Project intelligent management and analysis device based on security risk standardization

Abstract

The application relates to the technical field of security risk standardization management, and discloses a project intelligent management and analysis device based on security risk standardization, which comprises the following modules: a hierarchical semantic space representation model construction module, which constructs a multilevel semantic representation and outputs a hierarchical semantic tree; a meta-learning driven cross-domain mapping function construction module, which receives the hierarchical semantic tree and constructs a cross-domain mapping function to realize cross-domain knowledge transfer; a semantic hierarchical prototype network construction module, which receives the output of the cross-domain mapping function and maps a complex risk structure; a sample rapid adaptation implementation module, which realizes efficient knowledge transfer of the output of the construction module; a standardized risk knowledge base construction and application module, which receives the output of the sample rapid adaptation implementation module and realizes risk standardization and application; the application reduces the sample quantity and sample demand required for new field adaptation by combining meta-learning and a semantic hierarchical prototype network, and shortens the construction period of new field risk standardization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of safety risk standardization management, more particularly, it relates to a project intelligent management and analysis device based on safety risk standardization. BACKGROUND

[0002] With the deepening development of global economic integration and business diversification, more and more enterprises begin to involve in multiple different business fields, and cross-field safety risk management has become a key challenge for the sustainable development of enterprises. Under this background, it is increasingly important to establish a unified and adaptable safety risk standardization system.

[0003] Currently, safety risk standardization mainly relies on rule-based expert systems and traditional machine learning methods. Although rule-based methods perform well in specific fields, their rules are usually manually developed by domain experts and have strong domain specificity, making it difficult to directly apply to other fields. When an enterprise needs to expand into new business fields, it often needs to rebuild the entire rule system, resulting in high cost and long cycle.

[0004] Traditional machine learning methods also have obvious shortcomings in risk standardization. These methods usually use supervised learning methods and require a large amount of labeled data for training. According to existing research, traditional transfer learning methods usually require thousands or even tens of thousands of labeled samples to achieve acceptable performance levels when adapting to new fields. This high data requirement makes enterprises face huge data collection and labeling costs when rapidly expanding new businesses.

[0005] In the processing of complex risk structures, existing technologies generally have limitations. Most methods treat risk description texts as flat sequences, using bag-of-words models or simple sequence encoding methods, ignoring the rich hierarchical structure information in risk descriptions.

[0006] Domain adaptability is another key technical bottleneck. Existing cross-domain methods often face a dilemma: either completely follow the source domain standard for forced mapping, losing the characteristic expression methods of the target domain; or completely rebuild the target domain standard, making it difficult to be compatible with existing enterprise risk management systems. This technical limitation makes it difficult for enterprises to form a unified risk management system during business expansion.

[0007] In terms of semantic understanding and generalization ability, existing technologies mainly rely on text surface features 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 belong to the risk of device unavailability, but existing methods often cannot identify this deep semantic association, resulting in poor cross-domain knowledge transfer.

[0008] In addition, the prior art also has deficiencies in parameter efficiency and computational resource consumption. Whenever a new field needs to be adapted to, the traditional method usually needs to retrain a large number of parameters, which not only takes a long time to train, but also requires high computational resources. This high resource consumption feature makes it difficult for the prior art to be applied in resource-constrained environments, and is also not conducive to the rapid business expansion of enterprises.

[0009] In summary, the existing security risk standardization technology has deficiencies in cross-field adaptability, complex structure processing, semantic understanding generalization, and resource efficiency, and cannot meet the actual needs of modern enterprise diversification and rapid business expansion. Therefore, a new technical solution is urgently needed to solve these problems. SUMMARY

[0010] The application provides a project intelligent management analysis device based on security risk standardization, which solves the technical problems of poor cross-field risk standardization adaptability, high data demand, and weak complex structure processing capability in related technologies.

[0011] The application provides a project intelligent management analysis device based on security risk standardization, which includes:

[0012] A hierarchical semantic space representation model construction module constructs a multi-level semantic representation and outputs a hierarchical semantic tree;

[0013] A meta-learning driven cross-field mapping function construction module receives the hierarchical semantic tree and constructs a cross-field mapping function to realize cross-field knowledge transfer;

[0014] A semantic hierarchical prototype network construction module receives the output of the cross-field mapping function and maps complex risk structures;

[0015] A sample rapid adaptation implementation module performs efficient knowledge transfer based on the output of the semantic hierarchical prototype network construction module;

[0016] A standardized risk knowledge base construction and application module receives the output of the sample rapid adaptation implementation module and realizes risk standardization and application.

