Mussel production place tracing method fusing rare earth element characteristics and machine learning

Abstract

The invention provides a mussel production place tracing method fusing rare earth element characteristics and machine learning. The method comprises the following steps: forming a training data set containing production place labels; constructing a mussel origin traceability model, wherein a network structure of the mussel origin traceability model sequentially comprises a double-branch mixed feature engineering module, a feature attention module introducing a multi-head self-attention mechanism, a classification head module and a Softmax layer; obtaining a trained mussel origin traceability model; obtaining original rare earth element concentration data of a to-be-detected mussel sample, and forming a to-be-detected feature vector; and inputting the to-be-tested feature vector into the trained mussel origin traceability model to perform forward reasoning calculation so as to obtain the probability distribution of the to-be-tested mussel sample belonging to each preset origin, and taking the probability distribution as a final traceability result. According to the method, the overall accuracy and the model generalization ability of mussel origin traceability are remarkably improved.

Description

Technical Field

[0001] This invention relates to a method for tracing the origin of mussels, specifically a method that integrates rare earth element characteristics and machine learning, belonging to the field of seafood data processing and analysis technology. Background Technology

[0002] As an important economic shellfish, the quality and safety of mussels are closely related to the environment of their place of origin. Establishing an accurate and efficient traceability system is of great significance for protecting consumer rights, maintaining brand reputation, and implementing regional fisheries management.

[0003] Using chemical elements within organisms as fingerprint information for origin tracing is an effective method. Among them, rare earth elements, due to their unique geochemical properties, can retain significant regional characteristics during migration from the environment to organisms, and are therefore considered ideal origin tracers.

[0004] Existing rare earth element-based source tracing methods typically employ traditional multivariate statistical analysis techniques, such as principal component analysis (PCA) and linear discriminant analysis (LDA). These methods, when processing data, are primarily based on linear assumptions, making it difficult to effectively capture the complex and highly nonlinear intrinsic relationship between rare earth element concentration and origin. Furthermore, different rare earth elements do not contribute equally to distinguishing specific origins; traditional methods often treat all features equally, failing to dynamically focus on the most discriminative key feature combinations. This leads to a significant decrease in the accuracy and robustness of the source tracing model when faced with an increased number of origins or high similarity in origin characteristics.

[0005] On the other hand, researchers in the field have summarized some empirical formulas that can reflect the differentiation characteristics of rare earth elements, such as europium anomaly (δEu) and cerium anomaly (δCe). However, existing source tracing methods usually simply mix the derived features calculated by these formulas with the original concentration features, or rely on only one of them, failing to organically combine explicit expert knowledge with data-driven deep feature mining. The insufficient utilization of feature information leads to low source tracing accuracy. Summary of the Invention

[0006] Based on the above background, the purpose of this invention is to provide a mussel origin tracing method that integrates rare earth element characteristics and machine learning, thereby solving the problems of existing mussel origin tracing methods being unable to handle nonlinear relationships and insufficient attention to key features, resulting in low tracing accuracy.

[0007] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0008] A method for tracing the origin of mussels that integrates rare earth element characteristics and machine learning, the method comprising the following steps:

[0009] Raw rare earth element concentration data of mussel samples from multiple known origins were obtained to form a training dataset with origin labels.

[0010] A mussel origin traceability model is constructed. The network structure of the mussel origin traceability model sequentially includes a dual-branch hybrid feature engineering module, a feature attention module incorporating a multi-head self-attention mechanism, a classification head module, and a Softmax layer. The dual-branch hybrid feature engineering module receives the original feature vector and concatenates it after parallel processing by a formula calculation unit and a learnable branch to generate an enhanced feature vector. The feature attention module incorporating a multi-head self-attention mechanism receives the enhanced feature vector and generates a context-aware feature vector through a multi-head self-attention layer, subsequent residual summation, and layer normalization operations. The classification head module and the Softmax layer output classification probabilities based on the context-aware feature vector.

[0011] The training dataset is input into the network structure of the mussel origin traceability model. The network parameters are iteratively updated by optimizing the preset loss function until the model converges, and the trained mussel origin traceability model is obtained.

[0012] Obtain the original rare earth element concentration data of the mussel sample to be tested, and form the feature vector to be tested;

[0013] The feature vector to be tested is input into the trained mussel origin tracing model for forward inference calculation to obtain the probability distribution of the mussel sample to be tested belonging to each preset origin as the final tracing result.

