Project intelligent management analysis platform based on security risk standardization

Abstract

The invention relates to the technical field of security risk standardization management, and discloses an intelligent project management analysis platform based on security risk standardization, comprising: a hierarchical semantic space representation model construction module for constructing multi-level semantic representation and outputting a hierarchical semantic tree; the meta-learning-driven cross-domain mapping function construction module is used for receiving the hierarchical semantic tree and constructing a cross-domain mapping function to realize cross-domain knowledge migration; the semantic hierarchy prototype network construction module is used for receiving the output of the cross-domain mapping function and mapping a complex risk structure; efficient knowledge migration is carried out on output of the sample rapid adaptation implementation module and the construction module; the standardized risk knowledge base construction and application module is used for receiving the output of the sample rapid adaptation realization module and realizing risk standardization and application; through the combination of meta-learning and the semantic hierarchy prototype network, the sample number and the sample demand required by new field adaptation are reduced, and the new field risk standardization construction period is shortened.

Description

Technical Field

[0001] The present invention relates to the technical field of safety risk standardization management, and more specifically, to a project intelligent management and analysis platform based on safety risk standardization. Background Art

[0002] With the deepening development of global economic integration and the diversification of corporate business, more and more companies are entering multiple business areas. Cross-domain security risk management has become a key challenge for corporate sustainable development. In this context, the establishment of a unified and adaptable security risk standardization system has become increasingly important.

[0003] Currently, security risk standardization primarily relies on rule-based expert systems and traditional machine learning methods. While rule-based approaches perform well in specific domains, their rules, often manually developed by domain experts, are highly domain-specific and difficult to directly apply to other areas. When companies expand into new business areas, they often need to rebuild their entire rule system, resulting in high costs and a lengthy process.

[0004] Traditional machine learning methods also have significant shortcomings in risk standardization. These methods typically employ supervised learning, requiring extensive amounts of labeled data for training. Existing research shows that traditional transfer learning methods often require thousands or even tens of thousands of labeled examples to achieve acceptable performance when adapted to new domains. This high data demand creates significant data collection and labeling costs for companies rapidly expanding into new businesses.

[0005] Existing technologies generally have limitations when it comes to processing complex risk structures. Most methods treat risk description text as a flat sequence, using bag-of-words models or simple sequence encoding, ignoring the rich hierarchical structure information in risk descriptions.

[0006] Domain adaptability is another key technical bottleneck. Existing cross-domain approaches often face a dilemma: either strictly adhere to the source domain's standards, losing the unique expressions of the target domain; or completely rebuild the target domain's standards, making them incompatible with existing enterprise risk management systems. This technical limitation prevents enterprises from establishing a unified risk management system during business expansion.

[0007] In terms of semantic understanding and generalization, existing technologies primarily rely on surface features of text for risk identification and classification. When different fields use different terms to describe similar risk concepts, such as "equipment failure" in manufacturing and "system downtime" in IT, both essentially refer to equipment unavailability risks, existing methods often fail to identify these deeper semantic connections, resulting in poor cross-domain knowledge transfer.

[0008] Furthermore, existing technologies also suffer from shortcomings in parameter efficiency and computational resource consumption. Traditional methods often require retraining a large number of parameters to adapt to a new domain, which not only takes a long time to train but also places high demands on computing resources. This high resource consumption makes existing technologies difficult to apply in resource-constrained environments and hinders rapid business expansion for enterprises.

[0009] In summary, existing security risk standardization technologies have shortcomings in cross-domain adaptability, complex structure processing, semantic understanding and generalization, and resource efficiency. They are unable to meet the actual needs of modern enterprises for diversified development and rapid business expansion. Therefore, a new technical solution is urgently needed to address these issues. Summary of the Invention

[0010] The present invention provides a project intelligent management and analysis platform based on security risk standardization, which solves the technical problems in related technologies such as poor adaptability of cross-domain risk standardization, high data requirements, and weak ability to handle complex structures.

[0011] The present invention provides a project intelligent management and analysis platform based on security risk standardization, including: Hierarchical semantic space representation model building module, constructing multi-level semantic representation and outputting hierarchical semantic tree; A meta-learning-driven cross-domain mapping function construction module receives a hierarchical semantic tree and constructs a cross-domain mapping function to achieve cross-domain knowledge transfer; A semantic hierarchical prototype network building module that receives the output of the cross-domain mapping function and maps the complex risk structure; The sample rapid adaptation implementation module performs efficient knowledge transfer based on the output of the semantic hierarchical prototype network construction module; The standardized risk knowledge base construction and application module receives samples, quickly adapts the module output, and realizes risk standardization and application.

[0012] Furthermore, the hierarchical semantic space representation model construction module includes semantic hierarchy definition, recursive neural parser, hierarchical feature extraction algorithm and semantic space pre-training. The recursive neural parser parses the risk text into a hierarchical semantic tree structure based on a tree-structured long short-term memory network.

[0013] Furthermore, the meta-learning-driven cross-domain mapping function construction module includes task construction and sampling, model-independent meta-learning algorithm implementation, hierarchical adaptation loss function construction and gradient path optimization. The model-independent meta-learning algorithm is based on the MAML framework and trains the mapping function through a two-layer optimization process.

[0014] Furthermore, the loss function constructed by the hierarchical adaptation loss function is: ; in, Indicates that the parameter is The hierarchical adaptation loss function of the mapping function; represents the summation operator, is the semantic level index, is the total number of semantic levels; Indicates the The weights of the semantic levels are adaptively adjusted according to domain differences; For the Layer source domain semantic representation and target domain semantic representation Semantic distance measurement function between them; and Represents the source domain and target domain in Semantic representation of layers.

[0015] Furthermore, the semantic hierarchical prototype network construction module includes multi-level prototype representation construction, hierarchical attention method implementation, prototype update and maintenance, and cross-level relationship modeling. The multi-level prototype representation maintains a set of prototype vectors for each semantic level, and each prototype vector represents a typical semantic pattern.

[0016] Furthermore, the semantic hierarchical prototype network includes a prototype representation layer, a cross-level association layer and an adaptation mapping layer. The cross-level association layer adopts a graph attention network structure to organize the prototypes of each level into a directed graph structure, and the edges represent the association strength between prototypes of different levels.

