Rapid variant identification method based on multi-loss multi-feature fusion

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a rapid variant recognition method based on multi-loss multi-feature fusion, and the method comprises the steps: S1, collecting multi-source gene sequence data through a biosensor, including genome sequencing data, epigenetic data and transcriptome data; s2, performing data preprocessing and feature extraction processing on the multi-source gene sequence data to generate a multi-dimensional feature vector set; by integrating multi-source biological data and performing cross-modal feature fusion, the information limitation of a single data source is overcome, the comprehensiveness and richness of feature representation are improved, and finer variant feature differences are captured, so that the accuracy and robustness of variant recognition are improved, and misjudgment caused by data missing and feature singleness is avoided; and by introducing a loss function collaborative optimization mechanism, dynamic balance classification and a relationship between regression and regularization constraint, the reliability and consistency of a variant identification result are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a fast variant recognition method based on multi-loss and multi-feature fusion. Background Technology

[0002] Feature fusion is a technique that extracts different feature vectors from the same pattern and combines them in an optimized manner. It includes two typical implementation methods: serial and parallel. It is divided into data-level, feature-level and decision-level fusion layers, which improve the accuracy and robustness of the model by integrating multi-source information.

[0003] Currently, in the field of gene variant identification, traditional methods mainly rely on single-type biometric data for variant detection and generally use a single loss function for optimization. They lack a fusion mechanism for multi-source heterogeneous data and cannot balance the conflict between classification accuracy, generalization ability and regularization constraints in the identification task. This results in incomplete feature representation, difficulty in capturing complex biological signal associations, and easy overfitting or underfitting of the model, affecting the stability and reliability of variant identification.

[0004] Therefore, a fast variant recognition method based on multi-loss and multi-feature fusion is proposed to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a fast variant identification method based on multi-loss and multi-feature fusion, which solves the problem of incomplete feature representation and difficulty in capturing complex biological signal associations mentioned in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a fast variant recognition method based on multi-loss and multi-feature fusion, the method comprising the following steps: S1. Collect multi-source gene sequence data, including genome sequencing data, epigenetic data, and transcriptome data, through biosensors; S2. Perform data preprocessing and feature extraction on the multi-source gene sequence data to generate a multi-dimensional feature vector set; S3. Input the multi-dimensional feature vector set into the feature fusion module for cross-modal fusion processing to generate a fused feature tensor; S4. Construct a multi-loss prediction model, perform variant recognition training based on the fused feature tensor, and generate variant classification parameters. S5. Perform variant type identification processing based on the variant classification parameters to generate variant identification results; S6. The variant identification results are executed through the output module to generate a visual report; S7. Real-time monitoring of variant recognition accuracy and feedback to step S1 form an adaptive optimization closed-loop mechanism.

[0007] Preferably, the acquisition of multi-source gene sequence data in S1 includes the following steps: S11. Collect genome sequencing data using high-throughput sequencing equipment, generate raw reads using paired-end sequencing technology, and perform real-time quality filtering to remove low-confidence sequences. S12. Collect epigenetic data through methylation chips, generate a methylation level matrix, and apply a standardized protocol to correct for batch effects. S13. Collect transcriptome data using RNA sequencing equipment, extract gene expression feature vectors, and generate highly consistent expression profile data through normalization.

[0008] Preferably, in step S2, data preprocessing and feature extraction are performed to generate a multi-dimensional feature vector set, including the following steps: S21. Perform quality filtering and noise reduction on the genome sequencing data to generate standardized read data, and use a sliding window algorithm to detect sequencing errors; The specific implementation steps of the sliding window algorithm are as follows: set a window of fixed length L, slide along the read segment sequence, and slide one base position at a time; Calculate the average quality score of all bases within the window. When the score is lower than a preset threshold, mark the bases in the window as low quality and correct or remove them. The formula for calculating the average quality score within the window is: ; in The average quality score of all bases within the window. The length of the window. This serves as an index for the bases within the window. The quality score of the i-th base within the window; S22. Principal component analysis is used to reduce the dimensionality of the methylation level matrix to generate a low-dimensional feature set, and the feature variance is calculated to retain the high-information dimension. S23. The gene expression feature vector is spatiotemporally aligned and fused with the low-dimensional feature set to generate a five-dimensional feature tensor, and then coordinate transformation is performed.

