Semen sample quality management system based on federated learning and dynamic time series prediction

Abstract

The application discloses a semen sample quality management system based on federal learning and dynamic timing prediction, wherein a data acquisition module (100) is used for cleaning and processing the received original semen sample detection data, a federal learning cooperation module (200) adopts an improved Paillier homomorphic encryption algorithm to encrypt and transmit local model gradient parameters, a dynamic timing prediction module (300) extracts quality features with timing dependence from the semen sample detection data, and performs online incremental training on a prediction model, a multi-modal data fusion module (400) is used for extracting a genomic feature sequence of the semen sample and analyzing stored environmental parameters of the sample, and a quality evaluation decision module (500) fuses the timing prediction model and the genomic feature sequence.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of medical artificial intelligence and data collaboration, and specifically relates to a semen sample quality management system based on federated learning and dynamic time series prediction, which solves the cross-institutional data island problem and realizes private and secure sharing of multi-modal data and collaborative monitoring of sample quality. BACKGROUND

[0002] Semen sample quality evaluation is a key link in assisted reproductive technology. Traditional evaluation methods mainly rely on laboratory manual detection and single-dimensional static data analysis. In the prior art, some systems have attempted to use machine learning methods to predict the quality of semen samples. These systems usually train models based on centralized data, and evaluate by analyzing static indicators such as sperm motility and morphological characteristics. In addition, some improved schemes introduce time series data analysis methods to track the trend of sample quality changes using historical detection records, in an attempt to improve the accuracy of prediction. However, these methods are still limited to single-institution data accumulation and lack multi-center collaborative joint modeling mechanisms.

[0003] The prior art has obvious deficiencies. First, centralized data processing methods cannot meet the requirements of medical data privacy protection, and data isolation between medical institutions is serious, making it impossible to effectively share and collaboratively model. Second, traditional prediction models often ignore the coupling effects of sample genetic characteristics and environmental parameters, resulting in single evaluation dimensions. Third, static models are difficult to adapt to the dynamic characteristics of sample quality and lack incremental learning capabilities for real-time data. Finally, the quality evaluation threshold usually uses a fixed standard, which cannot be adaptively adjusted according to changes in data distribution, affecting the accuracy and reliability of the evaluation results. SUMMARY

[0004] This section aims to summarize some aspects of the embodiments of the application and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of the specification to avoid obscuring the purpose of this section, abstract and title, and such simplifications or omissions cannot be used to limit the scope of the application.

[0005] In view of the above and / or existing problems in the semen sample quality management system based on federated learning and dynamic time series prediction, the present application is proposed.

[0006] To solve the above technical problems, the present application provides the following technical solutions: In a first aspect, a semen sample quality management system based on federated learning and dynamic time series prediction includes,

[0007] The system includes a data acquisition module, a federated learning collaboration module, a dynamic time series prediction module, a multi-modal data fusion module, and a quality evaluation and decision-making module.

[0008] The data acquisition module is configured to clean, normalize and format the received raw semen sample detection data to generate standardized time-series detection data.

[0009] The federated learning collaboration module inputs the time-series detection data output by the data acquisition module,

[0010] The federated learning collaboration module comprises a homomorphic encryption unit and a dynamic weight distribution unit. The homomorphic encryption unit uses an improved Paillier homomorphic encryption algorithm to encrypt and transmit the local model gradient parameters. The dynamic weight distribution unit dynamically calculates the aggregation weight based on the semen data quality score and the data timeliness index through an adaptive weight distribution algorithm.

[0011] The federated learning module outputs the global model parameters to the dynamic time-series prediction module as initialization weights, and combines the time-series detection data provided by the data acquisition module to perform quality trend prediction.

[0012] The dynamic time-series prediction module comprises a time-series feature extraction unit and an incremental learning unit. The time-series feature extraction unit uses a long short-term memory network with a fusion attention mechanism to extract quality features with time-series dependencies from the detection data of the semen sample. The incremental learning unit performs online incremental training of the prediction model based on new sample data within a sliding time window.

[0013] The multi-modal data fusion module integrates a gene sequencing data analysis unit and an environmental parameter analysis unit. The gene sequencing data analysis unit is configured to extract genomic feature sequences of the semen sample, and the environmental parameter analysis unit is configured to analyze environmental parameters of the sample storage. The multi-modal data fusion module comprises a cross-modal alignment unit that uses a domain adaptive network based on adversarial training to align the distribution of the genomic feature sequences extracted by the gene sequencing data analysis unit and the environmental parameter feature vectors extracted by the environmental parameter analysis unit in a latent feature space.