[0014] Preferably, in the dual-branch hybrid feature engineering module, the learnable branch employs a multilayer perceptron with a single hidden layer structure, and the processing steps of the multilayer perceptron specifically include:

[0015] The input feature vector is subjected to a first linear transformation to generate a hidden vector of a preset dimension.

[0016] The hidden vector is nonlinearly activated using a nonlinear activation function;

[0017] The hidden vectors that have undergone nonlinear activation are subjected to a second linear transformation to generate hidden derived features of a preset dimension, which are then used as the output of the multilayer perceptron.

[0018] Preferably, the nonlinear activation function is a modified linear unit function or a Gaussian error linear unit function.

[0019] Preferably, the processing steps of the multi-head self-attention layer specifically include:

[0020] The input enhanced feature vector is linearly transformed into a query matrix, a key matrix, and a value matrix;

[0021] The query matrix, the key matrix, and the value matrix are input into multiple parallel attention heads, and each attention head independently performs scaled dot product attention calculation;

[0022] The outputs of the multiple attention heads are concatenated and then transformed through a final linear layer to obtain the output of the multi-head self-attention layer.

[0023] Preferably, the classification head module employs a multilayer perceptron, and the processing steps of the multilayer perceptron specifically include:

[0024] The context-aware feature vector is subjected to a first linear transformation to generate a 16-dimensional classification intermediate vector.

[0025] The 16-dimensional classification intermediate vector is nonlinearly activated using a nonlinear activation function;

[0026] The classification intermediate vector that has undergone nonlinear activation is subjected to a second linear transformation to generate a K-dimensional classification vector as the output of the classification head module, where K is the number of preset production locations and K is an integer greater than or equal to 2.

[0027] Preferably, the network structure of the mussel origin traceability model also includes a backbone network set before the dual-branch hybrid feature engineering module, which is used to perform preliminary feature extraction on the original feature vector.

[0028] Preferably, the backbone network employs a multilayer perceptron, and the processing steps of the multilayer perceptron specifically include:

[0029] The original 14-dimensional feature vector is subjected to a first linear transformation to generate a first 32-dimensional intermediate vector.

[0030] The 32-dimensional first intermediate vector is non-linearly activated using a non-linear activation function.

[0031] The first intermediate vector, which has undergone nonlinear activation, is subjected to a second linear transformation to generate a 64-dimensional second intermediate vector.

[0032] The 64-dimensional second intermediate vector is nonlinearly activated using a nonlinear activation function to obtain the final deep feature vector, which is then used as the output of the backbone network.

[0033] Preferably, in the dual-branch hybrid feature engineering module, the geochemical formula preset by the formula calculation unit includes at least one of the following: total ratio of light and heavy rare earth elements, europium anomaly, cerium anomaly, and degree of differentiation of light and heavy rare earth elements.

[0034] Preferably, the original rare earth element concentration data includes the concentration data of rare earth elements such as lanthanum, cerium, praseodymium, neodymium, samarium, europium, gadolinium, terbium, dysprosium, holmium, erbium, thulium, ytterbium, and lutetium.

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] This invention presents a mussel origin tracing method that integrates rare earth element features and machine learning. It constructs a complete and efficient tracing process from hybrid feature enhancement to key feature focusing. This method can not only effectively handle the complex nonlinear relationship between rare earth elements and origin, but also make full use of various types of feature information, significantly improving the overall accuracy and model generalization ability of mussel origin tracing. Even when facing the challenge of similar origin features, it can maintain high tracing performance.

[0037] This invention sets up a dual-branch hybrid feature engineering module to extract and fuse expert knowledge based on geochemical formulas with hidden features learned through neural networks in parallel. This not only utilizes known and effective prior knowledge, but also mines deeper nonlinear patterns from the data to make up for the possible limitations of expert knowledge, thus significantly enhancing the expressive power of features.

