Intelligent bid evaluation expert extraction and management system

Abstract

The invention discloses an intelligent bid evaluation expert extraction and management system, and relates to the technical field of expert extraction systems. Comprising a data cleaning and compression module, a database logic and table relation design module, an expert information collection standard and automation module, an information real-time updating and principal management module, an extraction algorithm design module, a safety and authority management module and a real-time monitoring and auditing function module. Through a self-adaptive sparse auto-encoder and a dynamic clustering technology in the data cleaning and compression module, abnormal values in a bid evaluation expert database can be detected and cleaned, data compression is performed by adopting a self-adaptive dictionary coding algorithm, and the storage efficiency and query performance of the database are improved; based on a multi-modal information fusion and reinforcement learning method, accurate collection, dynamic updating and intelligent management of expert information are realized. The randomness and fairness of the bid evaluation expert extraction process are ensured by using a mode of combining deep reinforcement learning and a quantum random number generator in an extraction algorithm design module.

Description

Technical Field

[0001] The present invention relates to the technical field of bid evaluation expert extraction systems, and in particular to an intelligent bid evaluation expert extraction and management system. Background Art

[0002] In the current bid evaluation expert database management and expert extraction process, there are some urgent problems to be solved. For example, there are often a large number of null values ​​and redundant values ​​in the expert database, which leads to data redundancy and waste of storage space, affecting query efficiency and server performance. Traditional expert extraction methods are usually based on simple random number generation algorithms. Although basic randomness can be achieved, repeated series are easily generated, making it difficult to ensure true fairness. In addition, the expert extraction process may be interfered with by human factors to a certain extent, and complete independence and objectivity cannot be ensured. The maintenance and update of expert database information is also relatively lagging, and there is a high workload for manually updating information, and the updated information may not be accurate, resulting in poor consistency and real-time performance of expert database information. These problems make the current expert database system have many deficiencies in data management, extraction fairness and security. Summary of the invention

[0003] In view of the above problems, the present invention proposes an intelligent bid evaluation expert extraction and management system, which adopts advanced AI algorithms and technical architecture to achieve deep optimization of the database, ensure the accuracy and completeness of the data, and improve the query speed and recognition performance; in terms of expert extraction, it uses a variety of intelligent algorithms and strategies to eliminate human intervention factors and ensure the fairness and randomness of the extraction results; and can timely maintain and update the expert database information; thereby providing strong support for the smooth development of bid evaluation work, promoting technological progress and standardized development in the bid evaluation field, and improving the service quality and credibility of the entire industry.

[0004] An intelligent bidding evaluation expert extraction and management system, specifically comprising:

[0005] The data cleaning and compression module pre-processes the original bid evaluation expert database data and transmits the pre-processed compressed data to the database logic and table relationship design module, the expert information collection standard and automation module, and the information real-time update and person in charge management module; the pre-processing includes outlier detection and data compression;

[0006] The database logic and table relationship design module classifies and stores the received compressed data in different tables, and builds the relationship between different tables in the database based on semantic association and knowledge graph;

[0007] Expert information collection standard and automation module receives compressed data, fuses multimodal information in the compressed data to obtain expert information, and outputs expert information to the information real-time update and responsible person management module and the extraction algorithm design module;

[0008] The real-time information update and person-in-charge management module monitors the work dynamics and professional skill changes of bid evaluation experts in real time, and automatically updates the database based on the received compressed data and expert information;

[0009] The extraction algorithm design module receives expert information and learns the matching rules between project requirements and expert skills in expert information through a deep reinforcement learning algorithm. It uses a dynamic knowledge graph to associate expert skills and project requirements. It also combines the expert recommendation method based on a collaborative filtering algorithm to extract experts.

[0010] Preferably, the outlier detection is specifically as follows:

[0011] An autoencoder is used to capture the high-dimensional features in the original bid evaluation expert database data; the loss function of the autoencoder includes a sparse regularization term, and the formula of the sparse regularization term is expressed as:

[0012]

[0013] Where z represents the potential representation vector output by the encoder part of the autoencoder, that is, a representation vector in the low-dimensional space to which the input data is mapped after passing through the encoder; λ represents the sparsity regularization coefficient; m represents the dimension of the potential representation vector z, that is, the dimension of the low-dimensional space after the autoencoder compresses the input data; z i represents the absolute value of the representation of the i-th data point in the latent space; || represents the absolute value;

[0014] Perform dynamic clustering operations, dynamically adjust the number of clusters, and adaptively adjust the number and boundaries of clusters according to the structural changes of the data. The dynamic adjustment formula of the number of clusters is expressed as:

[0015]

[0016] Among them, α represents the control coefficient, represents density entropy; represents the change in density entropy; k old represents the number of clusters before adjustment; k new represents the adjusted number of clusters;

[0017] In each iteration of the dynamic clustering operation, the cluster center is dynamically updated according to the data distribution, and the formula is expressed as:

[0018]

[0019] Among them, wij Represents data point z i Similarity weight with cluster center; represents the updated cluster center; n represents the total number of data points; j represents the index of the current cluster, j = 1, 2, ..., k, k represents the number of clusters;

[0020] Calculate data point z i The minimum similarity score is expressed as: Among them, min represents the minimum value;

[0021] The minimum similarity score of all data points is calculated, and data points with a minimum similarity score lower than a preset similarity score threshold are considered outliers.

[0022] Preferably, the data compression is specifically as follows:

[0023] The data is compressed using an adaptive feature dictionary coding compression algorithm. The dictionary structure and coding strategy are dynamically adjusted by real-time analysis of the data's features and frequent patterns. The dictionary update rule is expressed as:

[0024]

[0025] Among them, η represents the learning rate; represents the gradient of the dictionary; C new represents the sparse coding coefficient of the new data; γ represents the smoothness control parameter of the dictionary update; X new Indicates a new batch of input data; D new represents the updated dictionary matrix; D represents the current dictionary matrix; D old Represents the old dictionary matrix; represents the square of the Frobenius norm;

[0026] In addition, a dictionary pruning strategy is adopted to reduce the size of the dictionary. By analyzing the contribution of each basis vector in the dictionary, redundant or unimportant basis vectors are removed, thereby optimizing the dictionary structure.

[0027] Preferably, the received compressed data is classified and stored in different tables, and the association relationship between different tables in the database is constructed based on semantic association and knowledge graph, as follows:

[0028] Design the table structure and classify and store the bid evaluation expert information according to its characteristics. Each classification is stored in a different table in the database. The classification includes basic information of the expert, professional skills and bid evaluation history.

[0029] Establish primary key-foreign key constraints for different tables in the database;

[0030] Establish cross-reference or index relationships between experts with high semantic association in the database;

[0031] Build a knowledge graph and use the adjacency matrix to represent the associations between tables in the knowledge graph;

[0032] The knowledge graph is represented as a directed graph G = (V, E); wherein the node set V includes experts, fields, skills and evaluation items; the edge set E represents the relationship between the nodes in the node set; when there is new data, a new node is added to the knowledge graph. If the similarity between the semantic representation of the new data and the nodes in the existing knowledge graph is higher than a preset dynamic threshold, it indicates that the new data has a semantic association with the existing graph nodes, and the edge set is updated; the method for calculating the similarity between the semantic representation of the new data and the nodes in the existing knowledge graph is the same as the method for calculating the semantic association.