[0017] Further, 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 is based on a tree structure long short-term memory network, and parses risk text into a hierarchical semantic tree structure.

[0018] Further, the meta-learning driven cross-field 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 double-layer optimization process.

[0019] Further, the loss function constructed by the hierarchical adaptation loss function is:

[0020]

[0021] wherein, represents the hierarchical adaptation loss function of the mapping function with parameter θ2; ∑ represents the summation operator, l1 is the index of the semantic hierarchy, and L2 is the total number of semantic hierarchies; represents the weight of the l1th semantic hierarchy, which is adaptively adjusted according to the domain difference; is the semantic representation of the l1th layer in the source domain; is the semantic representation of the target domain; is the semantic distance measurement function between the semantic representation of the l1th layer in the source domain and the semantic representation of the target domain; and respectively represent the semantic representation of the l1th layer in the source domain and the target domain.

[0022] Further, the semantic hierarchy prototype network construction module includes multi-level prototype representation construction, hierarchical attention method implementation, prototype updating and maintenance, and cross-level relationship modeling. The multi-level prototype representation maintains a set of prototype vectors for each semantic hierarchy, and each prototype vector represents a typical semantic pattern.

[0023] Further, the semantic hierarchy prototype network includes a prototype representation layer, a cross-level association layer, and an adaptation mapping layer. The cross-level association layer uses 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 different level prototypes.

[0024] Further, the sample rapid adaptation implementation module includes a small-scale sample learning strategy, a parameter efficient fine-tuning, a hierarchical progressive fine-tuning, and a zero-shot generalization ability enhancement. Through 5 to 10 risk description samples, the efficient migration of new domain knowledge is realized.

[0025] Further, the parameter efficient fine-tuning uses a low-rank adapter technology to realize the adaptation of specific domain features by inserting no more than 10% of the original network layers.

[0026] The hierarchical progressive fine-tuning realizes a top-down parameter adjustment method that starts from high-level semantics and adjusts parameters layer by layer to low-level semantics according to a preset level order.

[0027] Further, the standardized risk knowledge base construction and application module includes a bidirectional mapping relationship library construction, a risk automatic standardization processing flow, a man-machine collaborative verification and correction system, and a diversified project management application interface. The bidirectional mapping relationship library uses a hierarchical triple structure, and each mapping relationship contains a hierarchy identifier, a relationship type, and a credibility score.

[0028] The application provides a computer storage medium, comprising a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the project intelligent management analysis device based on security risk standardization.

[0029] The application has the beneficial effects that: through the combination of meta-learning and the semantic hierarchical prototype network, the number of samples required for new field adaptation is reduced from thousands in the traditional method to 5 to 10, the sample demand is reduced, and the new field risk standardization construction period is shortened;

[0030] Based on the recursive neural parser and the hierarchical semantic representation, the system can accurately process complex risk descriptions with multiple levels of nesting, and compared with the flat text processing method, the accuracy of complex risk structure is improved;

[0031] Through the hierarchical adaptation strategy, the risk standardization is realized while the field characteristics are retained, which ensures the comparability of cross-field risks and retains part of the field-specific representation method, providing support for diversified management;

[0032] Based on the semantic hierarchical prototype network and the zero-shot generalization module, the system can accurately identify similar risks expressed by different terminologies, and the mapping accuracy of unobserved risk concepts is improved;

[0033] Through efficient parameter fine-tuning and gradient path optimization, the system reduces the number of parameters required for adapting to new fields, reduces the demand for computing resources, and supports running in resource-limited environments;

[0034] Combining meta-learning and hierarchical progressive fine-tuning, the new field adaptation time is shortened, and the adaptation speed is improved;

[0035] Through the man-machine collaborative verification and correction system, the overall accuracy of the standardization result is improved;

[0036] Based on the hierarchical semantic representation and the semantic hierarchical prototype network, the system can provide a clear risk mapping path and basis, and enhance the transparency and explainability of the standardization process. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 is a flowchart of a project intelligent management analysis device based on security risk standardization in the application;

[0038] Figure 2 is a column chart comparing the accuracy of the application method and the flat text processing method when processing complex risk structures;

[0039] Figure 3 is a column chart comparing the unobserved risk concept mapping accuracy of the application method and the traditional method;

[0040] Figure 4 is a bar chart of the adaptation effect of the method of the present application in the migration task among five different fields;