[0038] This invention introduces a feature attention module with a multi-head self-attention mechanism, which can dynamically evaluate the importance of each feature in the enhanced feature vector and assign higher weights to more important features. This enables the traceability model to focus on the most critical feature combinations that distinguish different origins, suppress interference from irrelevant or redundant features, and thus make more accurate judgments in complex feature spaces. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0040] Figure 1 This is a schematic diagram of the network structure of the mussel origin traceability model in this invention;

[0041] Figure 2 This is a schematic diagram of the structure of the dual-branch hybrid feature engineering module in this invention;

[0042] Figure 3 This is a schematic diagram of the feature attention module that introduces a multi-head self-attention mechanism in this invention. Detailed Implementation

[0043] The technical solution of the present invention will be further described in detail below through specific embodiments and in conjunction with the accompanying drawings. It should be understood that the implementation of the present invention is not limited to the following embodiments, and any modifications and / or alterations made to the present invention will fall within the protection scope of the present invention.

[0044] In this invention, unless otherwise specified, all parts and percentages are by weight, and the equipment and raw materials used are commercially available or commonly used in the art. Unless otherwise specified, the methods in the following embodiments are conventional methods in the art. Unless otherwise specified, the components or equipment in the following embodiments are general standard parts or components known to those skilled in the art, and their structures and principles can be learned by those skilled in the art through technical manuals or conventional experimental methods.

[0045] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings. In this detailed description, numerous specific details are set forth to facilitate explanation and provide a thorough understanding of the embodiments of the present invention. However, one or more embodiments may be practiced by those skilled in the art without these specific details.

[0046] The present invention discloses a mussel origin tracing method that integrates rare earth element characteristics and machine learning, which mainly includes data acquisition and preprocessing, model building, model training and inference tracing. The specific steps are described in detail below.

[0047] I. Data Acquisition and Preprocessing

[0048] Mussel samples were collected from K known geographical locations. The concentrations of 14 rare earth elements in the soft tissue of each mussel sample were determined using chemical analysis techniques such as inductively coupled plasma mass spectrometry (ICP-MS). These elements included: lanthanum (La), cerium (Ce), praseodymium (Pr), neodymium (Nd), samarium (Sm), europium (Eu), gadolinium (Gd), terbium (Tb), dysprosium (Dy), holmium (Ho), erbium (Er), thulium (Tm), ytterbium (Yb), and lutetium (Lu).

[0049] The concentration data of the 14 rare earth elements in each sample were used to construct a 14-dimensional original feature vector, and the corresponding place of origin was attached to it. The sample data from all known places of origin together constituted the training dataset for model training.

[0050] To eliminate the impact of differences in concentration magnitudes between different elements and to make the model training process more stable and converge faster, the original feature vector can be normalized. Specifically, the Z-score normalization method can be used to convert the concentration data of each rare earth element in all samples into a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0051] II. Model Construction

[0052] like Figure 1 As shown, this invention constructs a mussel origin tracing model named GeoAtt-Net. This model is a deep neural network, and its network structure, from top to bottom, includes an input layer, a backbone network, a dual-branch hybrid feature engineering module (GFEM), a feature attention module (FCAM), a classification head module (Head), and a softmax layer.

[0053] The backbone network maps the original 14-dimensional feature vectors to a higher-dimensional space with richer feature representations, providing higher-quality input for subsequent modules. Specifically, the backbone network is implemented using a multilayer perceptron (MLP), and its processing steps include:

[0054] The original 14-dimensional feature vector is input into a fully connected layer, and a first 32-dimensional first intermediate vector is generated through the first linear transformation.

[0055] A nonlinear activation function (ReLU function is used in this embodiment) is applied to the first intermediate vector of 32 dimensions to increase the nonlinear expressive power of the network;

[0056] The activated 32-dimensional vector is input into the second fully connected layer, and a second linear transformation is performed to generate a 64-dimensional second intermediate vector.

[0057] The nonlinear activation function is applied again to the 64-dimensional second intermediate vector to obtain a 64-dimensional deep feature vector, which is then used as the output of the backbone network and passed to the GFEM module.

[0058] like Figure 2 As shown, the dual-branch hybrid feature engineering module (GFEM) receives a 64-dimensional deep feature vector from the backbone network and processes it through two parallel branches.

[0059] The formula calculation unit integrates domain expert knowledge. Based on preset geochemical formulas, it calculates the input feature vector to generate N-dimensional derived features. These formulas reflect the differentiation patterns of rare earth elements. Specifically, the preset formulas include:

[0060] Europium anomaly (δEu): The calculation formula is δEu = Eu N / sqrt(Sm N ×Gd N This indicator is used to reflect europium deficiency or enrichment.

[0061] Cerium anomaly (δCe): The calculation formula is δCe = Ce N / sqrt(La N×Pr N ).