[0033] Preferably, the semantic relevance is specifically as follows:

[0034] The semantic similarity between two bid evaluation experts is measured by the cosine similarity of the semantic vectors of the two bid evaluation experts; the semantic vector is a vector representation of the expert attributes mapped in the semantic space; the expert attributes include fields, skills and experience; when the cosine similarity is higher than the set similarity threshold, it is considered that the two experts are semantically associated.

[0035] Preferably, the multimodal information in the fused compressed data is specifically as follows:

[0036] Extract text features H from text modality expert information in compressed data using a pre-trained language model T , using convolutional neural networks to extract image features H from image modality expert information in compressed data I ;

[0037] The text features and image features are deeply integrated through a two-layer interaction mechanism:

[0038] At the low-level feature interaction layer, a cross-modal attention mechanism is introduced. The text features are used as the query parameters of the cross-modal attention mechanism. The image features are used as the key parameters and value parameters of the cross-modal attention mechanism for weighted calculation. The global attention information A of the text modality is obtained. TI ;

[0039] The image features are used as the query parameters of the cross-modal attention mechanism, and the text features are used as the key parameters and value parameters of the cross-modal attention mechanism for weighted calculation to obtain the global attention information A of the image modality. IT , which is expressed as follows

[0040]

[0041] Among them, d T Indicates H T The characteristic dimension of Indicates H T The transpose of ; softmax represents the softmax function;

[0042] Intermediate feature fusion layer, the intermediate fusion feature is expressed as:

[0043] F mid =α T H T +α I H I

[0044] Among them, α T =σ(W T H T ) represents the weight of the text modality, α I =σ(W I H I ) represents the weight of the image modality, W T and W I represents the learnable weight matrix, σ represents the Sigmoid activation function;

[0045] The modality reconstruction loss is introduced in the fusion process, and the reconstruction loss function is expressed as:

[0046]

[0047] in, g T represents a network for reconstructing text modalities; g I Represents a network for reconstructing image modality; ∥∥ 2 represents the square of the Frobenius norm;

[0048] The final fusion feature representation of text modality and image modality is obtained, and the formula is expressed as:

[0049] H final =F mid +A TI +A IT

[0050] Preferably, the automatically updated database is based on dynamic weight blockchain and smart contract technology, as follows:

[0051] Construct a dynamic weight mechanism for expert information so that blockchain nodes prioritize high-weight expert information update requests; the weight w of the expert’s j-th category of expert information at time t j (t) is expressed as:

[0052]

[0053] Among them, α j represents the initial weight; β j Controls how fast the weights change over time; Δt j Indicates the time interval since the last update of the same type of information;

[0054] An adaptive block generation strategy is introduced to adjust the time interval and size of block generation according to the current network load and the weight of expert information; the time interval ΔT of blockchain block generation is expressed as:

[0055]

[0056] Among them, T0 represents the basic block generation time interval of the system; γ1 represents the adjustment coefficient; N represents the total number of experts, and M represents the total number of information types;

[0057] Smart contracts are used to automatically control the update process of expert information. The triggering rules of smart contracts are based on the weight value priority of expert information. Each type of information of each expert corresponds to a smart contract. When the weight of a piece of information exceeds the preset weight threshold, the smart contract is automatically triggered to update the information and synchronize it to the blockchain. The smart contract adopts a batch processing mechanism to allow multiple low-weight information to be updated together at the same time.

[0058] Preferably, expert extraction is performed based on deep reinforcement learning and dynamic knowledge graph, as follows:

[0059] Construct a dynamic knowledge graph and define the edge weight w between the expert node and the project node to be evaluated in the dynamic knowledge graph ij , which is used to measure the degree of adaptation between expert nodes and project nodes to be evaluated; the edge weight calculation formula is expressed as:

[0060]

[0061] Among them, λ k represents the weight coefficient; It represents the similarity between the expert node and the project node to be evaluated in the kth dimension, which is calculated using cosine similarity; K represents the number of characteristic attributes of each expert;

[0062] Based on the constructed dynamic knowledge graph, deep reinforcement learning is used to model expert features and dynamically adjust the extraction strategy. A reward function is designed to evaluate the quality of expert selection. The reward function is expressed as:

[0063]

[0064] Among them, v irepresents the feature vector of the expert node; f(v i ,p j ) represents the expert node's evaluation of the bid evaluation project node p j The degree of matching; t represents the action space, that is, the set of experts that best suits the current project is selected at time t; C(a t ) represents the cost of selecting the expert set; δ represents the penalty coefficient; Q represents the number of evaluation items; s t It represents the state representation, i.e., the characteristic matrix of experts and evaluation items;

[0065] The deep reinforcement learning is optimized using a proximal strategy optimization method, and dynamically adjusts and selects expert strategies, so that the system can adaptively select a suitable expert combination under different business needs, thereby maximizing the compatibility between experts and the projects to be evaluated; in the process of dynamically adjusting and selecting expert strategies, a quantum random number generator is used to generate random numbers.

[0066] Preferably, it also includes a security and authority management module, which uses blockchain technology to homomorphically encrypt and distribute the expert information in the expert information collection standard and automation module and the evaluation expert extraction process in the extraction algorithm design module; and allocates operation permissions and data access permissions based on user roles and attributes.

[0067] Preferably, it also includes a real-time monitoring and auditing function module, which monitors the update of expert information and the process of extracting bid evaluation experts in real time, updates the data changes in the process of extracting bid evaluation experts through a deep incremental learning model, and performs intelligent auditing of expert data; during the real-time monitoring and auditing process, the discovered potential security vulnerabilities are transmitted to the security and authority management module through a feedback mechanism, and the discovered information updates and management problems are transmitted to the information real-time update and responsible person management module through a feedback mechanism for self-correction.

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] (1) The data cleaning and compression module of the present invention adopts an outlier detection and adaptive compression method based on autoencoders and dynamic clustering to adaptively capture complex features and potential laws in the data, thereby efficiently and accurately marking outliers that do not conform to normal data distribution; at the same time, the module uses a compression algorithm based on intelligent adaptive dictionary coding to compress and store data, and dynamically adjusts the dictionary structure and encoding strategy by analyzing the characteristics and frequent patterns of the data in real time to minimize data storage space and make the storage structure of the database more compact and optimized. Compared with traditional methods, it can better adapt to the diversity and dynamic variability of the bid evaluation expert database data and improve the efficiency and quality of data processing.

[0070] (2) The database logic and table relationship design module of the present invention adopts a database logic and table relationship design method based on the fusion of semantic association and knowledge graph. It can classify and store relevant information in different tables according to the characteristics of the bid evaluation expert information, and establish clear table-to-table relationships through primary key and foreign key associations. At the same time, it uses graph database technology to construct a complex relationship network between expert information to improve query efficiency.

[0071] (3) The expert information collection standard and automation module of the present invention utilizes multimodal information fusion technology to comprehensively and automatically collect expert information; the collected expert information serves as the data source of the database optimization subsystem, supplementing the comprehensiveness and timeliness of the expert information in the expert database. It provides basic data for the subsequent expert information real-time update module, and provides more accurate expert feature data for the expert extraction algorithm.

[0072] (4) The information real-time update and responsible person management module of the present invention automatically updates the expert information in the database, ensures the dynamic update of the expert information in the expert database, and ensures the immutability and security of the data through blockchain technology.

[0073] (5) The extraction algorithm design module of the present invention proposes an expert extraction algorithm based on deep reinforcement learning and dynamic knowledge graph, which is used to extract features of expert information and model and construct expert portraits; and an algorithm based on a quantum random number generator is used to replace the traditional random number generation function to ensure the fairness of extraction; in terms of recommendation extraction, it is combined with a recommendation method based on the analysis of historical evaluation data to improve the recommendation effect and the intelligence level of the system.