[0017] Furthermore, the sample rapid adaptation implementation module includes small-scale sample learning strategies, efficient parameter fine-tuning, hierarchical progressive fine-tuning and zero-sample generalization capability enhancement, achieving efficient transfer of new domain knowledge through 5 to 10 risk description samples.

[0018] Furthermore, the efficient parameter fine-tuning adopts low-rank adapter technology, which adapts to the characteristics of specific fields by inserting adaptation layers between the original network layers, with the number not exceeding 10% of the original network layers; The hierarchical progressive fine-tuning method implements a top-down parameter adjustment method that starts from high-level semantics and adjusts parameters layer by layer according to a preset hierarchical order to low-level semantics.

[0019] Furthermore, the standardized risk knowledge base construction and application module includes the construction of a two-way mapping relationship library, an automatic risk standardization processing flow, a human-computer collaborative verification and correction system, and a diversified project management application interface. The two-way mapping relationship library adopts a hierarchical triple structure, and each mapping relationship contains a hierarchical identifier, a relationship type, and a credibility score.

[0020] The present invention provides a computer storage medium comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the above-mentioned project intelligent management and analysis platform based on security risk standardization.

[0021] The beneficial effects of the present invention are: through the combination of meta-learning and semantic hierarchical prototype networks, the number of samples required for new domain adaptation is reduced from thousands of samples in traditional methods to 5 to 10, which reduces the sample demand and shortens the construction period of new domain risk standardization; Based on a recursive neural parser and hierarchical semantic representation, the system can accurately process complex risk descriptions with multiple levels of nesting, and compared with flat text processing methods, the accuracy of preserving complex risk structures is improved; Through a hierarchical adaptation strategy, risk standardization is achieved while retaining domain characteristics, ensuring the comparability of risks across domains while retaining some domain-specific representations to support diversified management. Based on the semantic hierarchical prototype network and zero-shot generalization module, the system can accurately identify similar risks expressed in different terms and improve the accuracy of mapping unseen risk concepts. Through efficient parameter fine-tuning and gradient path optimization, the system reduces the number of parameters required to adapt to new domains, reduces computing resource requirements, and supports operation in resource-limited environments; Combining meta-learning and hierarchical progressive fine-tuning, the adaptation time to new domains is shortened and the adaptation speed is improved; Through the human-machine collaborative verification and correction system, the overall accuracy of standardized results is improved; Based on hierarchical semantic representation and semantic hierarchical prototype network, the system can provide clear risk mapping paths and basis, enhancing the transparency and explainability of the standardization process. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flow chart of the project intelligent management and analysis platform based on security risk standardization in the present invention; Figure 2 This is a bar chart comparing the accuracy of the method of the present invention and the flat text processing method when processing complex risk structures; Figure 3 This is a bar chart comparing the accuracy of unseen risk concept mapping using the method of the present invention and traditional methods; Figure 4 is a bar chart showing the adaptation effect of the method of the present invention in the transfer tasks between five different fields; Figure 5 It is a bar chart of the weight distribution of each level in the migration of different fields of the present invention; Figure 6It is a bar chart comparing the number of samples required for the method of the present invention and the traditional method to adapt to the new field; Figure 7 It is a bar chart comparing the time taken to complete new domain adaptation by the method of the present invention and the traditional method; Figure 8 It is a bar chart comparing the accuracy of standardized results between the human-machine collaborative method of the present invention and the purely automated method. DETAILED DESCRIPTION

[0023] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0024] At least one embodiment of the present invention discloses a project intelligent management and analysis platform based on security risk standardization, such as Figure 1 Shown, including: Hierarchical semantic space representation model building module, constructing multi-level semantic representation and outputting hierarchical semantic tree; The method provided in this application first constructs a hierarchical semantic space representation model to provide a structured semantic representation foundation for subsequent cross-domain mapping. It specifically includes the following sub-steps: Step 1.1, semantic level definition; According to an embodiment of the present application, a multi-level semantic space structure is defined: ; in Representing a hierarchical semantic space; Represents the most abstract semantic layer, mainly containing general risk concepts that are independent of the domain; Indicates ratio A more specific semantic layer; Represents the most specific semantic layer, containing domain-specific detailed risk descriptions; the middle layer arrive It represents the transition level from abstract to concrete, and each layer carries more detailed semantic information, realizing the layer-by-layer refinement and decomposition of semantics. Indicates the total number of semantic levels, which can be flexibly set according to the actual application scenario.

[0025] The semantic representation of each layer is stored as a feature vector, and the dimension increases as the layer goes deeper to accommodate more detailed information.

[0026] It should be noted that, in some implementations, the number of semantic levels Adjustments can be made based on the complexity of the application scenario. For example, for simple risk descriptions, a three-layer structure (abstract layer, intermediate layer, and concrete layer) can be used; for complex risk descriptions, a five-layer or higher structure can be used to more finely model semantic hierarchical relationships. Furthermore, the vector dimensions at different levels can be flexibly configured to balance expressiveness and computational efficiency.

[0027] Optionally, in another embodiment of the present application, semantic hierarchical definitions can be combined with domain expert knowledge to design specialized hierarchical structures for specific industry domains. For example, in the financial risk domain, a five-layer structure consisting of "risk category - risk subcategory - risk factor - risk event - risk impact" can be defined; while in the manufacturing domain, a five-layer structure consisting of "production link - equipment system - failure type - failure manifestation - failure consequence" can be defined. Through customized hierarchical definitions, the model's adaptability to specific domains can be further improved.

[0028] Step 1.2, recursive neural parser construction; According to one embodiment of the present application, a recursive neural parser based on a tree-structured long short-term memory (Tree-LSTM) network is used to parse risk text into a hierarchical semantic tree. This parser receives the original risk description as input and, through three phases of lexical analysis, syntactic analysis, and semantic analysis, outputs a multi-layered semantic tree structure. The parser specifically focuses on causal, conditional, and modifier relationships within risk descriptions, breaking down complex, nested risk descriptions into a clear hierarchical structure.

[0029] The tree-structured long short-term memory network recursive neural parser provided in this application has the following structure and implementation: The Tree-LSTM recursive neural parser consists of an input layer, a tree-structured encoding layer, an attention layer, and an output layer. The input layer first tokenizes the text and performs word embedding, mapping each word into a dense vector representation. The tree-structured encoding layer is the core of the model, using an improved Child-SumTree-LSTM unit to explicitly model the dependencies in the syntactic tree. Each Tree-LSTM unit contains an input gate, an output gate, a forget gate, a memory cell, and a hidden state. Unlike traditional LSTMs, Tree-LSTMs equip each child node with a separate forget gate, enabling the model to more precisely control the retention and discarding of information from different child nodes.