[0041] Figure 5 is a column chart of the weight distribution of each level of the present application in the migration of different fields;

[0042] Figure 6 is a column chart of the comparison of the number of samples required for the adaptation of the method of the present application and the traditional method in a new field;

[0043] Figure 7 is a column chart of the comparison of the time for completing the adaptation of the method of the present application and the traditional method in a new field;

[0044] Figure 8 is a column chart of the comparison of the accuracy rate of the standardized results of the human-computer collaborative method of the present application and the pure automation method. DETAILED DESCRIPTION

[0045] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that discussions of these implementations are merely provided to enable those skilled in the art to better understand and utilize the subject matter described herein, and do not limit the scope of protection of the present specification. Various processes and components can be omitted, substituted, or added as desired. Additionally, features described with respect to one example can be combined in other examples.

[0046] In at least one embodiment of the present application, a project intelligent management analysis device based on security risk standardization is disclosed, as shown in Figure 1 The device comprises:

[0047] A hierarchical semantic space representation model construction module constructs a multi-level semantic representation and outputs a hierarchical semantic tree.

[0048] The method provided by the present application first constructs a hierarchical semantic space representation model, providing a structured semantic representation basis for subsequent cross-field mapping. Specifically, the following sub-steps are included:

[0049] Step 1.1, semantic level definition;

[0050] According to an embodiment of the present application, a multi-level semantic space structure is defined:

[0051]

[0052] where H represents the hierarchical semantic space; h1 represents the most abstract semantic level, mainly containing general risk concepts independent of the field; h2 represents a semantic level one level more specific than h1; represents the most specific semantic layer, containing field-specific detailed risk description; the intermediate layers h2 to represent the transition levels from abstract to specific in turn, each level carrying more refined semantic information, realizing the layer-by-layer refinement and decomposition of semantics; L1 represents the total number of semantic levels, which can be flexibly set according to the actual application scenario.

[0053] The semantic representation of each layer is stored as a feature vector, and the dimension increases with the depth of the level to accommodate more detailed information.

[0054] It should be noted that in some embodiments, the number of semantic levels L1 can be adjusted according to the complexity of the application scenario. For example, for a simple structure risk description, a 3-layer structure (abstract layer, intermediate layer, specific layer) can be used; while for a complex structure risk description, a 5-layer or more level structure can be used to model the semantic level relationship more finely. In addition, the vector dimension of different levels can also be flexibly configured to balance the expression ability and computational efficiency.

[0055] Alternatively, in another embodiment of the present application, the semantic level definition can be combined with domain expert knowledge to design a special level structure for a specific industry field. For example, in the field of financial risk, a five-layer structure including "risk category-risk subclass-risk factor-risk event-risk impact" can be defined; while in the field of manufacturing industry, a five-layer structure including "production link-equipment system-failure type-failure performance-failure consequence" can be defined. Through customized level definition, the adaptability of the model to specific fields is further improved.

[0056] Step 1.2, recursive neural parser construction;

[0057] According to one embodiment of the present application, a recursive neural parser based on tree structure long short-term memory network (Tree-LSTM, Tree-Long Short-Term Memory) is used to parse the risk text into hierarchical semantic tree. The parser receives the original risk description text as input, and outputs a multi-level semantic tree structure through three stages of morphological analysis, syntactic analysis and semantic analysis. The parser pays special attention to the cause-effect relationship, conditional relationship and modification relationship in the risk description, and decomposes the complex nested risk description into a clear hierarchical structure.

[0058] The tree structure long short-term memory network recursive neural parser provided by the present application has the following structure and implementation mode:

[0059] The Tree-LSTM recursive neural parser consists of an input layer, a tree structure encoding layer, an attention layer, and an output layer. The input layer first tokenizes the text and performs word embedding processing, mapping each word to a dense vector representation. The tree structure encoding layer is the core of the model, which uses an improved Child-Sum Tree-LSTM unit to explicitly model the dependency relationships in the syntax tree. Each Tree-LSTM unit contains an input gate, an output gate, a forget gate, a memory cell, and a hidden state. Unlike traditional LSTM, Tree-LSTM is equipped with a separate forget gate for each child node, allowing the model to more accurately control the retention and discard of information from different child nodes.