[0009] Preferably, generating the fused feature tensor in S3 includes the following steps: S31. The feature fusion module is designed as a multi-head attention mechanism module, with the input layer dimension set to 256 dimensions and learnable weight parameters configured. S32. Input the five-dimensional feature tensor into the multi-head attention mechanism module for weight allocation processing, and integrate cross-modal information by calculating cross-attention scores; S33. Output the fused feature tensor, integrate cross-modal information through the attention weight matrix, and add residual connections.

[0010] Preferably, the construction of the multi-loss prediction model in S4 includes the following steps: S41. Construct a multi-loss prediction model using a hybrid architecture of convolutional neural network and long short-term memory network, and embed batch normalization layer to accelerate convergence; S42. Define a set of multiple loss functions, including classification loss function, regression loss function and regularization loss function, and set dynamic weight coefficients; S43. Based on the fused feature tensor, train the model to generate variant classification parameters, wherein the classification loss function is calculated using the following formula: ; in For classification loss value, For the actual label vector, For the predicted probability vector, This represents the number of samples.

[0011] Preferably, the variant type identification process in S5 includes the following steps: S51. Analyze variant type atlas data, establish a variant-disease association model, and integrate public databases; S52. Calculate the variant confidence index based on the variant classification parameters, and use the Bayesian method to quantify the uncertainty. S53. Generate a variant identification report matrix based on the association relationship model and confidence index, and filter low-confidence variants by thresholding.

[0012] Preferably, the visualization report generation in step S6 includes the following steps: S61. Encode the variant identification result into a standard BAM file format and add metadata tags; S62. Data transmission is secured by TLS protocol and transmitted to the terminal device via an encrypted cloud platform. S63. Generates real-time visual interactive reports, triggers model retraining when the recognition confidence level is below 90%, and records exception logs to optimize system resilience.

[0013] Preferably, a real-time optimization mechanism is also included: S81. Real-time monitoring of feature fusion efficiency; when the data dimension conflict rate exceeds the threshold, switch to backup fusion mode. S82. Employ gradient descent algorithm to dynamically adjust the weight parameters of multiple losses and update the coefficients based on validation set performance feedback. S83. Optimize the model architecture by matching similar scenarios based on historical variant databases, and integrate new data by applying an incremental learning mechanism.

[0014] Preferably, it also includes an exception handling mechanism: S91. Detect the abnormal rate of data collection. When the missing rate exceeds 15%, activate the imputation algorithm and use multiple imputation techniques to fill in the missing values. S92. Use the K-nearest neighbor algorithm to fill in missing features and calculate sample similarity to select the best neighbor; S93. Initiate the continuity of the redundant computing node identification process and deploy a failover mechanism to automatically switch over when the master node fails.

[0015] Preferably, this also includes a model lightweighting mechanism: S101. Perform knowledge distillation and compression on the multi-loss prediction model, and use the teacher-student framework to transfer key knowledge. S102. Generate a lightweight model parameter set and quantize the weights to 8-bit integers; S103, deployed to edge devices to achieve real-time variant recognition, and supports mobile applications through API interfaces.

[0016] (III) Beneficial Effects Compared with existing technologies, this invention provides a fast variant recognition method based on multi-loss and multi-feature fusion, which has the following advantages: 1. In this invention, when performing rapid identification of gene variants, by integrating multi-source biological data and performing cross-modal feature fusion, the information limitations of a single data source are overcome, the comprehensiveness and richness of feature representation are improved, and more subtle variant feature differences are captured, thereby improving the accuracy and robustness of variant identification and avoiding misjudgments caused by missing data or single features.

[0017] 2. In this invention, when training the variant recognition model, a loss function collaborative optimization mechanism is introduced to dynamically balance the relationship between classification, regression and regularization constraints, so that the model can simultaneously take into account the fitting ability and generalization performance, avoid overfitting, enhance the adaptability and stability of the model in multiple samples and multiple scenarios, and improve the reliability and consistency of variant recognition results.