[0030] In the specific scenario of risk text processing, this recursive neural parser first applies a pre-trained dependency parser to generate an initial syntactic tree. It then prunes and restructures the tree using a domain-specific lexicon and grammatical rules to enhance its ability to recognize risk terminology and unique sentence structures. Furthermore, the model incorporates a hierarchical attention approach, calculating attention weights both vertically (tree depth) and horizontally (nodes on the same level), enabling the model to simultaneously focus on local details and global structure. The lowest level of the tree (leaf nodes) processes specific entities and terms, the middle levels handle conditional and attribute information, and the top level handles abstract concepts such as risk category and risk severity.

[0031] To enhance the model's ability to handle cross-domain risk descriptions, the parser also incorporates a semantic role labeling module specifically for risk representations. This module identifies key components in risk descriptions, such as event triggers, risk objects, consequences, conditions, and responsible parties, and encodes this information into the corresponding nodes of the tree. To handle complex nested risk descriptions, the model employs a recursive tree synthesis strategy, first processing smaller semantic units and then gradually combining them into larger semantic structures. This effectively addresses the difficulties traditional sequence models face in handling long-range dependencies.

[0032] Through the above structure and implementation method, the Tree-LSTM recursive neural parser can accurately extract hierarchical risk knowledge from complex risk descriptions such as the text "When the production line in workshop 5 loses power, if the backup power supply fails to start and the operator does not switch to manual mode in time, the high-temperature equipment cooling system may stop working, causing equipment damage and production interruption." This includes event conditions (power outage, backup power supply failure, failure to switch to manual mode), risk objects (high-temperature equipment cooling system), risk consequences (equipment damage, production interruption), etc., forming a structured semantic representation.

[0033] It should be understood that in other embodiments, other types of recursive neural network structures, such as recursive neural tensor networks (RNTN) or graph convolutional networks (GCN) can be used to construct recursive neural parsers, as long as they can effectively parse risk text into a hierarchical semantic tree structure.

[0034] Step 1.3, semantic representation extraction; According to the embodiments of this application, a hierarchical feature extraction algorithm is applied to the semantic tree generated by parsing, extracting corresponding semantic representations at different levels. The bottom layer extracts specific entities, attributes, and numerical features; the middle layer extracts domain concepts, relationships, and conditional features; and the top layer extracts abstract risk categories and pattern features. Features from all levels are integrated using an attention-weighted algorithm to form a complete multi-level semantic representation.

[0035] In some implementations, feature fusion strategies can be employed to combine features at different levels through residual connections or gating units to enhance the completeness and robustness of the representation. Furthermore, external knowledge bases (such as domain risk ontologies or specialized dictionaries) can be introduced to assist in feature extraction, improving the accuracy and richness of the semantic representation.

[0036] like Figure 2 As shown, the accuracy comparison between the method of the present invention and the flat text processing method in processing complex risk structures is shown. The accuracy of the method of the present invention in processing complex risk structures reaches 81%, while the flat text processing method is only 45%, an improvement of 80%.

[0037] Step 1.4, semantic space pre-training; According to one embodiment of the present application, a semantic space representation model is pre-trained using a large-scale, cross-domain risk corpus, enabling the model to capture common risk semantic features. A contrastive learning approach is employed to enable the model to learn to distinguish between similar and unrelated risks. Furthermore, a masked language modeling task is used to enhance the model's semantic understanding capabilities.

[0038] Optionally, adversarial training strategies can be employed during pre-training to enhance the model's robustness through adversarial examples, enabling it to handle subtle changes in semantic expression. Furthermore, a multi-task learning framework can be combined to simultaneously perform multiple pre-training tasks (such as context prediction and relationship prediction) to further enhance the model's semantic understanding capabilities.

[0039] A meta-learning-driven cross-domain mapping function construction module receives a hierarchical semantic tree and constructs a cross-domain mapping function to achieve cross-domain knowledge transfer; This application further provides a meta-learning driven cross-domain mapping function construction method for realizing knowledge transfer that can quickly adapt to new domains. It should be understood that this module uses a meta-learning strategy to train the model to quickly adapt to new tasks under few sample conditions. In addition, this module is closely integrated with the previous module, using hierarchical semantic representation as the input and output of the mapping function. Specifically, it includes the following sub-steps: Step 2.1, task construction and sampling; According to an embodiment of this application, the cross-domain risk normalization problem is formalized as a set of meta-learning tasks. Each task consists of a support set and a query set. The support set contains a small number of examples (typically 5 to 10) mappings from the source domain to the target domain, while the query set is used to evaluate the model adaptation effect. A large number of such tasks are constructed from historical domain adaptation data to form a meta-learning training dataset.

[0040] It should be noted that during the task construction process, a stratified sampling strategy can be adopted to ensure that the constructed tasks can cover different types of inter-domain transfer scenarios (such as transfer between similar domains, transfer between domains with significant differences, etc.). In addition, a difficult sample mining method can be introduced to prioritize sample construction tasks that are difficult to map, thereby enhancing the generalization ability of the model.

[0041] Step 2.2, implementation of model-independent meta-learning algorithm; According to one embodiment of the present application, a meta-learning framework is constructed based on the MAML (Model-Agnostic Meta-Learning) algorithm. The framework trains the mapping function through a two-layer optimization process: the inner layer optimizes to quickly adapt the parameters to a specific task, and the outer layer optimizes to adjust the initial parameters of the model to make it adapt to the new task as quickly as possible. Receive a hierarchical semantic representation of the source domain As input, it outputs the corresponding representation of the target domain .

[0042] in Indicates that the parameter is The mapping function of Represents the parameters of the mapping function; Representing the hierarchical semantic representation of the source domain; Represents a hierarchical semantic representation of the target domain.

[0043] The MAML-based cross-domain mapping function provided in this application has the following implementation details: Mapping Function This transformer is composed of a multi-layer neural network, consisting of a feature encoding layer, a mapping transformation layer, and a feature decoding layer. The feature encoding layer uses a self-attention module to process the hierarchical semantic representation of the input, capturing dependencies within and between layers. The mapping transformation layer consists of multiple residual blocks, each of which includes a domain adaptation layer and a domain-invariant feature extraction layer. The feature decoding layer reconstructs the semantic representation of the target domain and maintains the consistency of the hierarchical structure through a cross-layer attention module.