[0060] In the specific scenario of risk text processing, the recursive neural parser first applies a pre-trained dependency syntax analyzer to generate an initial syntax tree, and then combines professional domain dictionaries and grammar rules to prune and recombine the tree, enhancing the ability to recognize risk terms and special sentence patterns. Further, the model introduces a hierarchical attention method, calculating attention weights in the vertical direction (depth of the tree) and horizontal direction (nodes at the same level), allowing the model to focus on both local details and global structure. At the bottom layer of the tree (leaf nodes), specific entities and terms are processed, intermediate layers process condition and attribute information, and the top layer processes abstract concepts such as risk categories and risk levels.

[0061] To enhance the model's ability to handle cross-domain risk descriptions, the parser also integrates a semantic role labeling module specifically designed for risk expressions, which can identify key components such as event triggers, risk objects, consequences, conditions, and responsible parties in risk descriptions, and encode this information into the corresponding nodes of the tree. In the processing of complex nested risk descriptions, the model uses a recursive tree synthesis strategy, first processing smaller semantic units and then gradually combining them into larger semantic structures, effectively addressing the difficulties of traditional sequence models in handling long-distance dependencies.

[0062] Through the above structure and implementation, the Tree-LSTM recursive neural parser can accurately extract hierarchical risk knowledge from complex risk descriptions such as "When the power supply of production line 5 in the workshop fails, if the backup power supply fails to start and the operator fails to switch to manual mode in time, it may cause the cooling system of high-temperature equipment to stop working, resulting in equipment damage and production interruption", including event conditions (power failure, 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.

[0063] It should be understood that in other embodiments, other types of recurrent neural network structures, such as a recursive neural tensor network (RNTN) or a graph convolutional network (GCN), can be employed to construct the recurrent neural parser as long as they can effectively parse the risk text into a hierarchical semantic tree structure.

[0064] Step 1.3, semantic representation extraction;

[0065] According to an embodiment of the present application, a hierarchical feature extraction algorithm is applied to the parsed semantic tree to extract corresponding semantic representations from 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. All levels of features are integrated through an attention weighting algorithm to form a complete multi-level semantic representation.

[0066] In some embodiments, a feature fusion strategy can be employed to combine features from different levels through residual connections or gating units to enhance the completeness and robustness of the representation. In addition, an external knowledge base (such as a domain risk ontology or a professional dictionary) can be introduced to assist feature extraction, improving the accuracy and richness of the semantic representation.

[0067] As shown in Figure 2 The accuracy of the method of the present application and the flat text processing method in processing complex risk structures is compared, and the accuracy of the method of the present application in processing complex risk structures reaches 81%, while the accuracy of the flat text processing method is only 45%, which is improved by 80%.

[0068] Step 1.4, semantic space pre-training;

[0069] According to an embodiment of the present application, a large-scale cross-domain risk corpus is used to pre-train the semantic space representation model, so that the model can capture general risk semantic features. A contrastive learning method is used to enable the model to learn the ability to distinguish similar risks and unrelated risks, and at the same time, a mask language modeling task is used to enhance the semantic understanding ability of the model.

[0070] Optionally, the pre-training stage can employ an adversarial training strategy to enhance the robustness of the model through adversarial samples, so that it can handle minor changes in semantic expression. In addition, a multi-task learning framework can be combined to perform multiple pre-training tasks (such as context prediction, relationship prediction, etc.) at the same time, further enhancing the semantic understanding ability of the model.

[0071] The meta-learning driven cross-domain mapping function construction module receives the hierarchical semantic tree and constructs a cross-domain mapping function to realize cross-domain knowledge transfer;

[0072] The application further provides a meta-learning driven cross-domain mapping function construction method for realizing knowledge transfer capable of quickly adapting to new domains. It should be understood that this module adopts a meta-learning strategy to train the model to quickly adapt to new tasks under the condition of few samples. In addition, this module is closely combined with the previous module, and the hierarchical semantic representation is taken as the input and output of the mapping function. Specifically, the following sub-steps are included:

[0073] Step 2.1, task construction and sampling;

[0074] According to an embodiment of the application, the cross-domain risk standardization problem is formalized as a set of meta-learning tasks. Each task contains a support set and a query set, the support set contains a small number (usually 5 to 10) of mapping samples from the source domain to the target domain, and 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 data set.

[0075] It should be noted that in the task construction process, a hierarchical sampling strategy can be adopted to ensure that the constructed tasks can cover different types of inter-domain transfer scenarios (such as similar inter-domain transfer and large difference inter-domain transfer). In addition, a difficult sample mining method can be introduced to preferentially select samples that are difficult to map to construct tasks, so as to enhance the generalization ability of the model.