[0074] (6) The security and authority management module of the present invention introduces blockchain technology to encrypt and distribute the expert information and extraction process to ensure the security and integrity of the data. At the same time, it implements an authority management strategy that combines role-based access control and attribute-based encryption, and assigns different authorities according to user roles and attributes to perform operations and data access control.

[0075] (7) The real-time monitoring and auditing function module of the present invention monitors the expert extraction process in real time, establishes a monitoring indicator system to promptly detect abnormal behaviors and potential risks, and automatically audits the extraction records and bid evaluation process to ensure the openness and transparency of the bid evaluation process. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] The present invention is further described in detail below in conjunction with the accompanying drawings;

[0077] Figure 1 It is a structural block diagram of the intelligent bidding evaluation expert extraction and management system according to an embodiment of the present invention;

[0078] Figure 2The present invention is a flowchart of an intelligent bidding expert extraction and management method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0079] The present invention is further described below through specific implementation modes.

[0080] like Figure 1 As shown, the present invention discloses an intelligent bid evaluation expert extraction and management system, including a database optimization subsystem 100, a bid evaluation expert database new model construction subsystem 200, and a bid evaluation expert intelligent extraction subsystem 300; each subsystem contains multiple modules, and each module cooperates with each other to form a complete process chain. The details are as follows:

[0081] (I) Database Optimization Subsystem 100

[0082] (1) Data cleaning and compression module 101:

[0083] In this module, an outlier detection and adaptive compression method based on autoencoder and dynamic clustering is proposed.

[0084] First, an outlier detection method is designed. The technology based on autoencoder and dynamic clustering is used to adaptively capture the complex features and potential rules in the bid evaluation expert database data. Specifically, in order to solve the problem that traditional autoencoders may capture redundant information when representing features, thereby affecting the sensitivity of outlier detection, an adaptive sparsity constraint is proposed. The constructed adaptive sparse autoencoder can select only a small number of important features in the low-dimensional latent space and exclude noise and redundant data. Adaptive sparsity is achieved by introducing a sparse regularization term to the latent variable z, and its sparsity loss function is It can be expressed as:

[0085]

[0086] Where z is the potential representation of the output of the encoder part of the autoencoder, which is a representation vector in the low-dimensional space to which the input data is mapped after passing through the encoder. λ is the sparsity regularization coefficient, which determines the strength of the sparsity constraint; m is the dimension of the potential representation vector z, that is, the dimension of the low-dimensional space after the autoencoder compresses the input data.

[0087] After automatically learning the high-dimensional features of the data through the autoencoder, a dynamic clustering framework is constructed, which can adaptively adjust the number and boundaries of clusters according to the structural changes of the data. The dynamic adjustment formula of the number of clusters k is:

[0088]

[0089] Among them, α is the control coefficient, Hρ and ΔHρ represent the density entropy and its change, kold and k new Represents the number of clusters before and after adjustment. Through this mechanism, the number of clusters can change adaptively to better fit the structure of the data.

[0090] Once the number of clusters k is determined, the cluster centers will be dynamically updated according to the data distribution. In each iteration, the cluster centers can be dynamically adjusted through the weighted center point update mechanism. The update formula is:

[0091]

[0092] Among them, w ij Represents data point z i and cluster center μ j The similarity weight of .

[0093] Outliers are usually those points that are far away from all cluster centers. In the latent space, the minimum similarity score can be used to determine whether a point is an outlier. For a data point z i , and define its minimum similarity score as:

[0094]

[0095] If S min (z i ) is lower than the similarity score threshold τ (set according to the statistical distribution of the similarity score), then the point is considered an outlier. That is, the condition for outlier detection can be expressed as:

[0096] IfS min (z i )<τ,z i is an outlier.

[0097] In this way, outliers that do not conform to normal data distribution can be efficiently and accurately marked, including null values ​​and redundant values. Compared with traditional outlier detection methods, this method better adapts to the diversity and dynamic changes of bid evaluation expert database data and improves the quality of data processing.

[0098] Design a data compression algorithm, and compress and store data based on an intelligent adaptive dictionary coding compression algorithm. This algorithm analyzes the characteristics and frequent patterns of the data in real time, and dynamically adjusts the dictionary structure and coding strategy, thereby minimizing the data storage space and making the database storage structure more compact and optimized. Specifically, design an adaptive feature dictionary coding method, combining the high-dimensional feature representation of the data with the adaptive capture of frequent patterns, and achieve the optimal data compression effect by dynamically adjusting the dictionary structure and coding strategy in real time. Assume that the new input data batch is (where n newis the number of new data points, d is the feature dimension), the dictionary D needs to be updated, and the update rules are as follows:

[0099]

[0100] Where η is the learning rate, represents the gradient of the dictionary, C new is the sparse coding coefficient of the new data, γ is the smoothness control parameter of the dictionary update, which is used to prevent the dictionary from fluctuating violently during the update process. D is the current dictionary matrix, which contains the dictionary elements used in the encoding process and is used to represent the characteristics of the input data. This is the dictionary that is dynamically adjusted and optimized in the algorithm and is continuously updated as new data is input. old is the old dictionary matrix, which is the previous state of the dictionary. It is used to provide a reference when updating the dictionary so that the updated dictionary will not deviate too much from the original structure to ensure the continuity and stability of the dictionary. new is the updated dictionary matrix, that is, after the new data batch X new The input and the dictionary matrix obtained after adaptive adjustment. Represents the square of the Frobenius norm (that is, the square root of the sum of the squares of the matrix elements).

[0101] In the dictionary optimization process, a dictionary pruning strategy is adopted to reduce the size of the dictionary while retaining the most informative basis vectors. By analyzing the contribution of each basis vector in the dictionary, redundant or unimportant basis vectors are removed to optimize the dictionary structure. The compressed data size S comp Including the storage cost of the dictionary S D and the encoded data storage overhead Specifically:

[0102]

[0103] The designed data compression algorithm not only improves the storage efficiency of the database, but also provides a solid foundation for the efficient operation of the database, reducing storage costs and data access time.

[0104] (2) Database logic and table relationship design module 102:

[0105] In this module, a database logic and table relationship design method based on the fusion of semantic association and knowledge graph is proposed.

[0106] Table structure design strategy: According to the characteristics of bid evaluation expert information, relevant information is classified and stored in different tables, including expert basic information table, professional skills table, bid evaluation history table, etc. This classified storage method makes data management clearer and more organized, and facilitates data query and maintenance.

[0107] Specifically, we first conduct semantic association analysis to mine the potential semantic relationship of expert information in the database. This requires establishing associations through semantic representation and similarity measurement of expert basic information, skills, bid evaluation history, and other data. Assume that the attribute vector of expert information is X i =(x i1 ,x i2 ,…,x im ), where x ij Represents the jth attribute of the i-th expert (such as field, skills, experience, etc.), and maps these attributes into vector representations v(x ij ). For example, the semantic similarity between evaluation experts i and j is measured by cosine similarity:

[0108]

[0109] Among them, v(x ik ) represents the semantic vector of expert i on the kth attribute.

[0110] When the similarity S(i,j) is higher than the set similarity threshold of 0.6, it is considered that experts i and j have a certain semantic association, so a semantic relationship can be established for these two experts in the database, which will affect the design of the following table relationship.

