A product data verification method and system based on big data analysis

Abstract

This invention discloses a product data verification method and system based on big data analysis. The method includes: performing initial verification to obtain initial verification data; performing anomaly detection and determining the verification time period and verification evaluation data based on the anomaly detection results; extracting data centers for products in the same category; calculating the centrality of each verification evaluation data point to obtain a data center degree vector; calculating the confidence of each verification evaluation data point based on a time-series dynamic confidence decay model to obtain a confidence matrix; extracting and normalizing user feedback data to obtain a user feedback vector; calculating the correlation between each verification evaluation data point and the remaining data and calculating the correlation deviation vector; constructing a semi-supervised adversarial verification prediction model; and inputting the product data to be verified into the semi-supervised adversarial verification prediction model to obtain the verification results. This method not only improves the efficiency and accuracy of product data verification but also has good interpretability and can be directly applied to product data verification systems.

Description

Technical Field

[0001] This invention relates to the field of data verification technology, and in particular to a product data verification method and system based on big data analysis. Background Technology

[0002] In today's digital economy, the production, distribution, and sales of products constitute a complex and data-intensive process chain. Accurate verification of product lifecycle data is a crucial link in ensuring product quality, tracing the root causes of problems, combating counterfeiting, and optimizing supply chain management. It is also of paramount importance for maintaining market order, protecting consumer rights, and enhancing corporate brand reputation.

[0003] Traditional product data verification methods mainly rely on static rule checks of a single data source or simple analysis based on isolated anomaly detection algorithms, which has obvious limitations: First, they lack collaborative analysis and correlation verification of data across the entire "production-logistics-sales" chain, resulting in data silos at each stage and making it difficult to detect cross-domain collaborative fraud. Second, the verification process is mostly static and reactive, unable to dynamically adjust the focus and granularity of verification based on data flow anomalies, leading to wasted computational resources and poor adaptability to new and complex fraud patterns. Finally, traditional methods rely on a large amount of precisely labeled data for fully supervised learning. In actual business, obtaining a large amount of labeled data is costly, and the model's generalization ability for unseen new anomaly patterns is insufficient, resulting in high false positive and false negative rates. This invention proposes a product data verification method and system based on big data analytics. By collecting data from the production, logistics, and sales chains and performing collaborative verification and hardware-level fingerprint auditing, a data authenticity baseline is constructed. A multi-sphere anomaly detection model is then used to dynamically define the verification time period, achieving precise time window extraction. Furthermore, a data density and a four-dimensional time-series dynamic confidence matrix are introduced to quantify data dispersion, process stability, and credibility evolution. Simultaneously, user feedback vectors and cross-node information entropy correlation analysis are integrated to construct a correlation deviation vector to capture anomaly propagation paths. Finally, based on a small amount of blockchain-verified verification annotations and massive amounts of unlabeled data, a semi-supervised adversarial graph verification network is constructed. Through generator-discriminator game theory and hard-negative sample comparative learning, high-precision verification results and auditable anomaly root cause paths are output. This constructs a closed-loop framework from data collection to intelligent decision-making, significantly reducing annotation dependence and achieving efficient, accurate, and adaptive verification of product data, thereby effectively improving the automation level and reliability of the verification system. Summary of the Invention

[0004] The purpose of this invention is to provide a product data verification method and system based on big data analysis.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution:

[0006] This invention includes the following steps:

[0007] The product process chain data is collected for initial verification to obtain initial verification data; the product process chain data includes production chain data, logistics chain data, and sales chain data; the initial verification includes collaborative verification and fingerprint auditing;

[0008] Anomaly detection is performed on the primary verification data. The verification time period is determined based on the anomaly detection results. All primary verification data within the verification time period is extracted as verification evaluation data. The anomaly detection results include anomaly score, abnormal data volume, and abnormal timestamp.

[0009] Extract the data centers of similar products, calculate the centrality of each verification and evaluation data to obtain the data center degree vector, and calculate the confidence of each verification and evaluation data according to the time-series dynamic confidence decay model to obtain the confidence matrix;

[0010] Extract user feedback data and normalize it to obtain a user feedback vector. Calculate the correlation between each verification and evaluation data and the remaining data to obtain a correlation vector. Calculate the correlation deviation vector based on the correlation vector of similar products.

[0011] A semi-supervised adversarial verification prediction model is constructed based on historical verification evaluation indicators and verification labels. The verification results are obtained by inputting the product data to be verified into the semi-supervised adversarial verification prediction model. The verification evaluation indicators include anomaly score, centrality vector, confidence matrix, user feedback vector, and correlation deviation vector.

[0012] Furthermore, the method for obtaining the initial verification data includes:

[0013] Collect product process chain data, format all data in a unified manner, align with timestamps and then perform entity parsing. Store the preprocessed product process chain data in a buffer for verification.

[0014] The client for federated learning is determined, the coordination server publishes a unified verification model architecture, and distributes the initialized model parameters to each client. Each client trains a local verification model using local process chain data and inputs it to the coordination server. The coordination server generates global model parameters based on the local model parameters and returns them to each client to configure the global verification model.

[0015] A global verification model is used to collaboratively verify the product process chain data to obtain preliminary verification results, and the preliminary verification results of each client are integrated to obtain a collaborative verification conclusion.

[0016] The encrypted hash value of each product process chain data unit is calculated as a data fingerprint. Each data provider uses its private key to sign the data fingerprint to generate a digital signature. The data fingerprint and digital signature are packaged into a transaction. Multiple transactions are aggregated into a Merkle root through a Merkle tree to construct the Light-Chain blockchain. The Light-Chain blockchain includes a production blockchain, a logistics blockchain, and a sales blockchain. The Light-Chain blockchain is composed of multiple trusted nodes. The block header contains the Merkle root, timestamp, previous block hash, and consensus information. The block body stores the transaction list.

[0017] Fingerprint auditing is performed on the Light-Chain blockchain corresponding to each product process chain. The specific steps include: querying the Merkel proof of the relevant data fingerprint from the Light-Chain blockchain, verifying the fingerprint consistency on the Light-Chain lightweight chain corresponding to different process chains, and combining the Merkel proof and fingerprint consistency results into fingerprint audit results.

[0018] Product process chain data that has been verified through collaborative validation and fingerprint auditing will be used as primary verification data.

[0019] Furthermore, the method for determining the verification time period includes:

[0020] A multidimensional spherical model is constructed based on the data units of each key operation node in each product process chain; the types of multidimensional spherical models include production node spheres, logistics node spheres, and sales node spheres; the center of the multidimensional spherical model is taken as the mean vector of each dimension feature of the corresponding key operation node in the process chain, and the radius of the sphere is determined based on Mahalanobis distance;

[0021] The initial verification data is input into the multidimensional sphere model according to the process chain and corresponding operation nodes for anomaly detection. The Euclidean distance from the abnormal data in the continuous time series to the center of the sphere is calculated. The difference between the Euclidean distance and the radius of the sphere is calculated to obtain the anomaly distance. The maximum value of the ratio of the anomaly distance to the radius of the sphere in the continuous time series is taken as the anomaly score. The timestamp of the corresponding time series is taken as the anomaly timestamp.

