mechanism are as follows: (1) Calculate weights: for a given query q and a set of keys k, calculate the similarity or relevance score d between them; (2) Normalization: normalize the score d using the softmax function to get the weight a corresponding to each key; (3) Weighted sum: according to the weight, the corresponding value a is weighted and summed to get the final attention output; In combination with the 1D-VDCNN constructed in step one, the BiLSTM constructed in step two and the attention mechanism, a 1D-VDCNN-BiLSTM hybrid architecture is established to realize the identification of melamine. 8.The melamine identification method of fusion attention feature 1D-VDCNN-BiLSTM according to claim 1 or 7, characterized in that, The core technology of VD structure design includes feature layer optimization and classification layer optimization. The hidden layer of 1D-CNN is divided into two parts: feature layer and classification layer. Feature layer optimization and classification layer optimization are to optimize the feature layer and the classification layer respectively. 9.A melamine discrimination system of fusion attention feature 1D-VDCNN-BiLSTM, characterized in that, The computer program instructions stored on the memory and capable of being executed by the processor can implement the steps of the method as claimed in any one of claims 1-8 when the processor executes the computer program instructions.

10. A computer readable storage medium having stored thereon computer program instructions capable of being executed by a processor, which can implement the steps of the method as claimed in any one of claims 1-8 when the processor executes the computer program instructions.

Images