[0111] Table relationship establishment method: Establish clear table-to-table relationships through primary key and foreign key associations, and use graph database technology to build a complex relationship network between expert information. This design can improve query efficiency, quickly obtain various information related to bid evaluation experts, and meet the multi-dimensional query needs for expert information during the bid evaluation process.

[0112] To this end, a knowledge graph fusion method with dynamic semantic expansion is proposed. This method not only constructs a static knowledge graph based on the expert database, but also can adaptively adjust and optimize the structure of the knowledge graph according to the introduction of new data or the update of knowledge through the dynamic semantic expansion mechanism, thereby enhancing the flexibility and self-update ability of the database. Let the knowledge graph be a directed graph G = (V, E), where V is a node set representing experts, fields, skills, bid evaluation items, etc., and E is an edge set representing the relationship between nodes. When the semantic representation of new data x new With the node v in the existing graph i The similarity is higher than the dynamic threshold τ ext (set to 0.7 in this example), it is considered that the new data is consistent with the existing graph node v iThere is a semantic association, and new nodes are introduced or edges are updated in the graph to expand the knowledge graph. The mathematical description of the expansion is:

[0113] V new =V∪{x new},E new =E∪{(v new ,v i )∣S(v new ,x i )>τ ext}

[0114] Through the above-mentioned dynamic semantic expansion mechanism, the graph structure will dynamically expand according to the introduction of new data, continuously enhancing the expressive power of the knowledge graph.

[0115] After integrating semantic associations and knowledge graphs, the core of table relationship design is to establish logical associations between different tables through primary key-foreign key relationships and the extensibility of graph structures. Specifically, it includes:

[0116] ●Primary key-foreign key constraint: Let Table 1 be the expert basic information table, and the primary key be the expert ID exp ; Table 2 is the expert skills table, external

[0117] Key ID skill The primary key of association table 1 represents the skills of each expert.

[0118] ●Semantic association constraints: Create cross-references or indexes in the database for experts with high semantic association through semantic similarity

[0119] relations to optimize the efficiency of semantic association queries.

[0120] ● Adjacency relationship in graph database: At the knowledge graph level, use the adjacency matrix A = [a ij ] indicates the association between tables.

[0121] where a ij =1 indicates that there is a relationship between table i and table j, and 0 indicates no relationship. In this way, the graph structure can describe the complex relationship between multiple tables.

[0122] (II) New model construction subsystem of bid evaluation expert database 200

[0123] (1) Expert Information Collection Standards and Automation Module 201:

[0124] In this module, a method for collecting expert information based on multimodal information fusion and reinforcement learning is proposed.

[0125] The first step is to formulate information collection standards, covering experts' professional fields, work experience, academic achievements, professional ethics, etc. This ensures that the collected expert information is standardized and consistent, providing a reliable data basis for subsequent expert evaluation and extraction.

[0126] Furthermore, a multimodal information fusion method is established to process expert data extracted from multiple information sources. Multimodality includes texts filled in by experts, existing records in the database, and provided image data. Through the semantic fusion of multimodal data, it is possible to more comprehensively describe information such as experts' professional fields, work experience, and academic achievements. Specifically, a deep semantic interactive fusion method based on text-image dual modality is established. First, feature extraction is performed on the text modality and image modality:

[0127] Text dataX T Extract text features H through the pre-trained language model BERT T , image data X I Extract image features H through convolutional neural network ResNet I , which can be expressed as:

[0128] H T =f T (X T )

[0129] H I =f I (X I )

[0130] Among them, f T and f I They are feature extraction networks for text and image modalities, and are the feature representations of text and image respectively, d T and d I is their characteristic dimension.

[0131] In view of the large semantic differences between text and image modalities, in order to improve their complementarity and information integration, this embodiment deeply fuses text and image features through a two-layer interaction mechanism: the first layer is low-level feature interaction, and the second layer is mid-level feature fusion, combined with an innovative modality reconstruction mechanism to retain the independence of modal features.

[0132] At the low-level feature interaction layer, a cross-modal attention mechanism is introduced to achieve global information sharing between text and image modalities. Specifically, the features of the text modality will be used as queries, and the features of the image modality will be used as keys and values ​​for weighted calculation to obtain the global attention information A obtained by the text modality from the image modality. TI .

[0133] Similarly, the image modality is weighted by attention to the text modality to obtain the global information A of the image modality IT :

[0134]

[0135] This cross-modal attention mechanism ensures that each modality can obtain global contextual information from the other modality, which helps to understand the semantic relationship between images and text.

[0136] In the intermediate feature fusion layer, an adaptive fusion network is designed to dynamically assign weights to text and image modalities, ensuring that in different scenarios, the system can flexibly decide which modality’s information is more important for the current task. Define the adaptive weight α T and α I are the weights of text and image modalities, respectively, and are calculated using the following formula:

[0137] α T =σ(W T H T ), α I =σ(W I H I )

[0138] Among them, W T and W I is a learnable weight matrix, and σ is a Sigmoid activation function, which is used to limit the weight value between 0 and 1. The final intermediate fusion feature is expressed as:

[0139] F mid =α T H T +α I H I

[0140] The adaptive fusion mechanism allows the system to flexibly adjust the weights between text and image modalities according to different task requirements, ensuring that the fused features can not only retain the important information of each modality but also form a comprehensive semantic representation.

[0141] In addition, in order to ensure the independence of text and image modalities and promote their deep fusion, this embodiment introduces a modality reconstruction loss. In this mechanism, the fused feature F mid It is necessary to be able to reconstruct the original text and image modal features. Suppose the network g that reconstructs the text modality T (using bidirectional long short-term memory network BiLSTM) and the network g that reconstructs the image modality I (Using U-Net), the reconstructed features are represented as:

[0142] H T =g T (F mid ), H I =g I (F mid )

[0143] The reconstruction loss function is defined as:

[0144]

[0145] By minimizing the reconstruction loss, the system will try to retain the feature information of the original modality during the fusion process, ensuring that the integrity of the multimodal features can be maintained during task processing.

[0146] After two layers of interactive fusion, the final fusion of text and image modalities is expressed as:

[0147] H final =F mid +A TI +A IT

[0148] H final It integrates the low-level global interaction information of text and image modalities, mid-level adaptive fusion information, and information guaranteed by modal reconstruction, and is the final feature representation after deep fusion of the two modalities.

[0149] (2) Information real-time update and person in charge management module 202

[0150] A method for real-time updating and intelligent management of expert information based on blockchain and edge computing is proposed. With the help of Internet of Things technology and sensors, the work dynamics and professional skills changes of experts are monitored in real time, so as to automatically update the expert information in the database. The timeliness and accuracy of expert information are ensured, so that the bid evaluation expert database can always reflect the latest status of experts and provide the most practical expert resources for bid evaluation. Specifically, a method for updating expert information based on edge computing driven by dynamic weight blockchain and smart contracts is constructed. The dynamic weight blockchain mechanism enables different types of expert information to have different update priorities and processing methods, thereby improving the flexibility and efficiency of the system. The update frequency and importance of bid evaluation expert information are different. For example, the bid evaluation work dynamics of experts (such as project participation) need to be updated frequently, while the qualifications of experts (such as academic qualifications or certificates) are updated less frequently. In order to efficiently manage this information, this embodiment assigns a dynamic weight w to each information type. j (t), used to determine the update priority of each piece of information. Define expert E i The weight w of the j-th information at time t j (t), its update formula is:

[0151]

[0152] Among them, α j is the initial weight, reflecting the basic importance of this type of information; β j Controls how fast the weights change over time; Δt j Is the time interval since the last update of this type of information.