[0022] Calculate the anomaly density within the data anomaly detection time period, and determine the dynamic verification length based on the anomaly density and the standard verification length; the anomaly density is taken as the ratio of the amount of abnormal data to the anomaly detection window length; the dynamic verification length is positively correlated with the anomaly density.

[0023] The abnormal timestamp is used as the midpoint of the verification period, and the verification period is determined based on the dynamic verification length.

[0024] Furthermore, the method for obtaining the data center degree vector includes:

[0025] Based on historical verification and evaluation data of similar products, cluster analysis was used to determine representative data centers for similar products. A weighted average method was then used to calculate the representative data centers for all similar products to obtain the data center data center data center data. ;

[0026] The expressions for distribution density centrality, statistical characteristic centrality, and volatility stability centrality, calculated based on data center and verification evaluation data, are as follows:

[0027] ;

[0028] ;

[0029] ;

[0030] ;

[0031] in Density centrality, Centrality is a statistical characteristic. For the wave stability centrality, Frequency weighting For range weights, To determine the frequency of occurrence of verification and evaluation data that are completely identical to the central data, This represents the total number of verification and evaluation data. For located The amount of data within the range For located Data volume outside the scope To aggregate weights, To verify the mean of the evaluation data, For discrete weights, To verify the standard deviation of the evaluation data, Weights for distribution shape To verify the skewness of the evaluation data, For data window weights, For fluctuation weights, The global standard deviation, For rolling window volatility, To verify the window length, For data points The value, Extreme value weights;

[0032] The centrality of the verification and evaluation data is obtained by weighting the distribution density centrality, statistical characteristic centrality, and fluctuation stability centrality. The centrality vector of the verification and evaluation data of each key operation node in each product process chain is composed of the centrality of the data.

[0033] Furthermore, the method for obtaining the confidence matrix includes:

[0034] The initial confidence level of the product process chain data at each operational node in each process chain during the collaborative verification period is used as the initial confidence level. The confidence level of the verification evaluation data for each operational node in each process chain is then calculated using a time-series dynamic confidence decay model. The expression is as follows:

[0035] ;

[0036] ;

[0037] in for Operation node verification and evaluation data in Confidence level at any given moment This is the initial confidence level. for Operation node verification and evaluation data in The temporal decay coefficient at time step, This is the initial time of the verification period. For the Sigmoid function, , As a weight for process stability, Within the verification period All data in the operation node is located in the corresponding The amount of data within the range For the corresponding Data volume outside the scope Within the verification period All data volume of the operation node. This represents the proportion of samples whose time-sensitivity distance to the corresponding data center exceeds a dynamic distance threshold. To verify the event impact value, including anomaly detection events and user feedback events, To verify the number of verification events within the specified time period. The moment when the verification event occurs. The event affects the half-life. Anchor indicator functions for blockchain events to ensure that events are non-repudiable. Based on the attenuation coefficient, The fluctuation sensitivity coefficient This represents time-series fluctuations, reflecting recent data volatility.

[0038] The expressions for the timeliness distance of the sample data to the corresponding data center and the dynamic distance threshold are as follows:

[0039] ;

[0040] ;

[0041] in For sample data To the corresponding data center Timeliness distance, The aging factor is... For sample data timestamp, For dynamic distance threshold, This is the initial distance threshold;

[0042] Within the verification period, the confidence scores of the verification evaluation data of the same operation node are arranged in chronological order to form a confidence vector;

[0043] The confidence vectors of each operation node are combined into a confidence matrix according to the process chain sequence.

[0044] Furthermore, the method for calculating the correlation deviation vector includes:

[0045] Calculate the probability distribution of verification and evaluation data across different dimensions within the same operation node to obtain the information entropy of that operation node. Based on the node's information entropy, calculate the mutual information between two operation nodes and perform temporal enhancement to obtain temporal mutual information. Calculate the mean conditional information entropy of a single operation node with all other operation nodes and perform multi-timescale fusion to obtain the enhanced correlation degree of the corresponding operation node. The expression is:

[0046] ;

[0047] ;

[0048] in For operation nodes and Timing-enhanced mutual information , For the data set of the corresponding operation node, For operation nodes and The length of the time period, For operation nodes exist Time period and operation node exist Inter-time information For operation nodes exist Information entropy over a period of time For operation nodes Enhance the correlation, , For time fusion weights, This represents the number of operation nodes. For operation nodes and During the same period Standardized mutual information, For operation nodes and During the same period mutual information, The longest time span. For operation nodes Timing mutual information with subsequent operation nodes;

[0049] The correlation vector is obtained by combining the enhanced correlation of each operation node according to the process chain sequence.

[0050] Extract the process chain data of qualified products in the same category, calculate the standardized mutual information of each operation node in the same time period to obtain the correlation vector of products in the same category, and average and weight all correlation vectors of products in the same category to obtain the correlation fusion vector of products in the same category. Calculate the deviation between the correlation vector of the product to be verified and the corresponding correlation fusion vector of products in the same category to obtain the correlation deviation vector.

[0051] Furthermore, the method for constructing a verification prediction model based on semi-supervised adversarial methods includes:

[0052] The first training set is formed by obtaining the verification labels corresponding to the historical verification evaluation indicators. The historical verification evaluation indicators without verification labels are actively learned and labeled to generate pseudo labels. The pseudo labels and the corresponding historical verification evaluation indicators are then combined to form the second training set. The verification labels include the verification pass status and the factors that affect the verification results.

[0053] The first and second training sets are randomly mixed to form a comprehensive set. The comprehensive set is then randomly divided into a training set and a test set in a 7:4 ratio to train and validate the verification prediction model based on semi-supervised adversarial methods. The verification prediction model based on semi-supervised adversarial methods includes an input layer, a feature processing layer, a shared encoding layer, a multi-task prediction layer, an adversarial discriminator, a feature sampler, and an output layer.

[0054] The feature processing layer uses a fully connected layer, BatchNorm, Dropout, and activation functions to perform feature normalization, missing value processing, and feature selection on the verification evaluation indicators to obtain standardized feature vectors.

[0055] The shared coding layer learns deep representations of the normalized feature vectors through a multilayer perceptron and residual connections to obtain coded features;

[0056] The multi-task prediction layer includes a verification predictor and a factor predictor; the verification predictor uses a BP neural network to perform regression prediction on the encoded features to obtain the verification pass rate and determine the verification pass result; the factor predictor performs multi-label classification on the encoded features and determines the factors that affect the verification result.

[0057] The adversarial discriminator is connected to the shared coding layer, inputs the encoded features into the lightweight discriminant network to distinguish the feature distribution of real labeled samples and pseudo-labeled samples and outputs the sample source probability;

[0058] The output of the feature sampler connected to the adversarial discriminator clearly marks the source of the samples, and positive and negative sample pairs are selected to train the multi-task prediction layer based on feature similarity and business rules;

[0059] The adversarial discriminator and the feature sampler are used only for training the multi-task prediction layer;

[0060] The output layer is connected to the multi-task prediction layer through a fully connected layer to output the verification result; the verification result includes the verification pass status and the factors that affect the verification result.