[0014] The fusion feature vector of the multi-modal fusion module is input to the quality assessment module as one of the two inputs, and the residual output of the time-series prediction module and the genomic features are jointly input to the anomaly detection engine.

[0015] The dynamic time-series prediction module and the multi-modal data fusion module work collaboratively to extract features of the semen sample from different angles and provide comprehensive information for quality assessment.

[0016] The quality assessment decision module comprises an anomaly detection engine and a dynamic threshold adjustment unit. The anomaly detection engine identifies the sample quality abnormal state by fusing the time-series prediction model and the genomic feature sequence. The dynamic threshold adjustment unit dynamically adjusts the judgment threshold of the quality assessment using an adaptive algorithm.

[0017] As a preferred scheme of the semen sample quality management system based on federated learning and dynamic time series prediction of the application, wherein: the adaptive weight distribution algorithm adopted in the dynamic weight distribution unit.

[0018] As a preferred scheme of the semen sample quality management system based on federated learning and dynamic time series prediction of the application, wherein: the dynamic weight distribution unit is further configured with a weight smoothing unit, which adopts a moving average algorithm to process the historical weight sequence calculated according to the adaptive weight distribution algorithm, to suppress the drastic change of aggregated weights caused by single-round data fluctuations.

[0019] As a preferred scheme of the semen sample quality management system based on federated learning and dynamic time series prediction of the application, wherein the attention mechanism in the time series feature extraction unit is a multi-head self-attention mechanism, configured to capture time series dependencies in different subspaces from the hidden state sequence output by the long short-term memory network in parallel, and to concatenate and linearly transform the outputs of each sub-space to form the final quality feature vector.

[0020] As a preferred scheme of the semen sample quality management system based on federated learning and dynamic time series prediction of the application, wherein: the online incremental training process of the incremental learning unit includes an elastic weight consolidation unit that identifies the importance of model parameters by calculating the Fisher information matrix of the historical data, and applies constraints during incremental updates.

[0021] As a preferred scheme of the semen sample quality management system based on federated learning and dynamic time series prediction of the application, wherein: the multi-modal data fusion module further includes a cross-modal alignment unit that uses a domain adaptation network based on adversarial training to align the distribution of the genomic feature sequence extracted by the genetic sequencing data analysis unit and the environmental parameter feature vector extracted by the environmental parameter analysis unit in the latent feature space.

[0022] The multi-modal data fusion module combines LSTM neural networks and the Transformer framework.

[0023] As a preferred scheme of the semen sample quality management system based on federated learning and dynamic time series prediction of the application, wherein: the anomaly detection engine includes a cascaded detection architecture that uses the prediction residual output by the time series prediction model for preliminary statistical anomaly detection, and inputs the preliminary detection result together with the genomic feature sequence into a graph neural network that builds a feature similarity graph between samples to mine abnormal patterns and make a final comprehensive judgment.

[0024] As a preferred scheme of the semen sample quality management system based on federated learning and dynamic time series prediction, the adaptive algorithm adopted by the dynamic threshold adjustment unit is based on statistical process control theory, and the index weighted moving average value and the index weighted moving standard deviation of the quality evaluation result are monitored in real time to dynamically calculate and update the determination threshold.

[0025] In a second aspect, some embodiments of the present application provide an electronic device, comprising: one or more processors; a storage device having one or more programs stored thereon, when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any of the implementation manners of the first aspect.

[0026] In a third aspect, some embodiments of the present application provide a computer readable medium having a computer program stored thereon, wherein the program is executed by a processor to implement the method described in any of the implementation manners of the first aspect.

[0027] The present application realizes multi-center data collaborative modeling through a federated learning architecture, effectively utilizes semen sample data scattered everywhere while strictly protecting data privacy. The system uses a dynamic time series prediction method to accurately capture the law of sample quality change over time, combines multi-modal data fusion technology, and comprehensively considers the influence of genetic characteristics and environmental factors, significantly improving the comprehensiveness and accuracy of quality evaluation. The innovative adaptive weight allocation mechanism ensures that the model gives priority to high-quality and time-sensitive data during aggregation, while the dynamic threshold adjustment function enables the system to adaptively optimize the determination standard according to real-time data distribution. In addition, the system has online incremental learning capability, which can continuously optimize the prediction model and effectively adapt to changes in sample characteristics, providing a more intelligent and reliable solution for semen sample quality management. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings; among them:

[0029] Figure 1 The flow structure diagram of the semen sample quality management system based on federated learning and dynamic time series prediction in embodiment 1.