[0153] The dynamic adjustment mechanism designed above allows higher priority information (such as work dynamics) to be processed and updated faster at the edge computing level, while low priority information can be processed with a delay to optimize system resource allocation.

[0154] Based on the above dynamic weight mechanism, blockchain nodes will give priority to high-weight expert information update requests. In order to further improve the real-time performance of information updates, this embodiment introduces an adaptive block generation strategy. This strategy allows the time interval and size of block generation to be adjusted according to the current network load and the weight of expert information. The block generation time interval ΔT is defined as a dynamic value that depends on the overall load and weight sum of the system. The formula is as follows:

[0155]

[0156] Among them, T0 is the basic block generation time interval of the system; γ1 is the adjustment coefficient, which reflects the sensitivity of the system to load changes; w j (t) is the weight of the j-th type of information of expert i at time t; N is the total number of experts, and M is the total number of information types.

[0157] The constructed adaptive block generation strategy can ensure that the block generation frequency is higher during the period of intensive information update (such as when a certain expert updates a large amount of information), and slow down the block generation when the information update is less, saving system resources.

[0158] In the management module, smart contracts are used to automatically control the update process of expert information and trigger operations according to preset rules to ensure the transparency and automation of the update process. The innovation lies in combining smart contracts with the dynamic priority mechanism of expert information to ensure fairness and timeliness in the information update process.

[0159] Design a trigger mechanism for dynamic smart contracts. Assume that each type of information from each expert corresponds to a smart contract. When the weight of a piece of information w j (t) exceeds the threshold value θ j , the smart contract will be automatically triggered, the information will be updated and synchronized to the blockchain. The triggering condition formula is:

[0160]

[0161] When Trigger(E i ,j)=1, the smart contract starts the update process and submits the latest information to the blockchain.

[0162] In order to further improve the efficiency of information update, the smart contract adopts a batch processing mechanism, allowing multiple low-weight information to be updated at the same time, avoiding frequent block updates triggered by a single piece of information. The optimization formula for batch processing is as follows:

[0163]

[0164] in, It means that at time t, expert E i is the set of low-weight information batch updates, η is the total weight threshold of batch updates, ensuring that batch updates are triggered only when the amount of information is large enough, further saving blockchain resources and computing burden.

[0165] At the edge computing layer, the expert information obtained in real time is preprocessed and the weights are dynamically updated. The processing formula for edge computing is:

[0166]

[0167] Among them, f edge (·) represents the preprocessing function of edge computing, x i (t) is the information of expert i at time t, and the updated information is obtained after processing

[0168] When the weight of a piece of information reaches the threshold, the edge node submits the information to the blockchain node, triggering the smart contract and generating a new block. This enables the collaborative work of edge computing and blockchain, ensuring both real-time data updates and data security and immutability.

[0169] In addition, the overall situation and changing trends of the expert database are displayed through data analysis visualization technology, which makes it convenient for the person in charge to intuitively understand the operation status of the expert database and make timely decisions and management adjustments to ensure the accuracy and completeness of information, thereby improving management efficiency and scientific decision-making.

[0170] (III) Intelligent extraction subsystem for bid evaluation experts 300

[0171] (1) Extraction algorithm design module 301

[0172] Expert feature extraction and modeling, an expert extraction algorithm based on deep reinforcement learning and dynamic knowledge graph is proposed to extract features of expert information and build expert portraits. The deep reinforcement learning algorithm can dynamically adjust the extraction strategy under different business demand scenarios, improve the flexibility and adaptability of extraction, and meet the different requirements of various bid evaluation projects for experts. The core of expert extraction is to extract the characteristics of experts and build expert portraits according to project requirements. This part adopts a method that combines deep reinforcement learning with dynamic knowledge graphs to realize intelligent processing and modeling of expert information. First, the expert feature vector is constructed. Each expert E i is represented as a high-dimensional feature vector v i , which contains various attribute information of experts, such as professional field, work experience, historical bid evaluation data, etc. Assuming that each expert has K characteristic attributes, then expert E i The eigenvector of is represented as:

[0173]

[0174] Each feature It is extracted from the expert database and mainly comes from information of different dimensions such as the experts’ academic qualifications, professional titles, work experience, and professional fields.

[0175] Secondly, construct a dynamic knowledge graph and define the expert node E in the graph. i and the node P of the project to be evaluated j The edge weight w ij , which is used to measure the degree of adaptation between the two. The edge weight is calculated as follows:

[0176]

[0177] Among them, λ k is the weight coefficient, which reflects the influence of different features on adaptability; It is the similarity between the expert and the project in the kth dimension, calculated using cosine similarity.

[0178] Based on the above knowledge graph, deep reinforcement learning is used to model expert features and dynamically adjust the extraction strategy. The core of reinforcement learning is to train an agent to select actions based on the state of the environment in order to maximize the cumulative return. The agent here is responsible for selecting the most suitable expert in different bid evaluation tasks.

[0179] First is the state representation, the state s of the system t Including the demand information of the current project and the characteristic information of the experts. Assuming there are Q projects and N experts, then the state s t Expressed as the feature matrix of experts and items:

[0180] st = {v i ,p j ∣i=1,…,N;j=1,…,Q}

[0181] Then the action space is determined, action a t It is the expert set that selects the most suitable one for the current project at time t. Let project P j Select Expert E i , then the action is expressed as:

[0182] a t ={E i |i=1,…,N}

[0183] Finally, the reward function is designed. The reward function R(s t ,a t ) is used to evaluate the quality of expert selection. Assume that the experts selected each time make the evaluation function f(v i ,p j ) is maximized, the reward function can be defined as:

[0184]

[0185] Among them, f(v i ,p j ) is an expert E i For Project P j The matching degree, C(a t ) is the cost of selecting the expert set, and δ is the penalty coefficient, which is used to balance the trade-off between matching degree and cost.

[0186] Next, the proximal policy optimization (PPO) method is used to dynamically adjust the strategy for selecting experts, so that the system can adaptively select the appropriate combination of experts under different business needs and maximize the fit between experts and projects.

[0187] In order to improve the fairness and unpredictability of random extraction, a quantum random number generator is used. Quantum random number generation relies on the uncertainty of quantum bits when measuring, and its output random number r q The process can be expressed as, where Measure(ψ) represents the measurement of the quantum state ψ, which will cause the quantum state to collapse to a classical random result:

[0188] R q ~Measure(ψ)

[0189] At the same time, the process of expert extraction using random numbers generated by a quantum random number generator can be expressed as:

[0190] Ej =Extract(R q ,N)

[0191] Where N is the total number of bid evaluation experts, E j is the extracted expert set, and Extract(·) represents the extraction function.

[0192] By building a relationship network between experts and projects, the collaborative filtering algorithm is used to analyze historical bid evaluation data. Define the rating matrix R of user u, where r ui Describes user u's opinion on bid evaluation expert E i Based on the ratings, a user-based collaborative filtering algorithm is used to recommend experts:

[0193]

[0194] Among them, U is the collection of similar users, and sim(u,v) is the similarity between users.

[0195] Introduce expert evaluation feedback mechanism to optimize the recommendation algorithm. Define the feedback update formula as:

[0196]

[0197] Where f(E i ,P j ) is the adaptation feedback of the project to the expert, and α is the feedback weight.