[0061] The semi-supervised adversarial verification prediction model is trained using a multi-task loss function, expressed as follows:

[0062] ;

[0063] ;

[0064] ;

[0065] ;

[0066] ;

[0067] ;

[0068] in For multi-task loss function, , , , , As task weight, To verify the predicted losses, To predict losses based on factors, To combat the losses, To combat the losses, For consistency loss, This represents the number of truly labeled samples. The number of pseudo-labeled samples. For the first Individual sample verification result label, For the first The probability that a sample passes verification. To verify the total number of influencing factor categories, For the first The verification results of the first sample are subject to the first Labels influenced by several factors For the first The verification results of the first sample are subject to the first The probability of being affected by each factor. To obtain data from the distribution of real labeled data Medium-sized samples The mathematical expectation, To distribute pseudo-labeled data Medium-sized samples The mathematical expectation, For the discriminator function, For feature extraction function, To counteract the weighting coefficient, For the set of positive sample pairs The number of corresponding sample pairs, For positive sample pairs negative sample set The number of corresponding sample pairs, For anchor point samples eigenvectors, Positive samples eigenvectors, negative samples eigenvectors, The cosine similarity function is used. For boundary parameters, For the original sample The verification prediction probability, These are the disturbance parameters;

[0069] The supervised adversarial verification prediction model employs a multi-stage training strategy for adversarial training, including supervised pre-training, semi-supervised adversarial training, and consistency refinement. The supervised pre-training optimizes the verification prediction loss using only real labeled data. Factors predicting loss The semi-supervised adversarial training uses real labeled data to optimize the multi-task loss function. Using pseudo-labeled data to optimize and mitigate loss and consistency loss The consistency refining focuses on optimizing consistency loss. ;

[0070] The product data to be verified is input into a semi-supervised adversarial graph verification network to obtain the verification results.

[0071] Secondly, a product data verification system based on big data analytics includes:

[0072] The primary verification module is used to configure a global verification model through federated learning, and to use the global verification model to collaboratively verify the product process chain data to obtain collaborative verification conclusions; it is also used to build a Light-Chain blockchain, perform fingerprint audits on the Light-Chain blockchain corresponding to each product process chain, and obtain fingerprint audit results.

[0073] Anomaly detection module: Used to detect anomalies in primary verification data, determine the verification time period based on the anomaly detection results, and extract all primary verification data within the verification time period as verification evaluation data;

[0074] Verification and evaluation index calculation module: used to calculate the centrality of each verification and evaluation data to obtain the centrality vector; used to calculate the confidence of each verification and evaluation data according to the time series dynamic confidence decay model to obtain the confidence matrix; used to extract user feedback data and normalize it to obtain the user feedback vector; used to calculate the correlation deviation vector.

[0075] Verification prediction module: This module is used to build a semi-supervised adversarial verification prediction model based on historical verification evaluation indicators and verification labels. The verification results are obtained by inputting the product data to be verified into the semi-supervised adversarial verification prediction model.

[0076] Management module: Used to store, manage, and view Light-Chain blockchain and verification results, and to process data based on the verification results.

[0077] The beneficial effects of this invention are:

[0078] This invention is a product data verification method and system based on big data analysis. Compared with existing technologies, this invention has the following technical advantages:

[0079] This invention enhances data preprocessing capabilities and improves model adaptability in product data verification through collaborative verification, fingerprint auditing, anomaly detection, construction of centrality vectors and confidence matrices, and model building steps. It improves the efficiency and accuracy of product data verification, optimizes the verification technology, significantly saves resources, increases work efficiency, and provides more reliable technical support for product data verification. Through semi-supervised learning, it greatly reduces reliance on labeled data and lowers model training costs, constructing a closed-loop framework from data collection to intelligent decision-making. This enables efficient, accurate, and adaptive verification of product data, effectively improving the automation level and reliability of the verification system. Attached Figure Description

[0080] Figure 1 This is a flowchart illustrating the steps of a product data verification method based on big data analysis according to the present invention. Detailed Implementation

[0081] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0082] The present invention discloses a product data verification method and system based on big data analysis, comprising the following steps:

[0083] like Figure 1 As shown, this embodiment includes the following steps:

[0084] The product process chain data is collected for initial verification to obtain initial verification data; the product process chain data includes production chain data, logistics chain data, and sales chain data; the initial verification includes collaborative verification and fingerprint auditing;

[0085] Anomaly detection is performed on the primary verification data. The verification time period is determined based on the anomaly detection results. All primary verification data within the verification time period is extracted as verification evaluation data. The anomaly detection results include anomaly score, abnormal data volume, and abnormal timestamp.

[0086] Extract the data centers of similar products, calculate the centrality of each verification and evaluation data to obtain the data center degree vector, and calculate the confidence of each verification and evaluation data according to the time-series dynamic confidence decay model to obtain the confidence matrix;

[0087] Extract user feedback data and normalize it to obtain a user feedback vector. Calculate the correlation between each verification and evaluation data and the remaining data to obtain a correlation vector. Calculate the correlation deviation vector based on the correlation vector of similar products.

[0088] A semi-supervised adversarial verification prediction model is constructed based on historical verification evaluation indicators and verification labels. The verification results are obtained by inputting the product data to be verified into the semi-supervised adversarial verification prediction model. The verification evaluation indicators include anomaly score, centrality vector, confidence matrix, user feedback vector, and correlation deviation vector.

[0089] In this embodiment, the method for obtaining primary verification data includes:

[0090] Collect product process chain data, format all data in a unified manner, align with timestamps and then perform entity parsing. Store the preprocessed product process chain data in a buffer for verification.

[0091] The client for federated learning is determined, the coordination server publishes a unified verification model architecture, and distributes the initialized model parameters to each client. Each client trains a local verification model using local process chain data and inputs it to the coordination server. The coordination server generates global model parameters based on the local model parameters and returns them to each client to configure the global verification model.

[0092] A global verification model is used to collaboratively verify the product process chain data to obtain preliminary verification results, and the preliminary verification results of each client are integrated to obtain a collaborative verification conclusion.

[0093] The encrypted hash value of each product process chain data unit is calculated as a data fingerprint. Each data provider uses its private key to sign the data fingerprint to generate a digital signature. The data fingerprint and digital signature are packaged into a transaction. Multiple transactions are aggregated into a Merkle root through a Merkle tree to construct the Light-Chain blockchain. The Light-Chain blockchain includes a production blockchain, a logistics blockchain, and a sales blockchain. The Light-Chain blockchain is composed of multiple trusted nodes. The block header contains the Merkle root, timestamp, previous block hash, and consensus information. The block body stores the transaction list.