[0062] Light rare earth element ratio (LREE / HREE): La to Gd are defined as light rare earth elements, and Tb to Lu are defined as heavy rare earth elements. The ratio of the sum of their concentrations is calculated.

[0063] Differentiation between light and heavy rare earth elements (La / Yb) N ): Reflects the differentiation between light and heavy rare earth elements.

[0064] Using these four formulas, the formula calculation unit outputs a 4-dimensional derived feature vector.

[0065] Learnable branches are used to learn hidden nonlinear feature combinations in data. Learnable branches are implemented using a multilayer perceptron with a single hidden layer, and their processing steps include:

[0066] The input 64-dimensional feature vector is subjected to a first linear transformation to generate a 32-dimensional hidden vector.

[0067] A nonlinear activation function is applied to the 32-dimensional hidden vector, specifically the Corrected Linear Unit (ReLU) function or the Gaussian Error Linear Unit (GELU) function;

[0068] The activated hidden vector is then subjected to a second linear transformation to generate a 16-dimensional hidden derived feature.

[0069] Finally, the 64-dimensional deep feature vector from the backbone network, the 4-dimensional derived feature vector output by the formula calculation unit, and the 16-dimensional hidden derived features output by the learnable branch are concatenated to obtain an 84-dimensional enhanced feature vector.

[0070] like Figure 3 As shown, the core of the Feature Attention Module (FCAM) ​​is a Multi-Head Self-Attention Layer (MHA Layer), followed by residual connections and layer normalization. The specific processing steps of the FCAM module are as follows.

[0071] The input 64-dimensional enhanced feature vector is copied three times and passed through three independent linear transformation layers to generate the query matrix Q, the key matrix K, and the value matrix V.

[0072] The Q, K, and V matrices are divided into h heads along the feature dimension, and these heads are fed into h parallel attention heads.

[0073] Within each attention head, scaling dot product attention is calculated independently. This process essentially calculates the similarity between the query vector and all key vectors, and uses this similarity as a weight to perform a weighted summation of the value vectors.

[0074] The outputs of all h attention heads are concatenated along the feature dimension;

[0075] The concatenated vectors are transformed through a final linear layer to obtain the output of the multi-head self-attention layer;

[0076] The output of the multi-head self-attention layer is added element-wise to its input (i.e., the 64-dimensional enhanced feature vector) to prevent gradient vanishing. Then, the result of the addition is normalized by layer to accelerate model convergence and improve training stability, and finally the context-aware feature vector is obtained.

[0077] The classification head module receives context-aware feature vectors from the FCAM module and maps them to the final classification space. The classification head module is implemented using a multilayer perceptron, and its processing steps specifically include:

[0078] The input context-aware feature vector is subjected to a first linear transformation to generate a 16-dimensional classification intermediate vector.

[0079] A nonlinear activation function is applied to the 16-dimensional classification intermediate vector;

[0080] The activated classification intermediate vector is subjected to a second linear transformation to generate a K-dimensional classification vector, where K is the total number of preset origins.

[0081] Finally, the Softmax layer processes the K-dimensional classification vector, transforming it into a probability distribution that sums to 1, where each value represents the probability that a sample belongs to the corresponding origin.

[0082] III. Model Training

[0083] All samples were randomly divided into a training set (80%), a validation set (10%), and a test set (10%) in a ratio of 8:1:1. The training set was used to calculate gradients and update model parameters, the validation set was used to monitor model performance and select the best model during training, and the test set was used to objectively evaluate the generalization ability of the final model after training was completed.

[0084] For a multi-class classification task with K origins, this invention employs a cross-entropy loss function with label smoothing. The standard cross-entropy loss function aims to make the model's prediction probability for the correct label infinitely close to 1. When the amount of data is limited or noisy, this can easily lead to model overconfidence, overfitting, and reduced generalization ability. Label smoothing, by softening the rigid one-hot labels into a distribution with slight noise, encourages the model to avoid making overly absolute predictions. This effectively suppresses overfitting and improves the model's robustness.

[0085] The optimizer is the AdamW optimizer. Compared to the traditional Adam optimizer, AdamW decouples the calculation of weight decay from gradient updates. This provides a more effective and stable regularization effect when training Transformer-like models with attention mechanisms, preventing the model parameters from becoming too large. The optimizer's hyperparameters are set as follows: beta1=0.9, beta2=0.999.