[0018] 3. In this invention, when deploying the variant recognition system, by integrating lightweight processing and real-time optimization mechanisms, the computational resource requirements and processing latency are reduced, enabling efficient operation on resource-constrained edge devices and achieving real-time variant recognition with millisecond-level response. At the same time, it supports online model updates and anomaly adaptive handling, improving the system's practicality, scalability, and engineering implementation capabilities. Attached Figure Description

[0019] Figure 1 This is a flowchart of a fast variant recognition method based on multi-loss and multi-feature fusion according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Specific Implementation: A fast variant recognition method based on multi-loss and multi-feature fusion, the method includes the following steps: S1. Collect multi-source gene sequence data, including genome sequencing data, epigenetic data, and transcriptome data, through biosensors; S2. Perform data preprocessing and feature extraction on multi-source gene sequence data to generate a multi-dimensional feature vector set; S3. Input the multi-dimensional feature vector set into the feature fusion module for cross-modal fusion processing to generate a fused feature tensor. S4. Construct a multi-loss prediction model, perform variant recognition training based on the fused feature tensor, and generate variant classification parameters. S5. Perform variant type identification processing based on variant classification parameters to generate variant identification results; S6. The variant recognition results are executed through the output module to generate a visual report; S7. Real-time monitoring of variant recognition accuracy and feedback to step S1 form an adaptive optimization closed-loop mechanism.

[0022] The collection of multi-source gene sequence data in S1 includes the following steps: S11. Collect genome sequencing data using high-throughput sequencing equipment, generate raw reads using paired-end sequencing technology, and perform real-time quality filtering to remove low-confidence sequences. S12. Collect epigenetic data through methylation chips, generate a methylation level matrix, and apply a standardized protocol to correct for batch effects. S13. Collect transcriptome data using RNA sequencing equipment, extract gene expression feature vectors, and generate highly consistent expression profile data through normalization.

[0023] In S2, data preprocessing and feature extraction are performed to generate a multi-dimensional feature vector set, including the following steps: S21. Perform quality filtering and noise reduction on the genome sequencing data to generate standardized read data, and use a sliding window algorithm to detect sequencing errors; The specific implementation steps of the sliding window algorithm are as follows: set a window of fixed length L, slide along the read segment sequence, and slide one base position at a time; Calculate the average quality score of all bases within the window. When the score is lower than a preset threshold, mark the bases in the window as low quality and correct or remove them. The formula for calculating the average quality score within the window is: ; in The average quality score of all bases within the window. The length of the window. This serves as an index for the bases within the window. The quality score of the i-th base within the window; S22. Principal component analysis is used to reduce the dimensionality of the methylation level matrix to generate a low-dimensional feature set. At the same time, the feature variance is calculated to retain the high-information dimension. The specific steps of principal component analysis are as follows: standardize the methylation level matrix so that its mean is 0 and its variance is 1; calculate the covariance matrix of the standardized data; calculate the eigenvalues ​​and eigenvectors of the covariance matrix; finally, sort the eigenvalues ​​from largest to smallest, select the eigenvectors corresponding to the first k eigenvalues ​​as the principal component directions, and project the original data onto these principal components to form a low-dimensional feature set.

[0024] Calculating feature variance to preserve high-information dimensions means determining the k value by calculating the variance contribution rate of each principal component, and retaining the first k principal components whose cumulative variance contribution rate is greater than a predetermined threshold. The formula for the variance contribution rate of the i-th principal component is: ; in For the i-th eigenvalue, For the j-th eigenvalue, The total number of original features. Index of the eigenvalue; S23. Spatiotemporally align and fuse the gene expression feature vector with the low-dimensional feature set to generate a five-dimensional feature tensor, and then perform coordinate transformation.

[0025] The steps involved in generating the fused feature tensor in S3 are as follows: S31. The feature fusion module is designed as a multi-head attention mechanism module, with the input layer dimension set to 256 dimensions and learnable weight parameters configured. S32. Input the five-dimensional feature tensor into the multi-head attention mechanism module for weight allocation processing, and integrate cross-modal information by calculating the cross-attention score; Calculating the cross-attention score is the core step of the attention mechanism. First, the input feature tensor is transformed linearly to generate query vector, key vector, and value vector respectively. Then, the dot product of the query vector and all key vectors is calculated and divided by a scaling factor to prevent gradient vanishing. Finally, the Softmax function is applied to obtain the attention weights, and the calculation formula is as follows: ; in For the fused feature tensor, For query vector, For key vectors, For value vectors, Let be the dimension of the key vector. To query the similarity between the key and the query result.

[0026] S33. Output the fused feature tensor, integrate cross-modal information through the attention weight matrix, and add residual connections.