[0030] Figure 2 The quality feature vector output schematic diagram of the semen sample quality management system based on federated learning and dynamic time series prediction in embodiment 1. DETAILED DESCRIPTION

[0031] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0032] Referring to Figure 1 With Figure 2 For the first embodiment of the present application, the embodiment provides a semen sample quality management system based on federated learning and dynamic time series prediction, which comprises,

[0033] The data acquisition module includes a data preprocessing unit for cleaning, normalizing and formatting the received raw semen sample detection data to generate standardized input data. The original reference data is obtained by automatic detection and manual assisted analysis using a semen analyzer (such as SQA-V).

[0034] Semen detection is collected once per sample submission; genetic sequencing data is collected once for initial group entry or specific research needs, which is static data; for storage environment, temperature and humidity are recorded every 1 hour; for transportation process, time series data of temperature are recorded every 15 minutes. Through 2000 samples from each of the 4 institutions, a total of 8000 semen sample data are formed to build a federated learning model. The preprocessed data is packaged into structured records according to a unified mode and is assigned a unique sample ID; the structured records are sent to the federated learning collaboration module through a standardized communication protocol.

[0035] In order to construct a detailed specific embodiment, three types of core data are needed, each type of data contains multiple dimensions to fully reflect various factors affecting semen quality. The data dimensions cover the key indicators recommended in the fifth edition of the World Health Organization (WHO) standard.

[0036] Table 1: Semen sample detection data dimensions

[0037] Data Category Specific Dimension / Feature Unit / Type Data Source / Note Basic Information Patient ID (De-identified) String Anonymized Sample ID String Unique Identifier Collection Date / Time Datetime Abstinence Time Days Sample Type Categorical Fresh, Frozen, Washed, etc. Macro Parameters Semen Volume ml pH Value Dimensionless Appearance Categorical Normal / Abnormal Viscosity Categorical Normal / Abnormal Liquefaction Status Categorical Normal / Abnormal Sperm Concentration and Motility Sperm Concentration M / ml Core Indicators Total Motility % PR + NP Progressive Motility (PR) % Non-progressive Motility (NP) % Immotile (IM) % Sperm Average Motility Velocity μm / sec Morphology and Count Normal Forms % Strict Criteria Motile Sperm Concentration (MSC) M / ml Concentration x Total Motility Progressive Motility Sperm Concentration (PMSC) M / ml Concentration x PR Functional Sperm Concentration (FSC) M / ml Calculated from Morphology and Motility Sperm# M Concentration x Volume Others White Blood Cell Count M / ml Inflammation Indicators Round Cell Count M / ml Sperm Motility Index (SMI) Dimensionless Instrument-derived Indicators

[0038] Genomic information is crucial for assessing the intrinsic quality of sperm, particularly DNA integrity. This part of data can include single nucleotide polymorphism (SNP) information related to spermatogenesis, DNA repair, and antioxidant capacity.

[0039] Table 2: Genetic sequencing data dimensions

[0040] Data Category Specific Dimension / Feature Unit / Type Data Source / Note Basic Information Sample ID String Associated with Test Data Sequencing Platform Categorical Illumina, MGI, etc. Sequencing Depth X DNA Integrity-related DNA Fragmentation Index (DFI) % Obtained through SCSA or Comet Assay High DNA Staining (HDS) % Key Gene SNPs CYP1A1 rs4646903 Genotype (AA / AG / GG) Detoxification Metabolism-related GSTM1 Absence / Presence Categorical Antioxidation-related GSTT1 Absence / Presence Categorical Antioxidation-related MTHFR C677T (rs1801133) Genotype (CC / CT / TT) Folic Acid Metabolism, Affecting DNA Synthesis MTHFR A1298C (rs1801131) Genotype (AA / AC / CC) Folic Acid Metabolism P53 codon 72 (rs1042522) Genotype (Arg / Pro) Tumor suppression, DNA repair PARP1 rs1136410 Genotype (TT / TC / CC) DNA repair related XRCC1 rs25487 Genotype (GG / GA / AA) DNA repair related SOD2 rs4880 Genotype (TT / TC / CC) Antioxidant related CAT rs1001179 Genotype (CC / CT / TT) Antioxidant related

[0041] The storage and handling environment of the sample has a significant impact on its quality, especially in the context of cross-institutional transportation and long-term preservation. This part of data records the environmental conditions of the sample from collection to analysis throughout the process.