[0094] Fingerprint auditing is performed on the Light-Chain blockchain corresponding to each product process chain. The specific steps include: querying the Merkel proof of the relevant data fingerprint from the Light-Chain blockchain, verifying the fingerprint consistency on the Light-Chain lightweight chain corresponding to different process chains, and combining the Merkel proof and fingerprint consistency results into fingerprint audit results.

[0095] Product process chain data that has undergone collaborative verification and fingerprint auditing will be used as primary verification data.

[0096] In actual evaluation, the production chain data (raw material batch number, production process record, quality inspection report, equipment operating status, etc.) of the manufacturing enterprise through the ERP (Enterprise Resource Planning) system, MES (Manufacturing Execution System) and IoT sensor are desensitized locally. Sensitive formula parameters are converted into ciphertext using homomorphic encryption algorithm, and non-sensitive fields are formatted to form a structured message body. The data is extracted in real time through the API gateway and transmitted to the client using TLS encryption.

[0097] Logistics chain data (warehousing and warehousing time, transportation trajectory coordinates, environmental monitoring values, transportation containers, and signed receipts) are collected from the logistics company's TMS, GPS tracking devices, RFID tags, and temperature and humidity sensors. Among them, the transportation trajectory coordinates are processed with differential privacy and Laplace noise is added.

[0098] Data is collected from e-commerce platforms, POS systems, and CRM systems (order ID, payment records, inventory changes, promotional activities, sales time, location, barcode scanner IMEI, anonymous consumer identifiers, and user reviews), and integrated in real time through the Flink stream processing platform;

[0099] All data is uniformly formatted using a JSON-LD-based semantic model, and the time of each system is synchronized using the NTP protocol. Through fuzzy matching and a rule engine, the same product identifier (serial number) in different data sources is associated to form a complete process chain view. The buffer retention policy is to save only 7 days of raw data.

[0100] Following the data generation process, manufacturers, logistics providers, and retailers are treated as clients. Each client generates an RSA asymmetric encryption key pair; the public key is uploaded to the server, while the private key is stored locally. A lightweight neural network serves as the unified verification model architecture. During training, clients periodically generate zero-knowledge proofs to demonstrate the correctness of the training process to the server without disclosing data. After training, clients encrypt the model update values ​​using homomorphic encryption (Paillier algorithm) and upload the ciphertext to the coordination server. The coordination server collects the local model parameters from each client and aggregates them in the ciphertext space using a weighted average algorithm (FedAvg) to generate global model parameters, ensuring that the original local model parameters are not decrypted. The coordination server distributes the updated global model parameters to each client, which decrypts and updates its local model.

[0101] When new product data flows in, each client uses a global model to perform local inference (inferring data from each node of the production chain, logistics chain, and sales chain separately), outputs preliminary verification results (consistency scores of data from each node of the production chain, logistics chain, and sales chain), and integrates them through a secure multi-party computation protocol (secret sharing) to form a final collaborative verification conclusion (verification pass status of production chain data / logistics chain data / sales chain data).

[0102] For each data unit in the process chain (corresponding to a node, such as a production event or logistics record), input the data content, timestamp, and data source identifier into SHA-256 to obtain the data fingerprint;

[0103] Each product's process chain data includes three sets of Light-Chain blockchains. The Light-Chain blockchains use a proof-of-authority consensus algorithm, where authorized nodes take turns producing blocks. The block-producing nodes verify the validity of the transaction signatures and quickly reach consensus through a lightweight voting mechanism. Once confirmed, the blocks are broadcast to the entire network, and the data fingerprints are permanently recorded. The Light-Chain blockchains only store Merkle roots (data fingerprints and digital signatures) and metadata; the original data is kept on the local system.

[0104] During the initial verification phase, the system automatically retrieves the Merkel proof of the relevant data fingerprint from the Light-Chain. By verifying the Merkel path and digital signature, it confirms whether the data has been tampered with. For multi-process chain data (such as production to sales), the consistency of fingerprints on different chains is verified through a lightweight cross-chain protocol (hash locking) to ensure the trustworthiness of the entire process. Auditors can enter the data identifier through a blockchain explorer to retrieve historical fingerprint records and generate an audit report.

[0105] When the fingerprints of the multi-process chain data are consistent through Merkel's proof, the corresponding product process chain data is deemed to have passed fingerprint auditing.

[0106] In this embodiment, the method for determining the verification time period includes:

[0107] A multidimensional spherical model is constructed based on the data units of each key operation node in each product process chain; the types of multidimensional spherical models include production node spheres, logistics node spheres, and sales node spheres; the center of the multidimensional spherical model is taken as the mean vector of each dimension feature of the corresponding key operation node in the process chain, and the radius of the sphere is determined based on Mahalanobis distance;

[0108] The initial verification data is input into the multidimensional sphere model according to the process chain and corresponding operation nodes for anomaly detection. The Euclidean distance from the abnormal data in the continuous time series to the center of the sphere is calculated. The difference between the Euclidean distance and the radius of the sphere is calculated to obtain the anomaly distance. The maximum value of the ratio of the anomaly distance to the radius of the sphere in the continuous time series is taken as the anomaly score. The timestamp of the corresponding time series is taken as the anomaly timestamp.

[0109] Calculate the anomaly density within the data anomaly detection time period, and determine the dynamic verification length based on the anomaly density and the standard verification length; the anomaly density is taken as the ratio of the amount of abnormal data to the anomaly detection window length; the dynamic verification length is positively correlated with the anomaly density.

[0110] The abnormal timestamp is used as the midpoint of the verification period, and the verification period is determined based on the dynamic verification length.

[0111] In actual assessments, the dynamic verification length is adjusted using a time sensitivity coefficient when calculating the dynamic verification length. The expression is as follows:

[0112] ;

[0113] ;

[0114] in For dynamic verification length, For standard verification length, The time sensitivity coefficient, It is an abnormal density. This is an abnormal amount of data. This is the length of the anomaly detection window;

[0115] The purpose of the anomaly detection is to identify data with potential anomaly risks. During anomaly detection, the radius of the sphere is set to be smaller than the actual sphere. The anomaly detection result does not necessarily mean that the data of the operation node is abnormal. The verification time period is different for each operation node.

[0116] In this embodiment, the method for obtaining the data center degree vector includes:

[0117] Based on historical verification and evaluation data of similar products, cluster analysis was used to determine representative data centers for similar products. A weighted average method was then used to calculate the representative data centers for all similar products to obtain the data center data center data center data. ;

[0118] The expressions for distribution density centrality, statistical characteristic centrality, and volatility stability centrality, calculated based on data center and verification evaluation data, are as follows:

[0119] ;

[0120] ;

[0121] ;

[0122] ;

[0123] in Density centrality, Centrality is a statistical characteristic. For the wave stability centrality, Frequency weighting For range weights, To determine the frequency of occurrence of verification and evaluation data that are completely identical to the central data, This represents the total number of verification and evaluation data. For located The amount of data within the range For located Data volume outside the scope To aggregate weights, To verify the mean of the evaluation data, For discrete weights, To verify the standard deviation of the evaluation data, Weights for distribution shape To verify the skewness of the evaluation data, For data window weights, For fluctuation weights, The global standard deviation, For rolling window volatility, To verify the window length, For data points The value, Extreme value weights;

[0124] The centrality of the verification and evaluation data is obtained by weighting the distribution density centrality, statistical characteristic centrality and fluctuation stability centrality. The centrality vector of the verification and evaluation data of each key operation node in each product process chain is composed of the centrality of the data center.