[0027] Building a multi-loss prediction model in S4 involves the following steps: S41. Construct a multi-loss prediction model using a hybrid architecture of convolutional neural network and long short-term memory network, and embed batch normalization layer to accelerate convergence; S42. Define a set of multiple loss functions, including classification loss function, regression loss function and regularization loss function, and set dynamic weight coefficients; S43. Train the model based on the fused feature tensor to generate variant classification parameters, where the classification loss function is calculated using the following formula: ; in For classification loss value, For the actual label vector, For the predicted probability vector, This represents the number of samples.

[0028] Variant type identification processing in S5 includes the following steps: S51. Analyze variant type atlas data, establish a variant-disease association model, and integrate public databases; S52. Calculate the variant confidence index based on the variant classification parameters and use the Bayesian method to quantify the uncertainty. The specific steps for calculating the variant confidence index using the Bayesian method are as follows: Treat the predicted probability output by the model as the likelihood; introduce a prior probability distribution based on historical data or expert knowledge; calculate the posterior probability using Bayes' theorem; and use this posterior probability as a calibrated, more reliable confidence index. The core formula is: ; in For posterior probability, For likelihood, For prior probability, The marginal probability of the data; S53. Generate a variant identification report matrix based on the association relationship model and confidence index, and filter low-confidence variants by thresholding.

[0029] Generating visual reports in S6 involves the following steps: S61. Encode the variant identification results into a standard BAM file format and add metadata tags; S62. Data transmission is secured by TLS protocol and transmitted to the terminal device via an encrypted cloud platform. S63. Generates real-time visual interactive reports, triggers model retraining when the recognition confidence level is below 90%, and records exception logs to optimize system resilience.

[0030] It also includes a real-time optimization mechanism: S81. Real-time monitoring of feature fusion efficiency; when the data dimension conflict rate exceeds the threshold, switch to backup fusion mode. S82. Employ gradient descent algorithm to dynamically adjust the weight parameters of multiple losses and update the coefficients based on validation set performance feedback. The specific steps of dynamically adjusting the weight parameters of multiple loss functions using the gradient descent algorithm are as follows: ; in For the first A loss of size, Its weight; The overall performance metric on the validation set is used as the optimization objective. The gradient of this metric with respect to the weight parameters is calculated, and a small update is performed in the reverse direction of the gradient. ; in The weight of the i-th sub-loss at the next time step. The weight of the i-th sub-loss at the current time. For learning rate, For the objective function Weights The partial derivatives; The appropriate step size for control updates is used to adaptively adjust the weight coefficients through iterative updates, thereby maximizing the performance of the validation set. S83. Optimize the model architecture by matching similar scenarios based on historical variant databases, and integrate new data by applying an incremental learning mechanism.

[0031] It also includes an exception handling mechanism: S91. Detect the abnormal rate of data collection. When the missing rate exceeds 15%, activate the imputation algorithm and use multiple imputation techniques to fill in the missing values. The specific implementation steps of the multiple interpolation technique are as follows: Based on the distribution characteristics of existing data, an imputation model is constructed. This model is used to randomly generate multiple imputation values, thereby creating multiple complete datasets. Standard analysis is performed on each imputed complete dataset. The results of multiple analyses are merged according to Rubin's Rules to obtain the final parameter estimates and statistical inferences. This result also takes into account the uncertainty brought about by the missing data itself.

[0032] S92. Use the K-nearest neighbor algorithm to fill in missing features and calculate sample similarity to select the best neighbor; The specific steps of the K-nearest neighbor algorithm for filling missing values ​​are as follows: For samples with missing values, in the feature space, based on the values ​​of their existing features, calculate the Euclidean distance to all other complete samples, and select the K nearest complete samples. For consecutive missing features, use the mean or weighted mean of these K neighbors for that feature to fill in the missing values; for categorical features, use the most frequent category among the K neighbors to fill in the missing values. The Euclidean distance is calculated as follows: ; in The Euclidean distance between sample x and sample y This represents the number of common non-missing features among the samples. Index of features Let x be the value of the sample on the i-th feature. Let y be the value of the sample on the i-th feature; S93. Initiate the continuity of the redundant computing node identification process and deploy a failover mechanism to automatically switch over when the master node fails.