[0044] In the meta-training phase, for each task , the algorithm performs the following steps: Sampled support set: ; in Indicates the The support set of each task; Indicates the Hierarchical semantic representation of source domain samples; Indicates the Hierarchical semantic representation of target domain samples; Indicates the number of samples in the support set, usually 5-10; represents the support set sample index; Based on the support set and the current mapping function parameters, the system calculates the task-specific parameter updates through gradient descent. First, the system calculates the loss function value of the current mapping function on the task-specific support set, which represents the performance of the current model on the task. Then, the gradient of the loss function with respect to the model parameters is calculated. This gradient indicates the direction of change of the loss function in the parameter space. Next, the system shifts the parameters in the opposite direction of the gradient by a small step, with the step size controlled by the inner learning rate. This step adjusts the parameters in the direction of reducing the task-specific loss function. Finally, the system obtains the optimized parameter values ​​for the specific task, which are more suitable for handling the domain mapping requirements of the current task. Sample query set to evaluate the updated parameter performance: ; in Indicates the The query set of tasks; Indicates the number of samples in the query set; Indicates the starting value of the query set sample index; Indicates the end value of the query set sample index; Indicates the Hierarchical semantic representation of source domain samples; Indicates the Hierarchical semantic representation of target domain samples.

[0045] Based on the query set performance of multiple tasks, the initial parameters are updated: the system first calculates the loss function value on the query set of each task separately. These values ​​reflect the generalization ability of the model parameters optimized for a specific task on unseen samples; then the loss function values ​​of all tasks are accumulated to obtain a comprehensive loss value, which represents the overall adaptability of the model's initial parameters; then the gradient of this comprehensive loss with respect to the initial parameters is calculated, indicating how to adjust the initial parameters to improve the overall performance of the model on multiple tasks; finally, the system updates the initial parameters in the opposite direction of the gradient, with the step size controlled by the outer learning rate, completing a meta-learning update iteration; this two-layer optimization design enables the model to learn a good parameter initial point, so that it can quickly adapt to new tasks with only a small number of samples.

[0046] This application makes two key improvements to the standard MAML algorithm: Introducing a task similarity perception module to set dynamic weights for different tasks, so that the model pays more attention to source tasks that are similar to the target domain; A hierarchical optimization strategy is adopted, and parameters at different semantic levels use different learning rates to more flexibly adapt to the migration difficulty of each level.

[0047] In a specific application scenario, when migrating from the petrochemical industry to the natural gas processing sector, the algorithm can identify both similarities and differences in risk classification systems and terminology between the two fields. For example, from a small number of mapping examples, it can learn the semantic correspondence between "valve leakage leading to flammable gas accumulation" and "pressure vessel seal failure leading to gas escape." It can then generalize to unseen examples, such as correctly mapping "aging of oil tank floating roof seals leading to oil and gas volatilization" to "deterioration of gas tank gaskets leading to gas loss" in the natural gas sector.

[0048] In other embodiments, other meta-learning algorithms such as Prototypical Networks or Relation Networks can be used, with the most appropriate method selected based on the characteristics of the specific application scenario. Furthermore, these algorithms can be combined with traditional domain adaptation methods (such as domain adversarial training) to further enhance the model's cross-domain transfer capabilities.

[0049] like Figure 3 The figure shows a comparison of the accuracy of the proposed method and the traditional method when processing unseen risk concept mapping. The proposed method has an accuracy of 65% for unseen risk concept mapping, while the traditional method only has an accuracy of 25%, which demonstrates the generalization ability of the proposed method. like Figure 4 The figure shows the adaptability of the proposed method in five different domain migration tasks. As can be seen from the figure, the proposed method demonstrates high adaptability in migration tasks across diverse domains, including manufacturing to intelligent manufacturing, petrochemicals to natural gas processing, construction engineering to road engineering, financial risk to insurance risk, and medical equipment to biopharmaceuticals, with adaptation accuracy ranging from 83% to 94%.

[0050] Step 2.3, constructing the hierarchical adaptation loss function; According to an embodiment of the present application, a hierarchical perception adaptation loss function is constructed: ; in, Indicates that the parameter is The hierarchical adaptation loss function of the mapping function; represents the summation operator, is the semantic level index, is the total number of semantic levels; Indicates the The weights of each semantic level can be adaptively adjusted according to domain differences; For the Layer source domain semantic representation and target domain semantic representation Semantic distance measurement function between them; and Represents the source domain and target domain in Semantic representation of layers.

[0051] This loss function achieves comprehensive adaptation optimization of different semantic levels by weighted summation of semantic differences at each level, thereby improving the accuracy and robustness of cross-domain knowledge transfer.

[0052] Distance metric function Different calculation methods are used according to the different characteristics of the semantic level, and the specific implementation is as follows: For the most abstract top-level semantics, the calculation method based on Wasserstein distance is used, that is, when hour: ; in A distance metric representing top-level semantics; Represents the most abstract top-level semantic representation of the source domain; Represents the most abstract top-level semantic representation of the target domain; represents the 2-Wasserstein distance function; represents the probability distribution of the source domain; Represents the probability distribution of the target domain; For the intermediate semantic level, a structure-sensitive graph distance metric is used, that is, when When the form is: ; in Indicates the Layer source domain semantic representation and target domain semantic representation The semantic distance metric function between ); Indicates the source domain Semantic representation of layers; Indicates the target area Semantic representation of layers; Represents the adjacency matrix of the source domain; An adjacency matrix representing the target domain; represents the Frobenius norm; represents the balance coefficient; The feature matrix representing the source domain; A feature matrix representing the target domain; represents the L2 norm; For the most specific underlying semantics, the context-enhanced cosine distance is used, that is, when When , it is expressed as: ; in A distance metric representing the underlying semantics; Represents the most specific underlying semantic representation of the source domain; Represents the most specific underlying semantic representation of the target domain; represents the cosine similarity function; represents the context similarity adjustment factor function; Represents contextual information of the source domain; Represents the contextual information of the target domain.

[0053] By adaptively adjusting the weights of different layers This allows the model to dynamically focus on different semantic levels based on domain differences. For example, for domains with significant terminology differences but similar risk logic, the bottom layer weights are increased; for domains with similar basic concepts but significant differences in risk frameworks, the top layer weights are increased.