[0125] In practical assessments, frequency weighting is used when calculating the centrality of the distribution density. / range weight We set the clustering weights to 0.4 / 0.6 when calculating the statistical characteristic centrality. / Discrete weights / Distribution Shape Weight The fluctuation weights are 0.5 / 0.3 / 0.2 when calculating the fluctuation stability centrality. / Extremum Weight It is 0.7 / 0.3;

[0126] The dimensions of the data center degree vector are consistent with all key operation nodes in the product process chain, and the elements of the data center degree vector are arranged in the order of the product process chain.

[0127] In this embodiment, the method for obtaining the confidence matrix includes:

[0128] The initial confidence level of the product process chain data at each operational node in each process chain during the collaborative verification period is used as the initial confidence level. The confidence level of the verification evaluation data for each operational node in each process chain is then calculated using a time-series dynamic confidence decay model. The expression is as follows:

[0129] ;

[0130] ;

[0131] in for Operation node verification and evaluation data in Confidence level at any given moment This is the initial confidence level. for Operation node verification and evaluation data in The temporal decay coefficient at time step, This is the initial time of the verification period. For the Sigmoid function, , As a weight for process stability, Within the verification period All data in the operation node is located in the corresponding The amount of data within the range For the corresponding Data volume outside the scope Within the verification period All data volume of the operation node. This represents the proportion of samples whose time-sensitivity distance to the corresponding data center exceeds a dynamic distance threshold. To verify the event impact value, including anomaly detection events and user feedback events, To verify the number of verification events within the specified time period. The moment when the verification event occurs. The event affects the half-life. Anchor indicator functions for blockchain events to ensure that events are non-repudiable. Based on the attenuation coefficient, The fluctuation sensitivity coefficient This represents time-series fluctuations, reflecting recent data volatility.

[0132] The expressions for the timeliness distance of the sample data to the corresponding data center and the dynamic distance threshold are as follows:

[0133] ;

[0134] ;

[0135] in For sample data To the corresponding data center Timeliness distance, The aging factor is... For sample data timestamp, For dynamic distance threshold, This is the initial distance threshold;

[0136] Within the verification period, the confidence scores of the verification evaluation data of the same operation node are arranged in chronological order to form a confidence vector;

[0137] The confidence vectors of each operation node are combined into a confidence matrix according to the process chain sequence;

[0138] In actual evaluation, the confidence level of the product process chain data input into the global verification model at each operation node is directly extracted as the initial confidence level. At this time, the verification evaluation data of the verification period is only a part of the product process chain data input into the global verification model. Therefore, it is necessary to dynamically adjust the initial confidence level through time decay, data distribution, anomaly detection and verification events.

[0139] Calculating the timing attenuation coefficient At that time, the fluctuation sensitivity coefficient Taking 0.5, recent time series fluctuations During periods of intense activity (such as when switching production shifts), the time-series decay coefficient... As the data increases, the confidence level of historical data decays more rapidly, forcing the system to rely more on recent verification data for decision-making, thus achieving timeliness adaptation.

[0140] Calculate the timeliness distance from sample data to the corresponding data center At that time, aging coefficient Take 0.01 / h, the proportion of time-sensitive samples. The larger the value (closer to 1), the more it indicates that a large amount of historical data continues to deviate from the center, the central benchmark is aging and ineffective, and a re-election mechanism needs to be triggered immediately to adjust the data center;

[0141] Verification events include anomaly detection events and user feedback events. The impact value of a verification event is determined based on a dynamic function of the event severity.

[0142] For the impact value of anomaly detection events, if the verification data If the abnormal score exceeds the abnormal score threshold (0.15), a negative shock is generated, expressed as follows:

[0143] ;

[0144] in This represents the impact value of an abnormal detection event. To verify the data The outlier scores are extracted using the aforementioned multidimensional sphere model. The threshold for abnormal scores;

[0145] For the impact value of user feedback events, sentiment analysis is used to map the feedback intensity (ranging from -1 to 1, where -1 represents strong negative and 1 represents strong positive) to impact, expressed as:

[0146] ;

[0147] in The impact value of user feedback events. For feedback strength, This is an indicator function that ensures feedback only takes effect when the cross-domain correlation deviation verification passes, thus preventing malicious users from manipulating reviews.

[0148] Event affects half-life The timeframes for routine verification events, major anomaly events, and systemic verification events are 24 / 168 / 720 hours, respectively.

[0149] After each verification event is generated, its event hash, timestamp, impact dimension, and impact value are encapsulated into a verification event package, which is then anchored through the aforementioned Light-Chain blockchain. Once anchored, the indicator function takes effect, ensuring that the event's modification of the confidence vector is non-repudiable and tamper-proof. If an event is not anchored (e.g., due to a network failure), its impact value is only temporarily stored in memory and participates in the confidence calculation only after the anchoring is confirmed, thus preventing replay attacks and event forgery.

[0150] In this embodiment, the method for calculating the correlation deviation vector includes:

[0151] Calculate the probability distribution of verification and evaluation data across different dimensions within the same operation node to obtain the information entropy of that operation node. Based on the node's information entropy, calculate the mutual information between two operation nodes and perform temporal enhancement to obtain temporal mutual information. Calculate the mean conditional information entropy of a single operation node with all other operation nodes and perform multi-timescale fusion to obtain the enhanced correlation degree of the corresponding operation node. The expression is:

[0152] ;

[0153] ;

[0154] in For operation nodes and Timing-enhanced mutual information , For the data set of the corresponding operation node, For operation nodes and The length of the time period, For operation nodes exist Time period and operation node exist Inter-time information For operation nodes exist Information entropy over a period of time For operation nodes Enhance the correlation, , For time fusion weights, This represents the number of operation nodes. For operation nodes and During the same period Standardized mutual information, For operation nodes and During the same period mutual information, The longest time span. For operation nodes Timing mutual information with subsequent operation nodes;

[0155] The correlation vector is obtained by combining the enhanced correlation of each operation node according to the process chain sequence.

[0156] Extract the process chain data of qualified products in the same category, calculate the standardized mutual information of each operation node in the same time period to obtain the correlation vector of products in the same category, average and weighted fuse all correlation vectors of products in the same category to obtain the correlation fusion vector of products in the same category, and calculate the deviation between the correlation vector of the product to be verified and the corresponding correlation fusion vector of products in the same category to obtain the correlation deviation vector.

[0157] In actual evaluation, the user feedback data obtained includes explicit feedback, implicit feedback, and social feedback. Explicit feedback includes user ratings, comment sentiment, complaint rate, and return rate. Implicit feedback includes user dwell time, clickstream data, number of shares, and repeat purchase rate. Social feedback includes social media mention frequency, sentiment analysis, and influence spread. The user feedback data is preprocessed as follows: time window alignment (unifying the time granularity of feedback data to ensure time sequence consistency), outlier filtering (removing extreme ratings and obviously malicious comments), and multilingual processing (unified semantic analysis for multilingual comments of international products).