[0198] The system implements the coordinated use of "real-time dynamic recommendation" (expert recommendation method using PPO to optimize deep reinforcement learning) and "historical preference recommendation" (expert recommendation method using collaborative filtering algorithm) to improve the recommendation effect and the intelligence level of the system: the prior application of the PPO method can adapt to changing project requirements and realize dynamic recommendation based on real-time features, so that the system can make high-quality recommendations even when project requirements may not be completely consistent with historical data; the subsequent use of collaborative filtering further optimizes recommendation decisions by analyzing historical bid evaluation data, especially when historical data has reference value for the current project, so that the recommendation results are more in line with historical preferences.

[0199] (2) Security and authority management module 302

[0200] A security method based on the fusion of blockchain and homomorphic encryption is proposed. Blockchain technology is used to encrypt and distribute the expert information and extraction process, ensuring the security and integrity of the data and preventing data tampering and leakage. Homomorphic encryption technology makes it possible to perform calculations on encrypted data, further protecting the privacy of the data. Specifically, a multi-chain architecture is introduced, combining the design of the main chain and the side chain to achieve hierarchical storage and management of expert information, improving efficiency and privacy protection. The main chain records the key identification information of the expert, and the side chain stores detailed expert data. Among them, the main chain B m Mainly stores the identification information of each expert and its corresponding side chain address:

[0201] B m =(H m-1 ,ID i ,Addr i ,Nonce m )

[0202] Among them, H m-1 Indicates the hash value of the previous block in the main chain, ID i The unique identifier of expert i, Addr i Indicates the sidechain address (pointing to the sidechain with detailed expert data), Nonce m A random number representing the proof of work.

[0203] The detailed information of each evaluation expert is stored in the side chain B s middle:

[0204] B s =(H s-1 ,D i ,T i ,Nonce s )

[0205] Among them, D i Indicates the detailed information of storage expert i (such as professional field, work experience, etc.), T i For timestamp, Nonce s Proof of Work random number for the sidechain.

[0206] This design ensures that only a small amount of key information is disclosed on the main chain, and detailed data is stored on the controlled side chain, reducing the data pressure on the main chain while enhancing the privacy protection of expert information.

[0207] In addition, a dynamic permission management strategy based on the integration of roles and attributes is designed, and a permission management method combining role-based access control and attribute-based encryption is implemented. Different permissions are assigned according to user roles and attributes for operation and data access control, which not only ensures the security of the system, but also improves the flexibility and refinement of permission management, meeting the access needs of different users to the system in different scenarios. Dynamic management is combined with roles and attributes to improve the flexibility and security of the system.

[0208] First, define the permission set P of user U u Depends on its role set R u and attribute set A u The weighted sum model is used to comprehensively consider the impact of user roles and attributes on permissions:

[0209]

[0210] in, and are the weights of roles and attributes. and Represents the set of permissions associated with a role and an attribute, respectively.

[0211] By combining dynamic permission management of roles and attributes, user permissions can be adjusted in real time. Let ΔP be the permission change, and the user permission update is:

[0212]

[0213] Dynamic permission changes are modeled as a set increment operation, described as:

[0214] ΔP={p k ∣p k ∈P and satisfy the new conditions}

[0215] Among them, P represents a permission set, that is, the set of all available permissions in the system.

[0216] In this way, in permission management, changes in user attributes or roles can be responded to quickly.

[0217] (3) Real-time monitoring and auditing function module 303

[0218] In order to cope with the dynamic changes of data in the process of extracting bid evaluation experts, especially when expert information and project requirements are constantly updated, the deep incremental learning model is based on the ability to efficiently adapt to new data while avoiding "catastrophic forgetting". Specifically, the proposed adaptive memory network model can retain the memory of old knowledge while learning new information. The model is divided into two parts: the "main model" and the "memory module". The main model is responsible for learning the current new data, while the memory module stores and maintains the key features of the old data to prevent the model from losing previous knowledge.

[0219] Among them, the "main model" processes and learns new data based on a multi-layer perceptron, and the "memory module" consists of multiple memory slots to store the feature representation of historical data. The content of the memory module is updated as follows:

[0220]

[0221] in, is the state of the memory module at the t+1th round; λ2 is the decay coefficient, which determines the ratio of new information to old information; h(X new ) is the new data X new Through this design, the memory module can retain the sensitivity to historical data for a long time and avoid forgetting previously learned knowledge.

[0222] “Real-time monitoring” is responsible for continuously monitoring the dynamic changes of data during the expert extraction process to ensure that the operations and data processing of each link are transparent. Specifically, a data change detector is first built. Every time the expert information or project requirements are updated, the data change detector captures these changes in real time and records the change information. Suppose at time t, the expert information E t and project requirements t As changes occur, the detector computes the difference vector of these changes on the fly:

[0223] ΔE t =E t+1 -E t , ΔP t =P t+1 -P t

[0224] These changes will trigger the corresponding memory module update mechanism, thus ensuring that the model adapts to new data in a timely manner. In addition, in order to prevent data anomalies or illegal operations, an anomaly detection mechanism is built based on the historical data distribution. By building a distribution of expert information and project requirements and Design anomaly scoring function S abn Determine the distribution of input data x, where N is the number of data samples. Expert information data distribution The mean vector of :

[0225]

[0226] If the anomaly score of a data update exceeds a preset threshold δ (in this embodiment, δ is 70), an alarm is triggered and the audit process begins.

[0227] Next, we design an audit function, the core of which is to ensure the compliance and fairness of the entire expert extraction process. To this end, we introduce a blockchain-based immutable log system to record all operations and data updates related to expert extraction, ensuring the traceability of operations and the transparency of audits. Each update of expert information or project requirements will generate an operation log and write it to the blockchain through the hash function H(log). Assume that the operation log at time t is log t , whose hash value is:

[0228] H(log t )=SHA256(E t ||P t ||ΔE t ||ΔP t )

[0229] The log will record the current expert information, project requirements and their changes, ensuring that the data cannot be tampered with and can be audited afterwards. At the same time, when an audit is required, the system can quickly locate the expert information updates and project requirements changes within a specific time period by querying the blockchain log. The audit module can trace the specific expert extraction process and update operations based on the blockchain records to ensure that every step is traceable.

[0230] The main purpose of the database optimization subsystem 100 is to improve the storage efficiency, processing capacity and query efficiency of the bid evaluation expert database data through data cleaning, compression and table relationship optimization, and provide a solid data foundation for subsequent expert management and extraction. It includes a data cleaning and compression module 101, which cleans, detects outliers, and compresses and stores the data in the database to ensure the storage efficiency and data quality of the database. The high-quality data after data cleaning will directly provide an accurate and simplified data foundation for the subsequent database logic table relationship design module; the compressed data reduces the storage space, improves the response speed of database queries, and provides support for the efficient operation of the entire system. The database logic and table relationship design module 102, the high-quality data transmitted from the data cleaning and compression module, will establish a more optimized table structure and table relationship through this module, and provide a clearer data organization structure for subsequent information retrieval and expert extraction. At the same time, the design of the table structure will provide a clearer data dependency relationship for the subsequent intelligent extraction algorithm.

[0231] The task of the new model construction subsystem 200 of the bid evaluation expert database is to ensure the integrity, accuracy and timely updating of expert information through standardization, multimodal information fusion, real-time updating and other means. It includes an expert information collection standard and automation module 201, which formulates expert information collection standards and uses multimodal information fusion technology to comprehensively and automatically collect expert information. The collected expert information will be used as the data source of the database optimization subsystem to supplement the comprehensiveness and timeliness of the expert information in the expert database. It provides basic data for the subsequent expert information real-time update module, and also provides more accurate expert feature data for the expert extraction algorithm. The information real-time update and responsible person management module 202 ensures the real-time update and security management of expert information through blockchain and edge computing technology. The module ensures the dynamic update of expert information in the expert database, and ensures the immutability and security of data through blockchain technology. The updated expert information will be automatically synchronized to the database optimization subsystem to ensure that the latest expert information is stored in the database. At the same time, it provides the latest and most accurate expert data for the subsequent intelligent extraction subsystem.