[0054] The layer adaptation loss function provided in this application has the following implementation details: The hierarchical adaptation loss function is an adaptive composite loss function that is specially constructed based on the characteristics of the semantic hierarchy. It adopts the form of weighted combination and integrates multiple measurement methods: for the top-level abstract semantics, Wasserstein distance is mainly used to measure the differences in risk category distribution; for the middle-level relational semantics, graph structure similarity measurement methods such as GraphMatchingLoss are used to evaluate the structural differences in risk relationship networks; for the bottom-level specific semantics, cosine similarity and context-related embedding distance are combined to accurately capture the differences in term and entity representations.

[0055] Adaptive adjustment of layer weights is achieved using an attention algorithm. The specific calculation process is as follows: First, the system calculates an importance score for each semantic layer, which reflects the current layer's importance in domain adaptation. Then, the importance scores of all layers are indexed, converting the scores to positive values ​​while amplifying the differences. Next, the sum of the indexed scores of all layers is calculated as a normalization factor. Finally, the indexed score of each layer is divided by the normalization factor to obtain the weight value of that layer, and the sum of all layer weights is 1. This soft allocation method enables the system to dynamically adjust the attention paid to different semantic layers, rather than using fixed weights or a hard selection mechanism.

[0056] The importance score of each level is calculated by a dedicated importance evaluation network, which comprehensively considers the semantic representation of the source and target domains at the current level, as well as the global state information of domain adaptation.

[0057] The evaluation network receives as input the semantic representations of the source and target domains at a specific level, as well as the global state of the current domain adaptation, and outputs an importance score for that level. The specific implementation utilizes a multi-layer perceptron network structure, with the following processing flow: First, the system compares the semantic representations of the source and target domains through various methods, including direct difference calculations, element-wise multiplication operations, and concatenation operations, to construct a comprehensive difference feature vector. These difference features are then concatenated with global state information (such as domain similarity, adaptation difficulty, and current adaptation progress) to form a complete input feature vector. These features are then nonlinearly transformed through a first-layer fully connected neural network, and the ReLU activation function is applied to extract key patterns. Finally, they are further processed by a second-layer neural network to output a scalar value as the importance score for the current level. The introduction of global state information enables the system to dynamically adjust the importance of each level based on the overall adaptation progress, achieving a more flexible and adaptive level weight distribution.

[0058] In specific application scenarios, when migrating from the construction engineering field to the road engineering field, the system will find that the two fields have many similarities in the underlying terminology (such as concrete strength, steel bar specifications, etc.), but there are differences in the risk logic framework. Therefore, the adaptive weight calculation will reduce the bottom-level weight and increase the top-level and middle-level weights, focusing on the adaptation of the risk framework and logical structure. For example, when mapping "uneven force on the support structure leading to local deformation" to the road field, the system will retain the underlying "uneven force" and "local deformation" terms, but adjust the risk causal logic of the middle layer and map it to "uneven roadbed compaction leading to local deformation of the road surface", accurately capturing the essential similarities and expression differences of the risk patterns of the two fields.

[0059] Optionally, in some embodiments, the loss function can be further extended to a multi-objective optimization form, taking into account the three objectives of semantic similarity, structure preservation, and domain feature retention, and finding the optimal balance between the three through weighted combination or Pareto front optimization method.

[0060] In addition, the semantic distance metric function You can flexibly choose according to the specific application scenario, such as cosine distance, Wasserstein distance, or a customized distance metric function combined with domain knowledge.

[0061] like Figure 5The figure below illustrates the weight distribution of each semantic layer in the proposed method for domain-specific migration. For similar domains, the bottom layer has a higher weight (0.4), focusing more on specific term mappings. For domains with significant differences, the top layer has a higher weight (0.45), focusing more on abstract concept mappings. This demonstrates that the proposed method can dynamically adjust the weights of different semantic layers based on the characteristics of different domains, employing different weight distribution strategies.

[0062] Step 2.4, gradient path optimization; According to the embodiments of this application, a gradient path optimization method is implemented to analyze the gradient flow path between different levels, identify parameter sensitivities, and perform targeted optimization. This method tracks key transition points in the gradient flow process, identifies the subset of parameters that have the greatest impact on adaptability, and prioritizes these parameters to accelerate convergence.

[0063] The gradient path optimization method provided in this application has the following implementation details: The gradient path optimization method consists of a gradient flow analysis unit, a parameter sensitivity evaluation unit, and a selective update unit. The gradient flow analysis unit uses visualization techniques and numerical analysis methods to track the gradient propagation paths and intensity changes in the network. Specifically, a gradient flow graph (GFH) is used to represent the gradient propagation relationship between network layers. Each node represents a network layer or parameter group, and the edge weight indicates the gradient propagation strength.

[0064] The parameter sensitivity evaluation unit calculates the influence of each parameter on the domain adaptation performance using a parameter-level perturbation analysis method. The specific calculation process is as follows: First, the system applies a small perturbation to each parameter to be evaluated to form a perturbed parameter set; then, the loss function values ​​of the original parameters and the perturbated parameters are calculated for each domain adaptation task; then, the loss change caused by the perturbation is calculated and divided by the perturbation size to obtain the approximate sensitivity of the parameter on the specific task; finally, the sensitivity results of all tasks are weighted averaged to obtain the overall sensitivity score of the parameter. During the weighted averaging process, the system assigns weights based on the representativeness and importance of different tasks to ensure that the evaluation results can reflect the degree of influence of the parameters on key tasks. Through this perturbation-based sensitivity analysis method, the system can efficiently identify the key parameters that have the greatest impact on adaptation performance.

[0065] The selective update unit implements a differentiated optimization strategy based on parameter sensitivity: using a larger learning rate and a precise optimization algorithm (such as Adam) for highly sensitive parameters; a standard learning rate and SGD for moderately sensitive parameters; and a smaller learning rate or complete freezing for less sensitive parameters. This differentiated update strategy improves optimization efficiency.

[0066] The system also implements a hierarchical optimization scheduling algorithm, dynamically adjusting resource allocation based on the convergence rate of different semantic layers. For example, in the initial optimization phase, the system prioritizes resources to optimize the mapping parameters of the top-level abstract semantics, establishing a framework correspondence between domains. As the top-level parameters gradually converge, the system gradually shifts the optimization focus to the middle and bottom layers, fine-tuning relationship and term mappings.