[0158] The improved S-curve function is used to normalize the rating data, the sentiment dictionary and deep learning model are used to calculate the sentiment score, and behavioral data such as click rate and conversion rate are used to normalize the user feedback data to obtain the user feedback vector.

[0159] In this embodiment, the method for constructing a verification prediction model based on semi-supervised adversarial methods includes:

[0160] The first training set is formed by obtaining the verification labels corresponding to the historical verification evaluation indicators. The historical verification evaluation indicators without verification labels are actively learned and labeled to generate pseudo labels. The pseudo labels and the corresponding historical verification evaluation indicators are then combined to form the second training set. The verification labels include the verification pass status and the factors that affect the verification results.

[0161] The first and second training sets are randomly mixed to form a comprehensive set. The comprehensive set is then randomly divided into a training set and a test set in a 7:4 ratio to train and validate the verification prediction model based on semi-supervised adversarial methods. The verification prediction model based on semi-supervised adversarial methods includes an input layer, a feature processing layer, a shared encoding layer, a multi-task prediction layer, an adversarial discriminator, a feature sampler, and an output layer.

[0162] The feature processing layer uses a fully connected layer, BatchNorm, Dropout, and activation functions to perform feature normalization, missing value processing, and feature selection on the verification evaluation indicators to obtain standardized feature vectors.

[0163] The shared coding layer learns deep representations of the normalized feature vectors through a multilayer perceptron and residual connections to obtain coded features;

[0164] The multi-task prediction layer includes a verification predictor and a factor predictor; the verification predictor uses a BP neural network to perform regression prediction on the encoded features to obtain the verification pass rate and determine the verification pass result; the factor predictor performs multi-label classification on the encoded features and determines the factors that affect the verification result.

[0165] The adversarial discriminator is connected to the shared coding layer, inputs the encoded features into the lightweight discriminant network to distinguish the feature distribution of real labeled samples and pseudo-labeled samples and outputs the sample source probability;

[0166] The output of the feature sampler connected to the adversarial discriminator clearly marks the source of the samples, and positive and negative sample pairs are selected to train the multi-task prediction layer based on feature similarity and business rules;

[0167] The adversarial discriminator and the feature sampler are used only for training the multi-task prediction layer;

[0168] The output layer is connected to the multi-task prediction layer through a fully connected layer to output the verification result; the verification result includes the verification pass status and the factors that affect the verification result.

[0169] The semi-supervised adversarial verification prediction model is trained using a multi-task loss function, expressed as follows:

[0170] ;

[0171] ;

[0172] ;

[0173] ;

[0174] ;

[0175] ;

[0176] in For multi-task loss function, , , , , As task weight, To verify the predicted losses, To predict losses based on factors, To combat the losses, To combat the losses, For consistency loss, This represents the number of truly labeled samples. The number of pseudo-labeled samples. For the first Individual sample verification result label, For the first The probability that a sample passes verification. To verify the total number of influencing factor categories, For the first The verification results of the first sample are subject to the first Labels influenced by several factors For the first The verification results of the first sample are subject to the first The probability of being affected by each factor. To obtain data from the distribution of real labeled data Medium-sized samples The mathematical expectation, To distribute pseudo-labeled data Medium-sized samples The mathematical expectation, For the discriminator function, For feature extraction function, To counteract the weighting coefficient, For the set of positive sample pairs The number of corresponding sample pairs, For positive sample pairs negative sample set The number of corresponding sample pairs, For anchor point samples eigenvectors, Positive samples eigenvectors, negative samples eigenvectors, The cosine similarity function is used. For boundary parameters, For the original sample The verification prediction probability, These are the disturbance parameters;

[0177] The supervised adversarial verification prediction model employs a multi-stage training strategy for adversarial training, including supervised pre-training, semi-supervised adversarial training, and consistency refinement. The supervised pre-training optimizes the verification prediction loss using only real labeled data. Factors predicting loss The semi-supervised adversarial training uses real labeled data to optimize the multi-task loss function. Using pseudo-labeled data to optimize and mitigate loss and consistency loss The consistency refining focuses on optimizing consistency loss. ;

[0178] The product data to be verified is input into a semi-supervised adversarial graph verification network to obtain the verification results.

[0179] In actual evaluation, the specific method for obtaining the second training set is as follows: select the samples with the most uncertainty of the model for manual labeling, select samples according to the prediction entropy and embedding space density, generate pseudo labels for high confidence prediction entropy, filter unreliable pseudo labels according to time consistency, and form the second training set with the remaining pseudo labels and corresponding historical verification evaluation indicators.

[0180] The factor predictor performs multi-label classification on the encoded features, extracts the data identifiers corresponding to the labels, traces the data on the Light-Chain blockchain based on the data identifiers, extracts the verification event packets anchored on the trusted nodes of the Light-Chain blockchain (corresponding to the operation nodes of the process chain), and determines the factors affecting the verification results.

[0181] Verification result label The value is 1 if the verification passes and 0 if the verification fails. The verification results of the first sample are subject to the first When each factor has an impact Select 1 if the input is 1, otherwise select 0; feature extraction function Mapping input samples to feature representations, the discriminator function Given the same input features, the output probability and boundary parameters are derived from the real labeled data. Take 0.5;

[0182] The specific strategies for constructing positive and negative sample pairs include:

[0183] Feature similarity-based sampling: When selecting positive sample pairs, calculate the cosine similarity matrix between all samples, and select the k samples with the highest similarity for each sample as positive samples (the similarity should be higher than the similarity threshold); when selecting negative sample pairs, randomly select random negative samples from samples of different categories, and select samples with moderate similarity (similarity within the similarity range) but different verification results as hard negative samples.

[0184] Sampling based on business rules: When selecting positive sample pairs, different products from the same supplier, similar products within the same time period, and products with similar user feedback patterns are used as positive samples; when selecting negative sample pairs, products with opposite verification results, products with large geographical differences, and products with large price range differences are used as negative samples.

[0185] Secondly, a product data verification system based on big data analytics includes:

[0186] The primary verification module is used to configure a global verification model through federated learning, and to use the global verification model to collaboratively verify the product process chain data to obtain collaborative verification conclusions; it is also used to build a Light-Chain blockchain, perform fingerprint audits on the Light-Chain blockchain corresponding to each product process chain, and obtain fingerprint audit results.

[0187] Anomaly detection module: Used to detect anomalies in primary verification data, determine the verification time period based on the anomaly detection results, and extract all primary verification data within the verification time period as verification evaluation data;

[0188] Verification and evaluation index calculation module: used to calculate the centrality of each verification and evaluation data to obtain the centrality vector; used to calculate the confidence of each verification and evaluation data according to the time series dynamic confidence decay model to obtain the confidence matrix; used to extract user feedback data and normalize it to obtain the user feedback vector; used to calculate the correlation deviation vector.