[0033] It also includes a model lightweighting mechanism: S101. Perform knowledge distillation and compression on the multi-loss prediction model, and use the teacher-student framework to transfer key knowledge. S102. Generate a lightweight model parameter set and quantize the weights to 8-bit integers; S103, deployed to edge devices to achieve real-time variant recognition, and supports mobile applications through API interfaces.

[0034] The steps of this method are as follows: Step 1: Multi-source biological data acquisition and feature extraction Genome sequencing data, epigenetic data, and transcriptome data were simultaneously acquired using biosensors. The raw data underwent standardized preprocessing. Genomic data: Sliding window technology was used to scan read sequences, dynamically detect and correct low-quality bases, and eliminate sequencing noise; Epigenetic data: Principal component analysis was used for dimensionality reduction to retain high-information features and reduce data redundancy; Transcriptome data: Gene expression vectors were extracted and normalized and aligned to ensure spatiotemporal consistency across modal data; The final result is a multi-dimensional feature vector set containing genomic variation features, methylation patterns, and gene expression profiles.

[0035] Step 2: Cross-modal feature fusion and tensor generation Input multi-dimensional features into the multi-head attention fusion module: The module integrates deep associations between genomic, epigenetic, and transcriptomic features by dynamically calculating cross-modal association weights.

[0036] Residual connections are used to preserve the original feature information and generate feature tensors that incorporate biological context information, thereby enhancing the discriminative power of variant features.

[0037] Step 3: Multi-loss collaborative model training and variant recognition Constructing a hybrid model of convolutional neural networks and long short-term memory networks: Multi-loss collaborative mechanism: Simultaneously optimize classification loss, regression loss and regularization loss, and balance model accuracy and generalization ability through dynamic weight adjustment.

[0038] Bayesian confidence calibration: By combining historical variant databases with disease association models, variant confidence indices are calculated and low-confidence results are filtered out, thereby improving the reliability of clinical decision-making.

[0039] Step 4: Closed-loop optimization and lightweight deployment Real-time feedback mechanism: Monitors recognition accuracy and automatically triggers model retraining, dynamically optimizing feature fusion weights and loss function parameters by combining gradient descent algorithm.

[0040] Fault tolerance: The K-nearest neighbor interpolation algorithm is activated for missing data, and the process continuity is ensured through redundant nodes and failover mechanisms.

[0041] Edge deployment: The model parameters are compressed through knowledge distillation, and the weights are quantized to an 8-bit integer format. The model can then achieve millisecond-level real-time recognition on mobile devices via an API interface and output an encrypted visualization report.

[0042] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0043] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fast variant recognition method based on multi-loss and multi-feature fusion, characterized in that: The method includes the following steps: S1. Collect multi-source gene sequence data, including genome sequencing data, epigenetic data, and transcriptome data, through biosensors; S2. Perform data preprocessing and feature extraction on the multi-source gene sequence data to generate a multi-dimensional feature vector set; S3. Input the multi-dimensional feature vector set into the feature fusion module for cross-modal fusion processing to generate a fused feature tensor; S4. Construct a multi-loss prediction model, perform variant recognition training based on the fused feature tensor, and generate variant classification parameters. S5. Perform variant type identification processing based on the variant classification parameters to generate variant identification results; S6. The variant identification results are executed through the output module to generate a visual report; S7. Real-time monitoring of variant recognition accuracy and feedback to step S1 form an adaptive optimization closed-loop mechanism.

2. The fast variant recognition method based on multi-loss and multi-feature fusion according to claim 1, characterized in that: The collection of multi-source gene sequence data in S1 includes the following steps: S11. Collect genome sequencing data using high-throughput sequencing equipment, generate raw reads using paired-end sequencing technology, and perform real-time quality filtering to remove low-confidence sequences. S12. Collect epigenetic data through methylation chips, generate a methylation level matrix, and apply a standardized protocol to correct for batch effects. S13. Collect transcriptome data using RNA sequencing equipment, extract gene expression feature vectors, and generate highly consistent expression profile data through normalization.