[0042] Table 3: Environmental parameter data dimensions

[0043] Data category Specific dimension / feature Unit / type Data source / remark Basic information Sample ID String Associated with test data Record timestamp Datetime Storage environment Storage temperature Celsius (°C) Ideal -196°C (liquid nitrogen) Storage humidity Percentage (%) Storage container type Categorical Liquid nitrogen tank, dry ice, etc. Storage location (institution ID) Categorical De-sensitization processing Transportation environment Transportation method Categorical Cold chain, room temperature Transportation duration Hours Temperature fluctuation during transportation Celsius (°C) Acquired by temperature recorder Processing environment Laboratory temperature Celsius (°C) Laboratory humidity Percentage (%) Processing operator ID (de-sensitized) String

[0044] The input of the federated learning collaboration module (200) is connected with the output of the data acquisition module (100), for receiving the preprocessed semen sample detection data from multiple institutions, and generating a global prediction model;

[0045] The federated learning collaboration module 200, the dynamic time series prediction module 300, the multi-modal data fusion module 400 and the quality evaluation decision module 500;

[0046] The federated learning collaboration module 200 includes a homomorphic encryption unit 201 and a dynamic weight distribution unit 202, the homomorphic encryption unit 201 uses an improved Paillier homomorphic encryption algorithm to encrypt and transmit the local model gradient parameters, and the dynamic weight distribution unit 202 dynamically calculates the aggregation weight through a self-adaptive weight distribution algorithm based on the semen data quality score and the data timeliness index;

[0047] The federated learning module outputs the global model parameters to the dynamic time series prediction module as the initialization weight, the fusion feature vector of the multi-modal fusion module is input to the quality evaluation module as one of the double inputs, and the residual output of the time series prediction module and the genomic features are jointly input to the anomaly detection engine.

[0048] The data acquisition module outputs the structured record M through a standardized interface unit,

[0049] ;

[0050] Wherein, ID is a unique sample identifier; InstID is an institution identifier; Q i is a quality score calculated for semen data, T i is a timeliness index calculated based on freshness, n∈[0,1]: n K The Kth normalized detection value;

[0051] The homomorphic encryption unit (201) takes (n1,…,n K ) in M as the local feature vector, and calculates the i-th institution local model gradient parameter ; the encryption function ;

[0052] The dynamic weight distribution unit (202) takes (Q i , T i ) in the same record M as input,

[0053] The dynamic timing prediction module 300 comprises a timing feature extraction unit 301 and an incremental learning unit 302. The timing feature extraction unit 301 adopts a long short-term memory network with a fusion attention mechanism to extract quality features with timing dependence from the detection data of the semen sample. The incremental learning unit 302 performs online incremental training on the prediction model based on new sample data within a sliding time window.

[0054] The homomorphic encryption unit 201 improves the calculation process of the Paillier homomorphic encryption algorithm by introducing a dynamic noise injection mechanism when encrypting the gradient parameters, and defines an encryption function :

[0055] ;

[0056] wherein the dynamic noise parameter ; ϵ is a noise coefficient, is the i-th local model gradient parameter of the institution; g, h, and n are public key parameters, ensuring that the encryption process has security and gradient effectiveness.

[0057] The dynamic weight allocation unit 202 calculates the adaptive dynamic weight allocation through double adjustment coefficient,

[0058] ;

[0059] wherein the exponential adjustment term enhances the timeliness influence, and the logarithmic term ln(1+Qi) alleviates the linear sensitivity of the quality score. Q i is the quality score of the semen data, T i is the timeliness index calculated based on the freshness of the data, α and β are the quality adjustment coefficient and the timeliness adjustment coefficient, respectively, λ is the freshness adjustment coefficient, and N is the total number of participating institutions. The quality score Q i ∈[0,1]; the timeliness T i ∈[0,100]; the adjustment coefficients α∈[0.3,0.7], β∈[0.2,0.5], and λ∈[0.01,0.1]; the weight is dynamically allocated according to the quality of the semen sample, such as integrity and noise level; and the timeliness, thereby optimizing the aggregation effect of federated learning.