[0189] Verification prediction module: This module is used to build a semi-supervised adversarial verification prediction model based on historical verification evaluation indicators and verification labels. The verification results are obtained by inputting the product data to be verified into the semi-supervised adversarial verification prediction model.

[0190] Management module: Used to store, manage, and view Light-Chain blockchain and verification results, and to process data based on the verification results.

[0191] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A product data verification method based on big data analysis, characterized in that, Includes the following steps: S1. Collect product process chain data for preliminary verification to obtain preliminary verification data; the product process chain data includes production chain data, logistics chain data, and sales chain data; the preliminary verification includes collaborative verification and fingerprint auditing; S2. Perform anomaly detection on the primary verification data, determine the verification time period based on the anomaly detection results, and extract all primary verification data within the verification time period as verification evaluation data; the anomaly detection results include anomaly score, anomaly data volume, and anomaly timestamp. S3. Extract the data centers of similar products, calculate the centrality of each verification and evaluation data to obtain the data center degree vector, and calculate the confidence of each verification and evaluation data according to the time-series dynamic confidence decay model to obtain the confidence matrix. S4. Extract user feedback data and normalize it to obtain a user feedback vector. Calculate the correlation between each verification and evaluation data and the remaining data to obtain a correlation vector. Calculate the correlation deviation vector based on the correlation vector of similar products. S5. Construct a semi-supervised adversarial verification prediction model based on historical verification evaluation indicators and verification labels. Input the product data to be verified into the semi-supervised adversarial verification prediction model to obtain the verification results. The verification evaluation indicators include anomaly score, centrality vector, confidence matrix, user feedback vector, and correlation deviation vector. The method for determining the verification time period includes: A multidimensional sphere model is constructed based on the data units of each key operation node in each product process chain; the types of multidimensional sphere models include production node spheres, logistics node spheres, and sales node spheres; the center of the multidimensional sphere model is taken as the mean vector of each dimension feature of the corresponding key operation node in the process chain, and the radius of the sphere is determined based on Mahalanobis distance; The initial verification data is input into the multidimensional sphere model according to the process chain and corresponding operation nodes for anomaly detection. The Euclidean distance from the abnormal data in the continuous time series to the center of the sphere is calculated. The difference between the Euclidean distance and the radius of the sphere is calculated to obtain the anomaly distance. The maximum value of the ratio of the anomaly distance to the radius of the sphere in the continuous time series is taken as the anomaly score, and the timestamp of the corresponding time series is taken as the anomaly timestamp. Calculate the anomaly density within the data anomaly detection time period, and determine the dynamic verification length based on the anomaly density and the standard verification length; the anomaly density is taken as the ratio of the amount of abnormal data to the anomaly detection window length; the dynamic verification length is positively correlated with the anomaly density. The abnormal timestamp is used as the midpoint of the verification period, and the verification period is determined based on the dynamic verification length.

2. The product data verification method based on big data analysis according to claim 1, characterized in that, The method for obtaining primary verification data includes: Collect product process chain data, format all data in a unified manner, align with timestamps and then perform entity parsing. Store the preprocessed product process chain data in a buffer for verification. The client for federated learning is determined, the coordination server publishes a unified verification model architecture, and distributes the initialized model parameters to each client. Each client trains a local verification model using local process chain data and inputs it to the coordination server. The coordination server generates global model parameters based on the local model parameters and returns them to each client to configure the global verification model. A global verification model is used to collaboratively verify the product process chain data to obtain preliminary verification results, and the preliminary verification results of each client are integrated to obtain a collaborative verification conclusion. The encrypted hash value of each product process chain data unit is calculated as a data fingerprint. Each data provider uses its private key to sign the data fingerprint to generate a digital signature. The data fingerprint and digital signature are packaged into a transaction. Multiple transactions are aggregated into a Merkle root through a Merkle tree to construct the Light-Chain blockchain. The Light-Chain blockchain includes a production blockchain, a logistics blockchain, and a sales blockchain. The Light-Chain blockchain is composed of multiple trusted nodes. The block header contains the Merkle root, timestamp, previous block hash, and consensus information. The block body stores the transaction list. Fingerprint auditing is performed on the Light-Chain blockchain corresponding to each product process chain. The specific steps include: querying the Merkel proof of the relevant data fingerprint from the Light-Chain blockchain, verifying the fingerprint consistency on the Light-Chain lightweight chain corresponding to different process chains, and combining the Merkel proof and fingerprint consistency results into fingerprint audit results. Product process chain data that has been verified through collaborative validation and fingerprint auditing will be used as primary verification data.

3. The product data verification method based on big data analysis according to claim 1, characterized in that, The method for obtaining the data center degree vector includes: Based on historical verification and evaluation data of similar products, cluster analysis was used to determine representative data centers for similar products. A weighted average method was then used to calculate the representative data centers for all similar products to obtain the data center data center data center data. ; The expressions for the distribution density centrality, statistical characteristic centrality, and volatility stability centrality, calculated based on the data center and verification evaluation data, are as follows: ； ； ； ； in Density centrality, Centrality is a statistical characteristic. For the wave stability centrality, Frequency weighting For range weights, To determine the frequency of occurrence of verification and evaluation data that are completely identical to the central data, This represents the total number of verification and evaluation data. For located The amount of data within the range For located Data volume outside the scope To aggregate weights, To verify the mean of the evaluation data, For discrete weights, To verify the standard deviation of the evaluation data, Weights for distribution shape To verify the skewness of the evaluation data, For data window weights, For fluctuation weights, The global standard deviation, For rolling window volatility, To verify the window length, For data points The value, Extreme value weights; The centrality of the verification and evaluation data is obtained by weighting the distribution density centrality, statistical characteristic centrality, and fluctuation stability centrality. The centrality vector of the verification and evaluation data of each key operation node in each product process chain is composed of the centrality of the data.

4. The product data verification method based on big data analysis according to claim 1, characterized in that, The method for obtaining the confidence matrix includes: The initial confidence level of the product process chain data at each operational node in each process chain during the collaborative verification period is used as the initial confidence level. The confidence level of the verification evaluation data for each operational node in each process chain is then calculated using a time-series dynamic confidence decay model. The expression is as follows: ； ； in for Operation node verification and evaluation data in Confidence level at any given moment This is the initial confidence level. for Operation node verification and evaluation data in The temporal decay coefficient at time step, This is the initial time of the verification period. For the Sigmoid function, , As a weight for process stability, Within the verification period All data in the operation node is located in the corresponding The amount of data within the range For the corresponding Data volume outside the scope Within the verification period All data volume of the operation node. This represents the proportion of samples whose time-sensitivity distance to the corresponding data center exceeds a dynamic distance threshold. To verify the event impact value, including anomaly detection events and user feedback events, To verify the number of verification events within the specified time period. The moment when the verification event occurs. The event affects the half-life. Anchor indicator functions for blockchain events to ensure that events are non-repudiable. Based on the attenuation coefficient, The fluctuation sensitivity coefficient This represents time-series fluctuations, reflecting recent data volatility. The expressions for the timeliness distance of the sample data to the corresponding data center and the dynamic distance threshold are as follows: ； ； in For sample data To the corresponding data center Timeliness distance, The aging factor is... For sample data timestamp, For dynamic distance threshold, This is the initial distance threshold; Within the verification period, the confidence scores of the verification evaluation data of the same operation node are arranged in chronological order to form a confidence vector; The confidence vectors of each operation node are combined into a confidence matrix according to the process chain sequence.