[0067] In practical application scenarios, when migrating from the financial risk domain to the insurance risk domain, the gradient path optimization method identifies core parameter groups: the top-level concept mapping matrix and risk association representation parameters, which have the greatest impact on adaptation performance. The system prioritizes these parameters for rapid adaptation. For example, when mapping "asset impairment due to derivative price fluctuations" to the insurance domain, the system was able to find the ideal parameter configuration within a very short number of iterations, accurately mapping it to "insufficient reserves due to fluctuations in the value of insured assets."

[0068] In other embodiments, second-order derivative information can be introduced to construct a Hessian matrix to guide the optimization direction and avoid falling into local optimal solutions. In addition, a meta-learning optimizer can be used to automatically learn the optimal parameter update strategy, further improving optimization efficiency.

[0069] A semantic hierarchical prototype network building module that receives the output of the cross-domain mapping function and maps the complex risk structure; Furthermore, this application provides a method for constructing a semantic hierarchical prototype network, specifically designed to achieve efficient mapping of complex risk structures. It should be noted that this network is an important supplement to the aforementioned meta-learning framework, particularly when dealing with complex, nested risk descriptions in the target domain. Therefore, this module includes the following sub-steps: Step 3.1, multi-level prototype representation construction; According to the embodiment of the present application, a multi-level prototype representation is constructed for the source domain and the target domain respectively. , representative risk patterns are extracted from historical data as prototypes through clustering methods, and each prototype represents a typical risk pattern. The prototype is represented as: ; in Indicates the A set of prototypes at the semantic level; 、 、 Respectively represent The first, second, and prototype vectors; Represents a semantic hierarchical index; Indicates the The number of prototypes for the layer.

[0070] The semantic hierarchical prototype network provided in this application has the following structure and implementation: The Semantic Hierarchical Prototype Network is a multi-layered, dual-branch network structure consisting of a prototype representation layer, a cross-layer association layer, and an adaptive mapping layer. The network's innovation lies in combining prototype learning with hierarchical representation learning to form a knowledge representation system that operates simultaneously at multiple levels of abstraction.

[0071] The prototype representation layer targets each semantic level Maintain a set of prototype vectors Each prototype vector represents a typical semantic pattern at that level. Unlike traditional prototype networks, this network uses a hierarchy-sensitive clustering algorithm during the prototype extraction phase. This algorithm considers the hierarchical position of nodes in the semantic tree, enabling prototypes at different levels to capture semantic patterns at corresponding levels of abstraction. For example, top-level prototypes capture broad risk categories (such as equipment failure risk and safety accident risk), middle-level prototypes capture risk forms (such as control system failure and operational errors), and bottom-level prototypes capture specific risk instances (such as specific failure modes of specific equipment).

[0072] The cross-level association layer establishes relationships between prototypes at different levels. Using a Graph Attention Network (GAN) structure, the prototypes at each level are organized into a directed graph. The edges in the graph represent the strength of the associations between prototypes at different levels, and are automatically learned through end-to-end training. This graph structure enables the network to model hierarchical dependencies between prototypes. For example, a top-level prototype for "equipment failure risk" might have a strong association with a middle-level prototype for "control system failure," but a weaker association with a prototype for "human error."

[0073] The adaptive mapping layer is responsible for mapping the prototypes from the source domain to the target domain. It uses a bidirectional alignment algorithm to optimize the consistency of the mappings from the source domain to the target domain and back to the target domain. This layer consists of a series of nonlinear transformation units, each of which performs a mapping operation on the prototypes at a specific level. To handle domain-specific concepts, the network introduces a residual connection structure, allowing certain features to be directly transferred, avoiding the information loss caused by forced mapping.

[0074] In practical applications, when processing a risk description such as "abnormal robot motion trajectory on an intelligent manufacturing production line may result in insufficient workpiece machining accuracy," the semantic hierarchical prototype network can activate corresponding prototypes at different levels of abstraction: the bottom layer activates prototypes related to entities such as "robot," "motion trajectory," and "workpiece machining"; the middle layer activates prototypes related to relational patterns such as "abnormal motion control" and "reduced precision"; and the top layer activates prototypes related to abstract risk categories such as "device malfunction" and "quality risk." Through this multi-level prototype activation pattern, the network can capture the complete semantic structure of the risk description and find a mapping relationship with the corresponding concepts in the traditional manufacturing field.

[0075] Optionally, in some embodiments, the prototype representation can adopt a dynamic prototype strategy to dynamically adjust the prototype representation based on contextual information to better adapt to specific scenarios. In addition, a hierarchical prototype generation network can be introduced to construct a hierarchical prototype representation from the bottom up by stacking multiple levels of generation modules.

[0076] Step 3.2, hierarchical attention method implementation; According to an embodiment of the present application, a hierarchical attention method is constructed to dynamically focus on relevant prototypes at different semantic levels during the mapping process. For the input risk description, its similarity with the prototypes at each level is calculated, and attention weights are generated based on the similarity distribution. The final mapping output is formed by weighted combination of the prototype mapping results at different levels. The attention weights are calculated using a multi-head attention module, which can focus on multiple relevant prototypes at the same time.

[0077] In addition, this application also provides an enhanced hierarchical attention method, which controls the transmission and fusion of information at different semantic levels by introducing a gated update unit, so that the model can more flexibly adjust the attention allocation strategy according to the input features.

[0078] Step 3.3, prototype update and maintenance method; According to the embodiments of the present application, an online prototype update system is implemented, enabling the prototype set to adapt to new data and domain changes. A prototype freshness evaluation metric is established to regularly assess the representativeness of prototypes, and the prototype set is updated through an incremental clustering method. Newly emerging risk patterns are added as new prototypes if their distance from existing prototypes exceeds a preset threshold. For rarely activated prototypes, their necessity is assessed and considered for removal.

[0079] Optionally, in some implementations, an uncertainty-based active learning strategy can be employed to proactively identify samples requiring expert annotation, effectively improving the efficiency and quality of prototype updates. Furthermore, a memory enhancement module can be introduced to specifically retain historically rare but important prototypes, ensuring the model's ability to handle rare risk types.

[0080] Step 3.4, cross-level relationship modeling; According to the embodiments of this application, a cross-hierarchical relationship model between prototypes is established to capture the dependencies between different abstraction levels. Using a Conditional Random Field (CRF) model, the transition probabilities between upper and lower-level prototypes are modeled, enabling collaborative inference between these hierarchies. This enables the model to handle complex, nested risk descriptions and maintain the integrity of the risk description hierarchy during cross-domain mapping.