[0232] The goal of the bid evaluation expert intelligent extraction subsystem 300 is to dynamically select the most suitable experts to participate in the bid evaluation through deep learning and reinforcement learning algorithms combined with project requirements. It includes an extraction algorithm design module 301, which dynamically extracts the most suitable bid evaluation experts through deep reinforcement learning and knowledge graph. The expert feature data required for expert extraction comes from the "expert information collection standard and automation module" and the "information real-time update module", which provide the latest and most comprehensive information of experts for the extraction algorithm. At the same time, the extraction results will be fed back to the database as bid evaluation history records to further improve the structure of the expert database. The security and authority management module 302 ensures data security and authority management during the expert extraction process through blockchain and homomorphic encryption technology. It works in conjunction with the expert information collection and update module to ensure that the update, storage and access of expert information are carried out in a secure environment. It provides security and privacy for the entire bid evaluation expert database, while ensuring that administrators can flexibly adjust permissions. The real-time monitoring and auditing function module 303 monitors the dynamic changes of data in the process of bid evaluation expert extraction in real time to prevent data loss and "catastrophic forgetting". The real-time monitoring module ensures the consistency and integrity of data during expert extraction and data update, and provides data auditing functions for the system. Combining edge computing and blockchain technology, the transparency and traceability of the expert extraction process are achieved.

[0233] In general, the database optimization subsystem provides an efficient storage and data management foundation for the entire system. The new model construction subsystem of the bid evaluation expert database ensures the comprehensiveness, accuracy and dynamic update of expert information, while the intelligent extraction subsystem uses this data to intelligently extract experts. The modules are organically integrated through data transmission, algorithm collaboration and authority management to ensure that the system can achieve efficient expert database management and intelligent extraction functions in all links.

[0234] like Figure 2As shown, the present invention also discloses a method for extracting and managing intelligent bid evaluation experts. Among them, 21 is an outlier detection and adaptive compression method based on autoencoder and dynamic clustering. This is a data preprocessing stage. It optimizes the original data through anomaly detection, clustering and compression. Its output is a key step for providing a high-quality data foundation for the follow-up. The output data is passed to 22 (database logic and table relationship design method based on semantic association and knowledge graph fusion), 23 (expert information collection method based on multimodal information fusion and reinforcement learning), and 24 (expert information real-time update and intelligent management method based on blockchain and edge computing). 22 is a database logic and table relationship design method based on semantic association and knowledge graph fusion. It receives compressed data from 21, reorganizes the logical relationships and table relationships in the database through the fusion of knowledge graphs, and designs the database logic and table relationships. 23 is an expert information collection method based on multimodal information fusion and reinforcement learning, which integrates information from different data sources to construct a comprehensive information portrait of experts, and the output expert information is transmitted to 24 (real-time update and intelligent management method of expert information based on blockchain and edge computing) and 25 (expert extraction algorithm based on deep reinforcement learning and dynamic knowledge graph). 24 is an expert information real-time update and intelligent management method based on blockchain and edge computing, which uses blockchain technology to ensure the authenticity and immutability of expert information, and performs real-time processing and updates expert information through edge computing. 25 is an expert extraction algorithm based on deep reinforcement learning and dynamic knowledge graph, which learns the matching rules between project requirements and expert skills through deep reinforcement learning algorithm; uses dynamic knowledge graph to associate expert skills with project requirements and perform expert matching. 26 is a security method based on the fusion of blockchain and homomorphic encryption. Blockchain technology ensures the integrity and traceability of data, and uses homomorphic encryption technology to encrypt sensitive data and realize data calculation without decryption. 27 Real-time monitoring and intelligent auditing method based on deep incremental learning, through the deep incremental learning model, dynamically monitors the expert information update and extraction process; monitors the operation of the system in real time, and conducts intelligent audits on expert data. In addition, during the real-time monitoring and auditing process, if potential security vulnerabilities are found, such as insufficient encryption mechanism, possible leakage risks in data transmission, inconsistencies in nodes in the blockchain, etc., the feedback mechanism will pass these problems to 26. If it is found that the expert information is not updated in time, the management is not perfect, or there is information lag in the edge computing node. These problems will be passed to 24 through the feedback mechanism, prompting it to self-correct.

[0235] The above is only a specific implementation of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial changes to the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. An intelligent bidding expert extraction and management system, characterized in that: include: The data cleaning and compression module pre-processes the original bid evaluation expert database data and transmits the pre-processed compressed data to the database logic and table relationship design module, the expert information collection standard and automation module, and the information real-time update and person in charge management module; The preprocessing includes outlier detection and data compression; The database logic and table relationship design module classifies and stores the received compressed data in different tables, and builds the relationship between different tables in the database based on semantic association and knowledge graph; Expert information collection standard and automation module receives compressed data, fuses multimodal information in the compressed data to obtain expert information, and outputs expert information to the information real-time update and person in charge management module and the extraction algorithm design module; The real-time information update and person-in-charge management module monitors the work dynamics and professional skill changes of bid evaluation experts in real time, and automatically updates the database based on the received compressed data and expert information; Extract algorithm design module, receive expert information, learn the matching rules between project requirements and expert skills in expert information through deep reinforcement learning algorithm; use dynamic knowledge graph to associate expert skills and project requirements; Experts are extracted by combining the expert recommendation method based on collaborative filtering algorithm.

2. The intelligent bidding expert extraction and management system according to claim 1 is characterized in that: The outlier detection is specifically as follows: An autoencoder is used to capture the high-dimensional features in the original bid evaluation expert database data; the loss function of the autoencoder includes a sparse regularization term, and the formula of the sparse regularization term is expressed as: Where z represents the potential representation vector output by the encoder part of the autoencoder, that is, a representation vector in the low-dimensional space to which the input data is mapped after passing through the encoder; λ represents the sparsity regularization coefficient; m represents the dimension of the potential representation vector z, that is, the dimension of the low-dimensional space after the autoencoder compresses the input data; z i represents the absolute value of the representation of the i-th data point in the latent space; || represents the absolute value; Perform dynamic clustering operations, dynamically adjust the number of clusters, and adaptively adjust the number and boundaries of clusters according to the structural changes of the data. The dynamic adjustment formula of the number of clusters is expressed as: Among them, α represents the control coefficient, represents density entropy; represents the change in density entropy; k old represents the number of clusters before adjustment; k new represents the adjusted number of clusters; In each iteration of the dynamic clustering operation, the cluster center is dynamically updated according to the data distribution, and the formula is expressed as: Among them, w ij Represents data point z i Similarity weight with cluster center; represents the updated cluster center; n represents the total number of data points; j represents the index of the current cluster, j = 1, 2, ..., k, k represents the number of clusters; Calculate data point z i The minimum similarity score is expressed as: Among them, min represents the minimum value; The minimum similarity score of all data points is calculated, and data points with a minimum similarity score lower than a preset similarity score threshold are considered outliers.