[0086] The learning rate strategy employed was a cosine annealing learning rate scheduler with warm-up. The total number of training epochs was set to 200, with a batch size of 64. At the end of each epoch, the accuracy of the current model was evaluated using the validation set. Throughout the entire training process, only the model parameters that achieved the highest accuracy on the validation set were saved as the final trained mussel origin tracing model.

[0087] The gradient of the loss function with respect to the model network parameters is calculated using the backpropagation algorithm, and the optimizer is used to iteratively update all learnable parameters based on the gradient until the value of the loss function converges to a stable state or reaches the preset number of training rounds, thus obtaining the trained mussel origin traceability model.

[0088] IV. Reasoning and Origin

[0089] When mussel samples from the place of origin to be tested are available, the concentration data of 14 rare earth elements are also measured and normalized to form the feature vector to be tested.

[0090] The feature vector to be tested is input into the trained mussel origin tracing model for a complete forward inference calculation. The model outputs a K-dimensional probability distribution vector, which represents the probability that the mussel sample belongs to each of the K known origins. The origin with the highest probability value is determined as the final tracing result.

[0091] To verify the effectiveness and advancement of the proposed GeoAtt-Net model, experiments were conducted on a dataset containing rare earth element data from 1200 mussel samples across five different origins. The experimental environment consisted of an Ubuntu 20.04 operating system, an NVIDIA RTX 3090 GPU, and implementation using the PyTorch deep learning framework. The evaluation metric was the overall accuracy (OA) of the traceability analysis.

[0092] To verify the effectiveness of the dual-branch hybrid feature engineering module (GFEM) and feature attention module (FCAM) ​​in this invention, the following ablation experiment was designed:

[0093] Model A (benchmark): Classification is performed by directly connecting the backbone network to the classification head.

[0094] Model B: Based on Model A, a dual-branch hybrid feature engineering module (GFEM) is added.

[0095] Model C (this invention): Based on Model B, a Feature Attention Module (FCAM) ​​is added to form the complete GeoAtt-Net.

[0096] Table 1 Comparison of ablation experimental performance of each module of the model

[0097]

[0098] As shown in Table 1, compared to the baseline model A which only uses deep features, model B, after adding the GFEM module, achieved an accuracy improvement of 2.7 percentage points. This fully demonstrates that the strategy of hybrid enhancement using expert knowledge (formula calculation unit) and data-driven hidden features (learnable branches) is effective, providing the model with richer and more discriminative features. Compared to model B, model C, after adding the FCAM module, further improved its accuracy by 1.3 percentage points, reaching a maximum of 97.5%. This strongly proves that introducing a multi-head self-attention mechanism allows the model to dynamically focus on the feature combination most important to the source tracing task, suppressing the interference of redundant information, thereby achieving more accurate classification and discrimination.

[0099] To further demonstrate the overall performance advantages of the method of this invention, two classic multivariate statistical analysis methods and a basic deep learning method were selected as comparative baselines.

[0100] LDA: Linear Discriminant Analysis, a classic supervised dimensionality reduction and classification method.

[0101] SVM: Support Vector Machine, a machine learning method for handling non-linear classification problems.

[0102] MLP: Basic Multilayer Perceptron, which directly uses a simple fully connected neural network to classify raw 14-dimensional data.

[0103] This invention (GeoAtt-Net): Employs the complete model of this invention.

[0104] Table 2. Performance comparison of different methods in mussel origin traceability tasks.

[0105]

[0106] As can be seen from the comparison results in Table 2, the GeoAtt-Net model proposed in this invention achieves an overall accuracy of 97.5%, significantly outperforming all compared methods. Compared to traditional LDA and SVM methods, the advantage of this invention's model lies in its ability to learn and capture the highly nonlinear relationship between rare earth elements and their origin through a deep network structure. Compared to basic MLP models, this invention achieves a significant performance improvement of over 5 percentage points by integrating expert knowledge with deep features and focusing on key information through GFEM and FCAM modules.

[0107] Although the GeoAtt-Net model of this invention, due to the introduction of new modules such as GFEM and FCAM, has increased parameters and training time compared to a simple MLP, the reasonable computational overhead incurred to achieve a decisive improvement in accuracy is entirely worthwhile for applications like product origin traceability, which demand extremely high accuracy. Experimental results fully demonstrate that the method of this invention can provide a high-precision and robust solution for mussel origin traceability that surpasses the level of existing technologies.