[0081] In other embodiments, recursive neural networks or graph neural networks can be used to build cross-level relationship models to capture more complex nonlinear dependencies. In addition, external knowledge graphs can be introduced to assist the model in understanding and reasoning about semantic relationships between levels, further enhancing the model's reasoning capabilities.

[0082] The sample rapid adaptation implementation module performs efficient knowledge transfer based on the output of the semantic hierarchical prototype network construction module; In addition, this application also proposes a method for implementing fast adaptation of a small number of samples, which is used to efficiently transfer knowledge in a new field using a very small number of samples. It should be understood that this module comprehensively utilizes all the components built previously and is the key innovation of this technical solution. Through this module, the system can achieve effective cross-domain adaptation with only 5 to 10 samples. It specifically includes the following sub-steps: Step 4.1, small sample learning strategy; According to the embodiments of this application, based on the previously constructed meta-learning framework and semantic hierarchical prototype network, a small sample learning strategy is implemented. When entering a new domain, only 5 to 10 representative risk description samples need to be collected, converted into a hierarchical semantic representation through a recursive neural parser, and paired with the standard representation of these samples in the source domain to form a small number of mapping examples.

[0083] It should be noted that during the sample selection phase, this application employed a representative sample screening algorithm, using clustering and diversity analysis to ensure that the small number of selected samples covered the main risk types and expressions in the target area. Furthermore, active learning strategies could be incorporated to select the most informative samples from the initial sample set for annotation, further improving the efficiency of small-sample learning.

[0084] like Figure 6 As shown in the figure, a comparison is shown between the number of samples required for new domain adaptation by the method of the present invention and the traditional method. The method of the present invention only requires 10 samples to complete new domain adaptation, while the traditional method requires more than 1,000 samples, reducing the sample requirement by 98%.

[0085] Step 4.2, efficient parameter fine-tuning; According to one embodiment of the present application, a parameter-efficient fine-tuning method is constructed, adjusting only the subset of parameters in the mapping function that are most sensitive to the new domain characteristics, while leaving the majority of pre-trained parameters unchanged. Low-Rank Adapters technology is employed to adapt to specific domain characteristics while maintaining the generalizability of the model by inserting a small number of adaptation layers between the original network layers. This approach reduces the number of parameters that need to be updated, thereby mitigating the risk of overfitting.

[0086] In some implementations, a parameter freezing strategy can optionally be employed. Based on the results of parameter sensitivity analysis, insensitive parameter layers are completely frozen, and only the highly sensitive parameter layers are updated, further improving fine-tuning efficiency. Furthermore, this application proposes a mixed-precision training strategy, which uses different computational precisions for parameters of varying importance, reducing computational resource requirements while maintaining performance.

[0087] Step 4.3, hierarchical progressive fine-tuning; According to the embodiments of the present application, a hierarchical and progressive fine-tuning strategy is implemented, starting from high-level semantics and gradually fine-tuning downward to low-level semantics. First, the mapping parameters of the abstract level are adjusted to establish the corresponding relationship between the risk frameworks of different domains; then the parameters of the middle layer are adjusted to adapt to the domain-specific risk relationship structure; finally, the parameters of the bottom layer are adjusted to process the mapping relationship between specific terms and entities. This top-down fine-tuning method can efficiently adapt to the differences between domains at different levels while maintaining consistency.

[0088] It should be understood that in other embodiments, a bidirectional fine-tuning strategy may be adopted, that is, fine-tuning starting from both the top and bottom layers and converging toward the middle layer. This strategy may be more effective when dealing with domains where semantic frameworks and specific terminology differences are equally important.

[0089] Step 4.4, zero-sample generalization ability is enhanced; According to one embodiment of the present application, a zero-shot generalization enhancement module is constructed, enabling the system to perform inference even when some samples are lacking. Based on known mapping relationships, this module uses transitive reasoning and semantic similarity propagation to predict unseen risk concept mappings. A graph neural network is used to model the relationship network between concepts, and a message passing algorithm is used to propagate mapping information across the graph, enabling mapping inference for unseen concepts.

[0090] Optionally, this application also proposes a hybrid reasoning framework that combines the strengths of symbolic and neural reasoning to enhance the model's interpretability and generalization capabilities. The symbolic reasoning component leverages the conceptual relationships within the risk ontology and domain knowledge base for explicit reasoning, while the neural reasoning component handles fuzzy matching and implicit associations. The combination of the two provides powerful zero-shot generalization capabilities.

[0091] like Figure 7 The figure shows a comparison of the time required to complete new domain adaptation using the inventive method and traditional methods. The inventive method completes new domain adaptation in just 24 hours, while the traditional method takes 336 hours (approximately 2 weeks), increasing adaptation speed by over 90%.

[0092] Standardized risk knowledge base construction and application module, receiving samples and quickly adapting the module output to achieve risk standardization and application; Finally, this application applies the technical achievements of the aforementioned modules to actual risk standardization scenarios, constructing a unified standardized risk knowledge base and implementing it in project management. It is worth noting that this module transforms theoretical innovation into practical application, demonstrating the practical value of this technical solution. Specifically, it includes the following sub-steps: Step 5.1, building a bidirectional mapping relationship library; According to the embodiments of this application, a bidirectional mapping relationship library between the source and target domains is constructed based on the results of semantic hierarchical adaptation learning. This relationship library not only contains conceptual mappings but also structural transformation rules, enabling bidirectional conversion of complex risk descriptions. The relationship library is stored in a graph database, facilitating the representation and efficient querying of complex relationships.

[0093] It should be noted that the bidirectional mapping relationship library of this application adopts a hierarchical triple structure. Unlike traditional flat knowledge graphs, each mapping relationship contains rich metadata such as the hierarchical identifier, relationship type, and credibility score, supporting accurate cross-hierarchical relationship query and reasoning. In addition, this application also provides a mapping consistency verification method, which evaluates the quality of the mapping relationship through a cyclic transformation test (source domain → target domain → source domain) to ensure the consistency of the bidirectional transformation.

[0094] Step 5.2, automatic standardization of risk processing process; According to one embodiment of the present application, a standardized processing pipeline is constructed to achieve automatic standardization of risk descriptions. The pipeline includes text preprocessing, recursive semantic parsing, hierarchical representation extraction, cross-domain mapping application, and result generation and verification. The system automatically determines the domain to which the input risk description belongs and applies corresponding mapping rules to convert it into a standardized format while preserving the original structure and semantic integrity.