[0060] The weight smoothing unit (203) adopts a sliding average algorithm to process the historical weight sequence in order to suppress the drastic change of the aggregation weight caused by the fluctuation of single-round data. Let the weight of the current round be , the weight of the previous round be , and the smoothed weight be wherein, gamma is a smoothing coefficient, the value range is (0, 1); the weight fluctuation is inhibited, the stability of the federated learning process is improved, the quality evaluation result of the semen sample is avoided, the weight fluctuation is inhibited through the exponential weighted moving average algorithm, and the stability of the federated learning process is improved. Reduce the weight fluctuation caused by data heterogeneity or local update, ensure the stable convergence of the global model in multiple iterations, avoid the influence of model performance caused by weight shock.

[0061] The dynamic time series prediction module (300) and the multi-modal data fusion module (400) work cooperatively, extract features of the semen sample from different angles, and provide comprehensive information for quality evaluation.

[0062] The dynamic time series prediction module 300 comprises a time series feature extraction unit 301, an incremental learning unit 302, and an elastic weight consolidation unit 303.

[0063] The time series feature extraction unit 301 adopts a hierarchical multi-head self-attention architecture:

[0064] ; the feature dimension d is in [128, 512]; the number of Transformer layers L is in [6, 12]; efficient extraction of semen genomic features is realized, and genetic defect screening and sperm quality evaluation are supported.

[0065] wherein Q is a query vector, K is a key vector, V is a value vector, is a linear projection of the LSTM hidden state H, d k is an attention dimension; multi-head output is integrated through a gating fusion mechanism:

[0066] ;

[0067] wherein, sigma is a sigmoid gating function, and Wg is a trainable parameter.

[0068] The elastic weight consolidation unit 303 is needed in the incremental learning unit, realizes the elastic constraint mechanism by introducing the Fisher information matrix constraint model as a regularization term,

[0069] ;

[0070] wherein, L total is the total loss function of incremental learning, L new is the loss value of the new sample, F k is the diagonal element of the Fisher information matrix, theta k is a model parameter, and tau is a regularization coefficient, realizing the balanced update of historical knowledge and new knowledge. The diagonal element F kThe importance of historical task contribution is to impose stronger regularization constraints on key parameters during updating, ensuring that the model retains the generalization ability for historical data while absorbing new data.

[0071] The multi-modal data fusion module 400 includes a cross-modal alignment unit 403, which uses a domain adaptation network based on adversarial training to align the distribution of the genomic feature sequence extracted by the genetic sequencing data analysis unit 401 and the environmental parameter feature vector extracted by the environmental parameter analysis unit 402 in the latent feature space.

[0072] The genetic sequencing data analysis unit 401 encodes the genomic sequence using the Transformer architecture after parsing the genetic sequencing data and placing it into the cross-modal alignment unit to generate an adversarial structure.

[0073] The environmental parameter analysis unit 402 generates an adversarial structure B by analyzing the distribution of the environmental parameter feature vector and placing it into the cross-modal alignment unit.

[0074] ;

[0075] where G is the generator, D is the discriminator, p G , p E are the distributions of the genomic features and environmental parameters, respectively, and joint training is achieved through a gradient reversal layer. This cross-modal alignment unit achieves cross-modal feature space alignment between genomic data and environmental parameters, improving the effectiveness of multi-modal data fusion. The generator (G) maps environmental parameter features to the genomic feature space, and the discriminator (D) distinguishes between real genomic features and mapped environmental features. Through adversarial training, the two feature distributions p G ≈p E are made to solve the fusion barrier caused by the distribution difference between cross-modal data, improve the collaborative representation ability of multi-modal features, achieve cross-modal alignment between genomic features and environmental parameters, and improve the robustness of semen quality evaluation.

[0076] By using a double-flow structure, LSTM is used to process time-series environmental parameters, and Transformer is used to process genomic sequences, and cross-attention mechanism is used for fusion:

[0077] Cross-attention function ;

[0078] where Z G , Z E are the feature vectors of the genomic and environmental parameters, respectively, and d is the feature dimension.

[0079] The quality assessment decision module (500) includes an anomaly detection engine (501) and a dynamic threshold adjustment unit (502). The anomaly detection engine (501) identifies abnormal sample quality by fusing a time-series prediction model with genomic feature sequences. The dynamic threshold adjustment unit (502) uses an adaptive algorithm to dynamically adjust the judgment threshold for quality assessment.

[0080] The anomaly detection engine employs a two-stage detection process.

[0081] In the first stage, based on predicted residuals Calculate the statistic ,when The initial exception was triggered at that time;

[0082] The second stage involves constructing a sample similarity graph.