[0108] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A method of mussel provenance tracing fusing rare earth element characteristics and machine learning, characterized in that: The method comprises the following steps: Obtaining original rare earth element concentration data of mussel samples from multiple known origins to form a training data set containing origin labels; Building a mussel origin traceability model, the network structure of the mussel origin traceability model comprising in turn a double-branch mixed feature engineering module, a feature attention module introducing a multi-head self-attention mechanism, a classification head module and a Softmax layer; the double-branch mixed feature engineering module is used to receive an original feature vector and generate an enhanced feature vector after parallel processing by a formula calculation unit and a learnable branch and splicing; the feature attention module introducing a multi-head self-attention mechanism is used to receive the enhanced feature vector and generate a context-aware feature vector after a multi-head self-attention layer and subsequent residual addition and layer normalization operations; the classification head module and the Softmax layer are used to output classification probabilities according to the context-aware feature vector; Inputting the training data set into the network structure of the mussel origin traceability model, iteratively updating network parameters by optimizing a preset loss function until the model converges, and obtaining a trained mussel origin traceability model; Obtaining original rare earth element concentration data of a to-be-tested mussel sample to form a to-be-tested feature vector; Inputting the to-be-tested feature vector into the trained mussel origin traceability model for forward inference calculation to obtain probability distributions of the to-be-tested mussel sample belonging to each preset origin as a final traceability result.

2. The method according to claim 1, wherein: In the double-branch mixed feature engineering module, the learnable branch adopts a multi-layer perceptron with a single hidden layer structure, and the processing steps of the multi-layer perceptron specifically include: Performing first linear transformation on the input feature vector to generate a hidden vector of a preset dimension; Performing nonlinear activation on the hidden vector by a nonlinear activation function; Performing second linear transformation on the nonlinearly activated hidden vector to generate hidden derivative features of a preset dimension as the output of the multi-layer perceptron.

3. The method according to claim 2, wherein: The nonlinear activation function is a rectified linear unit function or a Gaussian error linear unit function.

4. The method according to claim 1, wherein: The processing steps of the multi-head self-attention layer specifically include: Linearly transforming the input enhanced feature vector into a query matrix, a key matrix and a value matrix; Inputting the query matrix, the key matrix and the value matrix into multiple parallel attention heads, and each attention head independently performs scaled dot-product attention calculation; Splicing the outputs of the multiple attention heads, performing transformation through a final linear layer to obtain the output of the multi-head self-attention layer.

5. The method according to claim 1, wherein: The classification head module adopts a multi-layer perceptron, and the processing steps of the multi-layer perceptron specifically include: Performing first linear transformation on the context-aware feature vector to generate a classification intermediate vector of a preset dimension; Performing nonlinear activation on the classification intermediate vector by a nonlinear activation function; Performing second linear transformation on the nonlinearly activated classification intermediate vector to generate a K-dimensional classification vector as the output of the classification head module, wherein K is the number of preset origins, and K is an integer greater than or equal to 2.

6. The method according to claim 1, wherein: The network structure of the mussel origin traceability model further comprises a backbone network arranged before the double-branch mixed feature engineering module, and the backbone network is configured to perform preliminary feature extraction on the original feature vector.

7. The method according to claim 6, wherein: The backbone network adopts a multi-layer perceptron, and the processing steps of the multi-layer perceptron specifically include: performing first linear transformation on the 14-dimensional original feature vector to generate a 32-dimensional first intermediate vector; performing nonlinear activation on the 32-dimensional first intermediate vector through a nonlinear activation function; performing second linear transformation on the first intermediate vector after nonlinear activation to generate a 64-dimensional second intermediate vector; performing nonlinear activation on the 64-dimensional second intermediate vector through a nonlinear activation function to obtain a final deep feature vector as the output of the backbone network.

8. The method according to claim 1, wherein: In the double-branch mixed feature engineering module, the geochemical formula preconfigured in the formula calculation unit includes at least one of total light and heavy rare earth ratio, europium anomaly, cerium anomaly, and light and heavy rare earth differentiation degree.

9. The method according to claim 1, wherein: The original rare earth element concentration data includes concentration data of lanthanum, cerium, praseodymium, neodymium, samarium, europium, gadolinium, terbium, dysprosium, holmium, erbium, thulium, ytterbium, and lutetium rare earth elements.

Images