[0095] In some implementations, a multi-path processing strategy can optionally be introduced, simultaneously processing the same risk description using multiple different mapping paths. The optimal result is then selected through voting or fusion algorithms, improving the accuracy and robustness of the normalization process. Furthermore, this application provides incremental processing capabilities, enabling efficient processing of newly added risk description fragments without reprocessing the entire document.

[0096] Step 5.3, human-machine collaborative verification and correction system; According to the embodiments of this application, a human-machine collaborative verification and correction system is implemented, integrating expert knowledge to further improve standardization quality. The system calculates a confidence score for each automated standardization result. Low-confidence results are submitted to domain experts for verification and correction. Expert feedback is fed back to the system through active learning methods, continuously improving the mapping model and semantic hierarchical prototype network.

[0097] It should be understood that during the human-machine collaborative verification process, this application provides a hierarchical review strategy, assigning tasks to experts at different levels based on the importance and uncertainty of the results, thereby optimizing the efficiency of human resource utilization. Furthermore, the system possesses self-learning capabilities, capable of analyzing correction patterns from the experts' correction behaviors, gradually reducing the need for human intervention.

[0098] like Figure 8 The figure shows a comparison of the accuracy of the standardized results of the human-machine collaborative method of the present invention and the purely automated method. The accuracy of the standardized results of the human-machine collaborative method of the present invention reached 92%, while the purely automated method was 78%, an improvement of 18%.

[0099] Step 5.4, diversified project management application interface; According to the embodiments of this application, a diversified project management application interface is constructed, enabling the seamless integration of standardized risk knowledge into existing project management systems. The interface includes functional modules such as risk identification, risk classification, risk correlation analysis, and cross-domain risk comparison. Services are provided through APIs or microservices, supporting the invocation and integration of different project management platforms.

[0100] As can be seen, the diversified application interface proposed in this application not only supports standard REST API calls but also provides an event-driven interface that can respond to risk changes in real time within the project. Furthermore, the interface supports both batch and stream processing modes to accommodate different business scenarios.

[0101] A computer storage medium includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the above-mentioned project intelligent management and analysis platform based on security risk standardization.

[0102] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. Project intelligent management and analysis platform based on security risk standardization, characterized by: include: Hierarchical semantic space representation model building module, constructing multi-level semantic representation and outputting hierarchical semantic tree; A meta-learning-driven cross-domain mapping function construction module receives a hierarchical semantic tree and constructs a cross-domain mapping function to achieve cross-domain knowledge transfer; A semantic hierarchical prototype network building module that receives the output of the cross-domain mapping function and maps the complex risk structure; The sample rapid adaptation implementation module performs efficient knowledge transfer based on the output of the semantic hierarchical prototype network construction module; The standardized risk knowledge base construction and application module receives samples, quickly adapts the module output, and realizes risk standardization and application.

2. The project intelligent management and analysis platform based on security risk standardization according to claim 1 is characterized in that: The hierarchical semantic space representation model construction module includes semantic hierarchy definition, recursive neural parser, hierarchical feature extraction algorithm and semantic space pre-training. The recursive neural parser parses risk text into a hierarchical semantic tree structure based on a tree-structured long short-term memory network.

3. The project intelligent management and analysis platform based on security risk standardization according to claim 1 is characterized in that: The meta-learning-driven cross-domain mapping function construction module includes task construction and sampling, model-independent meta-learning algorithm implementation, hierarchical adaptation loss function construction and gradient path optimization. The model-independent meta-learning algorithm is based on the MAML framework and trains the mapping function through a two-layer optimization process.

4. The project intelligent management and analysis platform based on security risk standardization according to claim 3 is characterized in that: The loss function constructed by the hierarchical adaptation loss function is: ； in, Indicates that the parameter is The hierarchical adaptation loss function of the mapping function; represents the summation operator, is the semantic level index, is the total number of semantic levels; Indicates the The weights of the semantic levels are adaptively adjusted according to domain differences; For the Layer source domain semantic representation and target domain semantic representation Semantic distance measurement function between them; and Represents the source domain and target domain in Semantic representation of layers.

5. The project intelligent management and analysis platform based on security risk standardization according to claim 1 is characterized in that: The semantic hierarchical prototype network construction module includes multi-level prototype representation construction, multi-level attention method implementation, prototype update and maintenance, and cross-level relationship modeling. The multi-level prototype representation maintains a set of prototype vectors for each semantic level, and each prototype vector represents a typical semantic pattern.

6. The project intelligent management and analysis platform based on security risk standardization according to claim 5 is characterized in that: The semantic hierarchical prototype network includes a prototype representation layer, a cross-level association layer and an adaptation mapping layer. The cross-level association layer adopts a graph attention network structure to organize prototypes at each level into a directed graph structure, and the edges represent the association strength between prototypes at different levels.

7. The project intelligent management and analysis platform based on security risk standardization according to claim 1 is characterized in that: The sample rapid adaptation implementation module includes small-scale sample learning strategies, efficient parameter fine-tuning, hierarchical progressive fine-tuning and zero-sample generalization capability enhancement, and realizes efficient transfer of new domain knowledge through 5 to 10 risk description samples.

8. The project intelligent management and analysis platform based on security risk standardization according to claim 7 is characterized in that: The efficient parameter fine-tuning adopts low-rank adapter technology, which adapts to the characteristics of specific fields by inserting adaptation layers between the original network layers, with the number not exceeding 10% of the original network layers; The hierarchical progressive fine-tuning method implements a top-down parameter adjustment method that starts from high-level semantics and adjusts parameters layer by layer according to a preset hierarchical order to low-level semantics.

9. The project intelligent management and analysis platform based on security risk standardization according to claim 1 is characterized in that: The standardized risk knowledge base construction and application module includes the construction of a two-way mapping relationship library, an automatic risk standardization processing flow, a human-computer collaborative verification and correction system, and a diversified project management application interface. The two-way mapping relationship library adopts a hierarchical triple structure, and each mapping relationship contains a hierarchical identifier, a relationship type, and a credibility score.

10. A computer storage medium, characterized in that It includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the project intelligent management and analysis platform based on security risk standardization as described in any one of claims 1 to 9.

Images