[0083] Sample similarity matrix ; Propagate the anomaly score si=GNN(A) through a graph neural network. ij , F), when Confirm the abnormality in time. Let F be the cosine similarity function. i and F j This is the feature vector of the sample.

[0084] The dynamic threshold adjustment unit uses an exponentially weighted moving control limit:

[0085] Dynamic determination threshold Among them, the weighted average of the abnormal score index The square of the weighted standard deviation of the outlier score index , denoted as the mean of the outlier scores for the current batch of samples, and k is the control coefficient.

[0086] The quality assessment results are used to update the federated learning weight allocation strategy through backpropagation, forming a closed-loop optimization:

[0087] Next round of quality adjustment coefficient It is updated through a dynamic feedback mechanism; where MAEquality is the mean absolute error of the quality assessment results, and η is the federated learning rate.

[0088] The beneficial effects of this invention are as follows: By using a homomorphic encryption dynamic noise injection method and combining an improved Paillier algorithm with dynamic noise terms, the invention balances security and gradient validity when encrypting gradient parameters, preventing leakage of semen genomic data and meeting medical data privacy compliance requirements; through encrypted transmission and dynamic weight allocation, it enables collaborative modeling of semen sample data across institutions, avoids direct sharing of raw data, and protects patient privacy.

[0089] The exponential weighted moving average algorithm is used to suppress the fluctuation of federated learning aggregation weights, reduce the model oscillation caused by data heterogeneity, and ensure the stable convergence of global model. The Fisher information matrix constraint is used to update the parameters in incremental learning, balance the historical knowledge and new sample features, and prevent the model from forgetting key semen quality indicators such as sperm motility and DNA integrity.

[0090] The generative adversarial network realizes the space alignment of genomic features and environmental parameter features, solves the problem of modal distribution difference, and improves the dimension richness and accuracy of semen quality evaluation.

[0091] Based on the freshness index, the dynamic weight calculation is used to optimize the federated learning aggregation effect and adapt to the timeliness requirement of semen sample data. The new samples in the sliding time window are trained online, combined with elastic weight consolidation, to realize continuous model updating and adapt to the dynamic monitoring demand of semen quality. Through two-stage cascade detection and exponential weighted moving control limit adaptive adjustment of judgment threshold, the quality fluctuation characteristics of semen samples are adapted, and the false positive rate is reduced.

[0092] The present application significantly improves the security, stability, accuracy and dynamic adaptability of semen sample quality management through innovative dynamic noise encryption, double-regulation weight distribution, adversarial cross-modal alignment and modular connection mechanism, forms a unique and innovative whole-process technical solution, and is suitable for semen quality monitoring and evaluation requirements in medical institutions, reproductive health centers and other scenes.

[0093] Table 4: Federated learning weight calculation results

[0094] Parameter Institution A Institution B Institution C Institution D Data quality score Q i ]]> 0.9 0.85 0.75 0.95 Data timeliness indicator T i ]]> 1.0 0.8 0.6 1.0 Weighted sum 0.94 0.83 0.69 0.97 Aggregated weights w i ]] 0.253 0.223 0.186 0.338

[0095] Table 5: Time series prediction accuracy comparison (60-day average)

[0096] Indicator Control group (centralized) Experimental group (complete system) Experimental group (without attention mechanism) Experimental group (without incremental learning) Vitality prediction error (%) 4.2 1.8 3.1 3.5 Malformation rate prediction error (%) 2.8 0.9 2.1 2.4 Feature extraction time-consuming (ms) 156 189 142 165

[0097] Table 6: Multimodal data fusion effect

[0098] Evaluation index Control group (single data) Experimental group (multimodal fusion) Experimental group (without cross-modal alignment) Feature space distance 0.62 0.28 0.45 Fusion feature accuracy rate (%) 76.5 92.3 81.7 Data utilization rate (%) 65.0 98.2 82.5

[0099] Table 7: Anomaly detection performance

[0100] Indicator Control group (threshold method) Experimental group (cascade detection) Experimental group (without graph neural network) Abnormal identification accuracy rate (%) 72.3 95.6 83.1 False positive rate (%) 18.5 3.2 10.8 False negative rate (%) 15.2 1.8 8.3

[0101] Table 8: Model updating efficiency

[0102] Indicator Control group (full training) Experimental group (incremental learning) Experimental group (without elastic weight) Single update time-consuming (s) 45.2 8.7 9.1 Performance decay rate (%) 0 (full retraining) 2.3 8.5 Memory occupancy (GB) 8.6 3.2 3.4