5. The product data verification method based on big data analysis according to claim 1, characterized in that, The method for calculating the correlation deviation vector includes: Calculate the probability distribution of verification and evaluation data across different dimensions within the same operation node to obtain the information entropy of that operation node. Based on the node's information entropy, calculate the mutual information between two operation nodes and perform temporal enhancement to obtain temporal mutual information. Calculate the mean conditional information entropy of a single operation node with all other operation nodes and perform multi-timescale fusion to obtain the enhanced correlation degree of the corresponding operation node. The expression is: ； ； in For operation nodes and Timing-enhanced mutual information , For the data set of the corresponding operation node, For operation nodes and The length of the time period, For operation nodes exist Time period and operation node exist Inter-time information For operation nodes exist Information entropy over a period of time For operation nodes Enhance the correlation, , For time fusion weights, This represents the number of operation nodes. For operation nodes and During the same period Standardized mutual information, For operation nodes and During the same period mutual information, The longest time span. For operation nodes Timing mutual information with subsequent operation nodes; The correlation vector is obtained by combining the enhanced correlation of each operation node according to the process chain sequence. Extract the process chain data of qualified products in the same category, calculate the standardized mutual information of each operation node in the same time period to obtain the correlation vector of products in the same category, and average and weight all correlation vectors of products in the same category to obtain the correlation fusion vector of products in the same category. Calculate the deviation between the correlation vector of the product to be verified and the corresponding correlation fusion vector of products in the same category to obtain the correlation deviation vector.

6. The product data verification method based on big data analysis according to claim 1, characterized in that, The method for constructing a verification prediction model based on semi-supervised adversarial methods includes: The first training set is formed by obtaining the verification labels corresponding to the historical verification evaluation indicators. The second training set is formed by actively learning and labeling the historical verification evaluation indicators without verification labels and generating pseudo labels. The pseudo labels and the corresponding historical verification evaluation indicators are then combined to form the second training set. The verification labels include the verification pass status and the factors that affect the verification results. The first and second training sets are randomly mixed to form a comprehensive set. The comprehensive set is then randomly divided into a training set and a test set in a 7:4 ratio to train and validate the verification prediction model based on semi-supervised adversarial methods. The verification prediction model based on semi-supervised adversarial methods includes an input layer, a feature processing layer, a shared encoding layer, a multi-task prediction layer, an adversarial discriminator, a feature sampler, and an output layer. The feature processing layer uses a fully connected layer, BatchNorm, Dropout, and activation functions to perform feature normalization, missing value processing, and feature selection on the verification evaluation indicators to obtain standardized feature vectors. The shared coding layer learns deep representations of the normalized feature vectors through a multilayer perceptron and residual connections to obtain coded features; The multi-task prediction layer includes a verification predictor and a factor predictor; the verification predictor uses a BP neural network to perform regression prediction on the encoded features to obtain the verification pass rate and determine the verification pass result; the factor predictor performs multi-label classification on the encoded features and determines the factors that affect the verification result. The adversarial discriminator is connected to the shared coding layer, inputs the encoded features into the lightweight discriminant network to distinguish the feature distribution of real labeled samples and pseudo-labeled samples and outputs the sample source probability; The output of the feature sampler connected to the adversarial discriminator clearly marks the source of the samples, and positive and negative sample pairs are selected to train the multi-task prediction layer based on feature similarity and business rules; The adversarial discriminator and the feature sampler are used only for training the multi-task prediction layer; The output layer is connected to the multi-task prediction layer through a fully connected layer to output the verification result; the verification result includes the verification pass status and the factors that affect the verification result. The semi-supervised adversarial verification prediction model is trained using a multi-task loss function, expressed as follows: ； ； ； ； ； ； in For multi-task loss function, , , , , As task weight, To verify the predicted losses, To predict losses based on factors, To combat the losses, To combat the losses, For consistency loss, This represents the number of truly labeled samples. The number of pseudo-labeled samples. For the first Individual sample verification result label, For the first The probability that a sample passes verification. To verify the total number of influencing factor categories, For the first The verification results of the first sample are subject to the first Labels influenced by several factors For the first The verification results of the first sample are subject to the first The probability of being affected by each factor. To obtain data from the distribution of real labeled data Medium-sized samples The mathematical expectation, To distribute pseudo-labeled data Medium-sized samples The mathematical expectation, For the discriminator function, For feature extraction function, To counteract the weighting coefficient, For the set of positive sample pairs The number of corresponding sample pairs, For positive sample pairs negative sample set The number of corresponding sample pairs, For anchor point samples eigenvectors, Positive sample eigenvectors, negative samples eigenvectors, The cosine similarity function is used. For boundary parameters, For the original sample The verification prediction probability, These are the disturbance parameters; The semi-supervised adversarial verification prediction model employs a multi-stage training strategy for adversarial training, including supervised pre-training, semi-supervised adversarial training, and consistency refinement. The supervised pre-training optimizes the verification prediction loss using only real labeled data. Factors predicting loss The semi-supervised adversarial training uses real labeled data to optimize the multi-task loss function. Using pseudo-labeled data to optimize and combat loss and consistency loss The consistency refining focuses on optimizing consistency loss. ; The verification results are obtained by inputting the product data to be verified into a semi-supervised adversarial verification prediction model.

7. A product data verification system based on big data analysis, used to perform the method according to any one of claims 1-6, characterized in that, include: The primary verification module is used to configure a global verification model through federated learning, and to use the global verification model to collaboratively verify the product process chain data to obtain collaborative verification conclusions; it is also used to build a Light-Chain blockchain, perform fingerprint audits on the Light-Chain blockchain corresponding to each product process chain, and obtain fingerprint audit results. Anomaly detection module: Used to detect anomalies in primary verification data, determine the verification time period based on the anomaly detection results, and extract all primary verification data within the verification time period as verification evaluation data; Verification and evaluation index calculation module: used to calculate the centrality of each verification and evaluation data to obtain the centrality vector; used to calculate the confidence of each verification and evaluation data according to the time series dynamic confidence decay model to obtain the confidence matrix; used to extract user feedback data and normalize it to obtain the user feedback vector; used to calculate the correlation deviation vector. Verification prediction module: This module is used to build a semi-supervised adversarial verification prediction model based on historical verification evaluation indicators and verification labels. The verification results are obtained by inputting the product data to be verified into the semi-supervised adversarial verification prediction model. Management module: Used to store, manage, and view Light-Chain blockchain and verification results, and to process data based on the verification results.

Images