3. The intelligent bidding expert extraction and management system according to claim 1 is characterized in that: The data compression is specifically as follows: The data is compressed using an adaptive feature dictionary coding compression algorithm. The dictionary structure and coding strategy are dynamically adjusted by real-time analysis of the data's features and frequent patterns. The dictionary update rule is expressed as: Among them, η represents the learning rate; represents the gradient of the dictionary; C new represents the sparse coding coefficient of the new data; γ represents the smoothness control parameter of the dictionary update; x new Indicates a new batch of input data; D new represents the updated dictionary matrix; D represents the current dictionary matrix; D old Represents the old dictionary matrix; represents the square of the Frobenius norm; In addition, a dictionary pruning strategy is adopted to reduce the size of the dictionary. By analyzing the contribution of each basis vector in the dictionary, redundant or unimportant basis vectors are removed, thereby optimizing the dictionary structure.

4. The intelligent bidding expert extraction and management system according to claim 1 is characterized in that: The received compressed data is classified and stored in different tables, and the association relationship between different tables in the database is constructed based on semantic association and knowledge graph, as follows: Design the table structure and classify and store the bid evaluation expert information according to its characteristics. Each classification is stored in a different table in the database. The classification includes basic information of the expert, professional skills and bid evaluation history. Establish primary key-foreign key constraints for different tables in the database; Establish cross-reference or index relationships between experts with high semantic association in the database; Build a knowledge graph and use the adjacency matrix to represent the associations between tables in the knowledge graph; The knowledge graph is represented as a directed graph G = (V, E); wherein the node set V includes experts, fields, skills and evaluation items; the edge set E represents the relationship between the nodes in the node set; when there is new data, a new node is added to the knowledge graph. If the similarity between the semantic representation of the new data and the nodes in the existing knowledge graph is higher than a preset dynamic threshold, it indicates that the new data has a semantic association with the existing graph nodes, and the edge set is updated; the method for calculating the similarity between the semantic representation of the new data and the nodes in the existing knowledge graph is the same as the method for calculating the semantic association.

5. The intelligent bidding evaluation expert extraction and management system according to claim 4 is characterized in that: The semantic relevance is specifically as follows: The semantic similarity between two bid evaluation experts is measured by the cosine similarity of the semantic vectors of the two bid evaluation experts; the semantic vector is a vector representation of the expert attributes mapped in the semantic space; the expert attributes include fields, skills and experience; when the cosine similarity is higher than the set similarity threshold, it is considered that the two experts are semantically associated.

6. The intelligent bidding expert extraction and management system according to claim 1 is characterized in that: The multimodal information in the fused compressed data is specifically as follows: Extract text features H from text modality expert information in compressed data using a pre-trained language model T , using convolutional neural networks to extract image features H from image modality expert information in compressed data I ; The text features and image features are deeply integrated through a two-layer interaction mechanism: At the low-level feature interaction layer, a cross-modal attention mechanism is introduced. The text features are used as the query parameters of the cross-modal attention mechanism. The image features are used as the key parameters and value parameters of the cross-modal attention mechanism for weighted calculation. The global attention information A of the text modality is obtained. TI ; The image features are used as the query parameters of the cross-modal attention mechanism, and the text features are used as the key parameters and value parameters of the cross-modal attention mechanism for weighted calculation to obtain the global attention information A of the image modality. IT , which is expressed as follows Among them, d T Indicates H T The characteristic dimension of Indicates H T The transpose of ; softmax represents the softmax function; Intermediate feature fusion layer, the intermediate fusion feature is expressed as: Fmi d =a T H T +α I H I Among them, α T =σ(W T H T ) represents the weight of the text modality, α I =σ(W I H I ) represents the weight of the image modality, W T and W I represents the learnable weight matrix, σ represents the Sigmoid activation function; The modality reconstruction loss is introduced in the fusion process, and the reconstruction loss function is expressed as: in, g T represents a network for reconstructing text modalities; g I represents the network used to reconstruct the image modality; || || 2 represents the square of the Frobenius norm; The final fusion feature representation of text modality and image modality is obtained, and the formula is expressed as: H final =F mid +A TI +A IT 7. The intelligent bidding expert extraction and management system according to claim 1 is characterized in that: The automatically updated database is based on dynamic weight blockchain and smart contract technology, as follows: Construct a dynamic weight mechanism for expert information so that blockchain nodes prioritize high-weight expert information update requests; the weight w of the expert’s j-th category of expert information at time t j (t) is expressed as: Among them, α j represents the initial weight; β j Controls how fast the weights change over time; Δt j Indicates the time interval since the last update of the same type of information; An adaptive block generation strategy is introduced to adjust the time interval and size of block generation according to the current network load and the weight of expert information; the time interval ΔT of blockchain block generation is expressed as: Among them, T0 represents the basic block generation time interval of the system; γ1 represents the adjustment coefficient; N represents the total number of experts, and M represents the total number of information types; Smart contracts are used to automatically control the update process of expert information. The triggering rules of smart contracts are based on the weight value priority of expert information. Each type of information of each expert corresponds to a smart contract. When the weight of a piece of information exceeds the preset weight threshold, the smart contract is automatically triggered to update the information and synchronize it to the blockchain. The smart contract adopts a batch processing mechanism to allow multiple low-weight information to be updated together at the same time.

8. The intelligent bidding expert extraction and management system according to claim 1 is characterized in that: Expert extraction based on deep reinforcement learning and dynamic knowledge graph is as follows: Construct a dynamic knowledge graph and define the edge weight w between the expert node and the project node to be evaluated in the dynamic knowledge graph ij , which is used to measure the degree of adaptation between expert nodes and project nodes to be evaluated; the edge weight calculation formula is expressed as: Among them, λ k represents the weight coefficient; It represents the similarity between the expert node and the project node to be evaluated in the kth dimension, which is calculated using cosine similarity; K represents the number of characteristic attributes of each expert; Based on the constructed dynamic knowledge graph, deep reinforcement learning is used to model expert features and dynamically adjust the extraction strategy. A reward function is designed to evaluate the quality of expert selection. The reward function is expressed as: Among them, v i represents the feature vector of the expert node; f(v i , p j ) represents the expert node's evaluation of the bid evaluation project node p j The degree of matching; t represents the action space, that is, the set of experts that best suits the current project is selected at time t; C(a t ) represents the cost of selecting the expert set; δ represents the penalty coefficient; Q represents the number of evaluation items; s t It represents the state representation, i.e., the characteristic matrix of experts and evaluation items; The deep reinforcement learning is optimized using a proximal strategy optimization method, and dynamically adjusts and selects expert strategies, so that the system can adaptively select a suitable expert combination under different business needs, thereby maximizing the compatibility between experts and the projects to be evaluated; in the process of dynamically adjusting and selecting expert strategies, a quantum random number generator is used to generate random numbers.

9. The intelligent bidding expert extraction and management system according to claim 1 is characterized in that: It also includes a security and authority management module, which uses blockchain technology to homomorphically encrypt and distribute the expert information in the expert information collection standard and automation module and the evaluation expert extraction process in the extraction algorithm design module; and allocates operation permissions and data access permissions based on user roles and attributes.

10. The intelligent bidding expert extraction and management system according to claim 1 is characterized in that: It also includes a real-time monitoring and auditing function module, which monitors the update of expert information and the process of extracting bid evaluation experts in real time, updates the data changes in the process of extracting bid evaluation experts through a deep incremental learning model, and conducts intelligent auditing of expert data; During the real-time monitoring and auditing process, potential security vulnerabilities discovered will be transmitted to the security and authority management module through the feedback mechanism, and information updates and management issues discovered will be transmitted to the information real-time update and responsible person management module through the feedback mechanism for self-correction.

Images