[0076] Step 2.2, model-independent meta-learning algorithm implementation;

[0077] According to an embodiment of the application, a meta-learning framework is constructed based on the MAML (Model-Agnostic Meta-Learning) algorithm. The framework trains the mapping function through a double-layer optimization process: the inner-layer optimization quickly adapts the parameters to a specific task, and the outer-layer optimization adjusts the initial parameters of the model to make it fastest to adapt to new tasks. The mapping function Receives the hierarchical semantic representation of the source domain As input, output the corresponding representation of the target domain

[0078] Wherein represents the mapping function with parameters θ1; θ1 represents the parameters of the mapping function; represents the hierarchical semantic representation of the source domain; represents the hierarchical semantic representation of the target domain.

[0079] The cross-domain mapping function based on MAML provided by the application has the following implementation details:

[0080] Mapping function The transformer is composed of a multi-layer neural network, including 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 intra-layer and inter-layer dependencies; the mapping transformation layer is composed of multiple residual blocks, each of which includes a domain adaptive 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.

[0081] In the meta-training phase, for each task The algorithm performs the following steps:

[0082] Sampling the support set:

[0083]

[0084] wherein represents the support set of the ith task; represents the hierarchical semantic representation of the jth source domain sample; represents the hierarchical semantic representation of the jth target domain sample; n s1 represents the number of samples in the support set, usually 5-10; j1 represents the support set sample index;

[0085] Based on the support set and the current mapping function parameters, the task-specific parameter update is calculated by gradient descent: first, the system calculates the loss function value of the current mapping function on the specific task 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, which indicates the direction of change of the loss function in the parameter space; then, the system moves the parameters a small step in the opposite direction of the gradient, and the step size is controlled by the inner learning rate, which realizes the adjustment of the parameters in the direction of reducing the task-specific loss function; finally, the parameter values optimized for the specific task are obtained, which are more suitable for processing the domain mapping requirements of the current task;

[0086] Sampling the query set to evaluate the performance of the updated parameters:

[0087]

[0088] wherein represents the query set of the ith task; q1 represents the number of samples in the query set; n s1 +1 represents the starting value of the query set sample index; n s1 +q1 represents the end value of the query set sample index; represents the hierarchical semantic representation of the jth source domain sample; represents the hierarchical semantic representation of the jth target domain sample.

[0089] 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 respectively, which reflects 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 initial parameters of the model; then the gradient of the 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, and the step is controlled by the outer learning rate, completing one meta-learning update iteration; this double-layer optimization design enables the model to learn a good initial parameter point, so that it can quickly adapt to new tasks with a small number of samples

[0090] The present application makes two key improvements to the standard MAML algorithm:

[0091] A task similarity perception module is introduced to set dynamic weights for different tasks, so that the model pays more attention to source tasks similar to the target domain;

[0092] A hierarchical optimization strategy is adopted, and different learning rates are used for parameters at different semantic levels to adapt to the migration difficulty of each level more flexibly.

[0093] In a specific application scenario, when migrating from the petrochemical industry to the natural gas processing industry, the algorithm can identify the commonalities and differences between the two fields in the risk classification system and the expression of terms. For example, by learning from a small number of mapping samples that "valve leakage leading to the accumulation of flammable gas" and "pressure vessel sealing failure causing gas to escape" have corresponding relationships in the semantic structure, the algorithm can be generalized to unseen samples, such as correctly mapping "oil tank floating roof sealing ring aging leading to oil and gas volatilization" to "gas storage tank gasket deterioration causing gas loss" in the natural gas field.

[0094] In other embodiments, other meta-learning algorithms such as Prototypical Networks or Relation Networks can be used, and the most suitable method can be selected according to the characteristics of the specific application scenario. In addition, these algorithms can be combined with traditional domain adaptation methods (such as domain adversarial training) to further enhance the cross-domain migration ability of the model.

[0095] As shown in Figure 3 , the accuracy of the present method and the traditional method in processing unseen risk concept mapping is compared. The accuracy of the present method for unseen risk concept mapping reaches 65%, while the traditional method is only 25%, which reflects the generalization ability of the present method;

[0096] As shown in Figure 4As shown, the adaptation effect of the method of the application in the migration task among five different fields is shown. As can be seen from the figure, the method of the application shows high adaptation effect in the migration task among different fields such as manufacturing to intelligent manufacturing, petrochemical to natural gas processing, construction engineering to road engineering, financial risk to insurance risk, and medical equipment to biopharmaceutical, and the adaptation accuracy is between 83% and 94%.