[0103] Table 9: Quality evaluation stability

[0104] Indicator Control group (fixed threshold) Experimental group (dynamic threshold) Experimental group (without threshold adjustment) Threshold adaptation delay (days) - 1.2 - Evaluation consistency (%) 78.6 96.4 85.3 Qualified sample misjudgment rate (%) 12.3 2.1 7.8

[0105] From the table 4 of the federal learning weight calculation result, it can be seen that the high-quality and time-effective mechanism obtains higher weight through the dynamic weight algorithm, and the problem of "weight average allocation leading to low-quality data interference model" in the traditional federal learning is solved. The control group adopts fixed weight, and cannot distinguish the data value difference, which provides help for the subsequent model precision.

[0106] The table 5 of the time sequence prediction accuracy shows that the activity prediction error of the test group is only 42.9% of the control group, which benefits from the LSTM network with the fusion attention mechanism. The multi-head self-attention mechanism can capture the dependence relationship of different time scales, and the error of the test group without the mechanism increases to 3.1%. The effect of the incremental learning unit is also significant: the error of the test group without incremental learning is 3.5% because it cannot adapt to the new sample distribution, which verifies the necessity of the sliding window and the elastic weight consolidation.

[0107] The table 6 of the multi-modal data fusion effect highlights the creativity of cross-modal alignment: the test group reduces the spatial distance between gene features and environmental features to 0.28 through adversarial training, which is reduced by 37.8% compared with the test group without alignment mechanism, and the fusion feature accuracy is increased by 13.0%. The accuracy of the control group is only 76.5% because it only uses time sequence data and cannot associate genomic features, which proves that the application breaks through the limitation of "single data dimension analysis".

[0108] In the table 7 of the abnormality detection performance, the accuracy of the test group is increased to 95.6% by the cascade detection architecture, and the false positive rate is reduced to 3.2%. The graph neural network effectively identifies the "implicit abnormality" missed by the control group by mining the similarity between samples. The false negative rate of the test group without graph neural network increases to 8.3%, which shows that the traditional statistical method is difficult to capture complex correlation anomalies.

[0109] The table 8 of the model updating efficiency reflects the advantages of incremental learning: the time consumption of single update of the test group is only 19.2% of the full training of the control group, and the performance decay rate is much lower than that of the test group without elastic weight. The elastic weight consolidation unit protects the key parameters and avoids "catastrophic forgetting", solving the pain point of "new data covering old knowledge" in traditional incremental learning.

[0110] The table 9 of the quality evaluation stability shows that the dynamic threshold adjustment makes the evaluation consistency of the test group reach 96.4%, and the misjudgment rate of qualified samples is only 2.1%. When the systemic temperature fluctuation occurs on the 50th day, the dynamic threshold unit completes the threshold update within 1.2 days, while the misjudgment rate of the control group with fixed threshold increases to 12.3%, which proves that the application can adapt to systemic changes such as environmental fluctuations.

[0111] In summary, the present application realizes privacy protection and value aggregation of distributed data through federated learning, improves feature capture accuracy through dynamic time series prediction, breaks data silos through multi-modal fusion, and guarantees evaluation reliability through anomaly detection and dynamic threshold, thereby comprehensively solving the defects of traditional systems in data privacy, prediction accuracy, adaptability and the like, and having significant creativity and practical value.