3. The fast variant recognition method based on multi-loss and multi-feature fusion according to claim 1, characterized in that: In step S2, data preprocessing and feature extraction are performed to generate a multi-dimensional feature vector set, including the following steps: S21. Perform quality filtering and noise reduction on the genome sequencing data to generate standardized read data, and use a sliding window algorithm to detect sequencing errors; The specific implementation steps of the sliding window algorithm are as follows: set a window of fixed length L, slide along the read segment sequence, and slide one base position at a time; Calculate the average quality score of all bases within the window. When the score is lower than a preset threshold, mark the bases in the window as low quality and correct or remove them. The formula for calculating the average quality score within the window is: ； in The average quality score of all bases within the window. The length of the window. This serves as an index for the bases within the window. The quality score of the i-th base within the window; S22. Principal component analysis is used to reduce the dimensionality of the methylation level matrix to generate a low-dimensional feature set, and the feature variance is calculated to retain the high-information dimension. S23. The gene expression feature vector is spatiotemporally aligned and fused with the low-dimensional feature set to generate a five-dimensional feature tensor, and then coordinate transformation is performed.

4. The fast variant recognition method based on multi-loss and multi-feature fusion according to claim 1, characterized in that: The generation of the fused feature tensor in S3 includes the following steps: S31. The feature fusion module is designed as a multi-head attention mechanism module, with the input layer dimension set to 256 dimensions and learnable weight parameters configured. S32. Input the five-dimensional feature tensor into the multi-head attention mechanism module for weight allocation processing, and integrate cross-modal information by calculating cross-attention scores; S33. Output the fused feature tensor, integrate cross-modal information through the attention weight matrix, and add residual connections.

5. The fast variant recognition method based on multi-loss and multi-feature fusion according to claim 1, characterized in that: The construction of the multi-loss prediction model in S4 includes the following steps: S41. Construct a multi-loss prediction model using a hybrid architecture of convolutional neural network and long short-term memory network, and embed batch normalization layer to accelerate convergence; S42. Define a set of multiple loss functions, including classification loss function, regression loss function and regularization loss function, and set dynamic weight coefficients; S43. Based on the fused feature tensor, train the model to generate variant classification parameters, wherein the classification loss function is calculated using the following formula: ； in For classification loss value, For the actual label vector, For the predicted probability vector, This represents the number of samples.

6. The fast variant recognition method based on multi-loss and multi-feature fusion according to claim 1, characterized in that: The variant type identification process in S5 includes the following steps: S51. Analyze variant type atlas data, establish a variant-disease association model, and integrate public databases; S52. Calculate the variant confidence index based on the variant classification parameters, and use the Bayesian method to quantify the uncertainty. S53. Generate a variant identification report matrix based on the association relationship model and confidence index, and filter low-confidence variants by thresholding.

7. The fast variant recognition method based on multi-loss and multi-feature fusion according to claim 1, characterized in that: The process of generating a visual report in S6 includes the following steps: S61. Encode the variant identification result into a standard BAM file format and add metadata tags; S62. Data transmission is secured by TLS protocol and transmitted to the terminal device via an encrypted cloud platform. S63. Generates real-time visual interactive reports, triggers model retraining when the recognition confidence level is below 90%, and records exception logs to optimize system resilience.

8. The fast variant recognition method based on multi-loss and multi-feature fusion according to claim 1, characterized in that: It also includes a real-time optimization mechanism: S81. Real-time monitoring of feature fusion efficiency; when the data dimension conflict rate exceeds the threshold, switch to backup fusion mode. S82. Employ gradient descent algorithm to dynamically adjust the weight parameters of multiple losses and update the coefficients based on validation set performance feedback. S83. Optimize the model architecture by matching similar scenarios based on historical variant databases, and integrate new data by applying an incremental learning mechanism.

9. The fast variant recognition method based on multi-loss and multi-feature fusion according to claim 1, characterized in that: It also includes an exception handling mechanism: S91. Detect the abnormal rate of data collection. When the missing rate exceeds 15%, activate the imputation algorithm and use multiple imputation techniques to fill in the missing values. S92. Use the K-nearest neighbor algorithm to fill in missing features and calculate sample similarity to select the best neighbor; S93. Initiate the continuity of the redundant computing node identification process and deploy a failover mechanism to automatically switch over when the master node fails.

10. The fast variant recognition method based on multi-loss and multi-feature fusion according to claim 1, characterized in that: It also includes a model lightweighting mechanism: S101. Perform knowledge distillation and compression on the multi-loss prediction model, and use the teacher-student framework to transfer key knowledge. S102. Generate a lightweight model parameter set and quantize the weights to 8-bit integers; S103, deployed to edge devices to achieve real-time variant recognition, and supports mobile applications through API interfaces.

Images