[0112] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A semen sample quality management system based on federated learning and dynamic time series forecasting, characterized in that, The system comprises a data acquisition module (100), a federated learning collaboration module (200), a dynamic time series prediction module (300), a multi-modal data fusion module (400) and a quality evaluation decision module (500); The data acquisition module (100) is configured to clean, normalize and format the received raw semen sample detection data to generate standardized time series detection data; The federated learning collaboration module (200) inputs the time series detection data output by the data acquisition module, The federated learning collaboration module (200) comprises a homomorphic encryption unit (201) and a dynamic weight distribution unit (202), the homomorphic encryption unit (201) uses an improved Paillier homomorphic encryption algorithm to encrypt and transmit the local model gradient parameters, and the dynamic weight distribution unit (202) dynamically calculates the aggregation weight based on the semen data quality score and the data timeliness index through an adaptive weight distribution algorithm; The federated learning module outputs global model parameters to the dynamic time series prediction module as initialization weights, and combines the time series detection data provided by the data acquisition module to predict the quality trend; The dynamic time series prediction module (300) comprises a time series feature extraction unit (301) and an incremental learning unit (302), the time series feature extraction unit (301) uses a long short-term memory network with a fusion attention mechanism to extract quality features with time series dependence from the detection data of the semen sample, and the incremental learning unit (302) performs online incremental training on the prediction model based on new sample data in a sliding time window; The multi-modal data fusion module (400) integrates a gene sequencing data analysis unit (401) and an environmental parameter analysis unit (402), the gene sequencing data analysis unit (401) is configured to extract genomic feature sequences of the semen sample, and the environmental parameter analysis unit (402) is configured to analyze environmental parameters of the sample storage; the multi-modal data fusion module (400) comprises a cross-modal alignment unit (403), which uses a domain adaptive network based on adversarial training to align the distribution of the genomic feature sequences extracted by the gene sequencing data analysis unit (401) and the environmental parameter feature vectors extracted by the environmental parameter analysis unit (402) in a latent feature space; The fusion feature vector of the multi-modal fusion module is input into the quality evaluation module as one of the dual inputs, and the residual output of the time series prediction module and the genomic features are jointly input into the anomaly detection engine; The dynamic time series prediction module (300) and the multi-modal data fusion module (400) work collaboratively to extract features of the semen sample from different angles and provide comprehensive information for quality evaluation, The quality evaluation decision module (500) comprises an anomaly detection engine (501) and a dynamic threshold adjustment unit (502), the anomaly detection engine (501) identifies the sample quality abnormal state by fusing the time series prediction model and the genomic feature sequence, and the dynamic threshold adjustment unit (502) dynamically adjusts the judgment threshold of the quality evaluation by using an adaptive algorithm.

2. The semen sample quality management system based on federated learning and dynamic timing prediction according to claim 1, characterized in that, The adaptive weight allocation algorithm employed in the dynamic weight allocation unit (202). 3.The semen sample quality management system based on federated learning and dynamic timing prediction of claim 1, wherein, The dynamic weight allocation unit (202) is further configured with a weight smoothing unit (203) that employs a moving average algorithm to process the historical weight sequence calculated according to the adaptive weight allocation algorithm, for suppressing the drastic change of the aggregated weight caused by single-round data fluctuation. 4.The semen sample quality management system based on federated learning and dynamic timing prediction of claim 1, wherein, The attention mechanism in the time series feature extraction unit (301) is a multi-head self-attention mechanism, configured to capture the time series dependency in different subspaces from the hidden state sequence output by the long short-term memory network in parallel, and concatenate and linearly transform the outputs of each subspace to form the final quality feature vector. 5.The semen sample quality management system based on federated learning and dynamic timing prediction of claim 1, wherein, The online incremental training process of the incremental learning unit (302) includes an elastic weight consolidation unit (303) that identifies the importance of the model parameters by calculating the Fisher information matrix of the parameters on the historical data, and applies constraints during incremental update.

6. The semen sample quality management system based on federated learning and dynamic timing prediction according to claim 1, characterized in that, The multi-modal data fusion module (400) further includes a cross-modal alignment unit (403) that employs a domain adaptation network based on adversarial training to align the distribution of the genomic feature sequence extracted by the genomic sequencing data analysis unit (401) and the environmental parameter feature vector extracted by the environmental parameter analysis unit (402) in the latent feature space. The multi-modal data fusion module (400) combines LSTM neural network and Transformer framework.

7. The semen sample quality management system based on federated learning and dynamic time series prediction of claim 1, wherein, The anomaly detection engine (501) includes a cascaded detection architecture that performs preliminary statistical anomaly detection using the prediction residual output by the time series prediction model, and inputs the preliminary detection result together with the genomic feature sequence into a graph neural network that constructs a feature similarity graph between samples to mine abnormal patterns and make a final comprehensive determination.

8. The semen sample quality management system based on federated learning and dynamic timing prediction according to claim 1, characterized in that, The dynamic threshold adjustment unit (502) employs an adaptive algorithm based on statistical process control theory to dynamically calculate and update the determination threshold by real-time monitoring the exponentially weighted moving average value and exponentially weighted moving standard deviation of the quality evaluation result.

9. An electronic device, characterized in that... comprising: one or more processors; a memory device having stored thereon one or more programs; when the one or more programs are executed by the one or more processors, cause the one or more processors to carry out the system of any one of claims 1-8.

10. A computer-readable storage medium having stored thereon executable instructions that cause a processor-based system to perform steps comprising The instructions, when executed by the processor, cause the processor to carry out the system of any one of claims 1-8.

Images