[0097] Step 2.3, construction of hierarchical adaptation loss function;

[0098] According to the embodiments of the present application, the hierarchical perception adaptation loss function is constructed:

[0099]

[0100] wherein, represents the hierarchical adaptation loss function of the mapping function with parameter θ2; ∑ represents the summation operator, l1 is the index of the semantic hierarchy, and L2 is the total number of semantic hierarchies; represents the weight of the l1th semantic hierarchy, which can be adaptively adjusted according to the field difference; is the semantic representation of the l1th layer of the source field and the semantic representation of the target field between the semantic distance measurement function; and respectively represent the semantic representation of the source field and the target field at the l1th layer.

[0101] The loss function realizes the comprehensive adaptation optimization of different semantic hierarchies by weighted summation of the semantic difference of each hierarchy, thereby improving the accuracy and robustness of cross-field knowledge migration.

[0102] The distance measurement function d adopts different calculation methods according to the different characteristics of the semantic hierarchy, and the specific implementation is as follows:

[0103] For the most abstract top-level semantics, the calculation method based on Wasserstein distance is adopted, i.e. when l1=1:

[0104]

[0105] wherein represents the distance measurement of the top-level semantics; represents the most abstract top-level semantic representation of the source field; represents the most abstract top-level semantic representation of the target field; W2 represents the 2-Wasserstein distance function; represents the probability distribution of the source field; represents the probability distribution of the target field;

[0106] For the intermediate semantic level, a structure-sensitive graph distance metric is adopted, i.e., when 1 < l1 < L2, the form is:

[0107]

[0108] wherein denotes the distance metric of the intermediate level semantics; denotes the semantic representation of the source field l1 layer; denotes the semantic representation of the target field l1 layer; denotes the adjacency matrix of the source field; denotes the adjacency matrix of the target field; F denotes the Frobenius norm; λ1 denotes the balance coefficient; denotes the feature matrix of the source field; denotes the feature matrix of the target field; ||·||2 denotes the L2 norm;

[0109] For the most specific bottom-level semantics, a context-enhanced cosine distance is adopted, i.e., when l1 = L2, denoted as:

[0110]

[0111] wherein denotes the distance metric of the bottom-level semantics; denotes the most specific bottom-level semantic representation of the source field; denotes the most specific bottom-level semantic representation of the target field; cos denotes the cosine similarity function; γ denotes the context similarity adjustment factor function; denotes the context information of the source field; denotes the context information of the target field.

[0112] By adaptively adjusting the weights of different levels The model can dynamically focus on different semantic levels according to the characteristics of the field differences. For example, for fields with large term differences but similar risk logic, increase the bottom-level weight; for fields with similar basic concepts but large risk framework differences, increase the top-level weight.

[0113] The hierarchical adaptive loss function provided in the application has the following implementation details:

[0114] The hierarchical adaptation loss function is a self-adaptive composite loss function specially constructed for the characteristics of semantic hierarchy. The distance metric function d adopts a weighted combination form, which integrates multiple measurement methods: for the top-level abstract semantics, the Wasserstein distance is mainly used to measure the difference of risk category distribution; for the middle-level relationship semantics, the graph structure similarity measurement method such as GraphMatchingLoss is used to evaluate the structural difference of the risk relationship network; for the bottom-level specific semantics, the cosine similarity and context-related embedding distance are combined to accurately capture the difference between term and entity representations.

[0115] The adaptive adjustment of hierarchical weights is realized by attention operation, and the specific calculation process is as follows: first, the system calculates an importance score for each semantic level, which reflects the importance of the current level in domain adaptation; then, the importance scores of all levels are exponentially processed to convert the scores to positive values and amplify the differences; next, the sum of the exponential scores of all levels is calculated as a normalization factor; finally, the weight value of each level is obtained by dividing the exponential score of each level by the normalization factor, and the sum of all level weights is 1. This soft distribution method enables the system to dynamically adjust the attention to different semantic levels, rather than using fixed weights or hard selection mechanism

[0116] The importance score of each level is calculated by a dedicated importance evaluation network, which considers the semantic representation of the source and target domains at the current level and the global state information of domain adaptation

[0117] The evaluation network receives the semantic representation of the source and target domains at a specific level and the global state of the current domain adaptation as input, and outputs the importance score of the level. The specific implementation adopts a multi-layer perceptron network structure, and the processing flow is as follows: first, the system compares the semantic representation of the source and target domains in multiple ways, including direct difference calculation, element-level product operation, and concatenation operation, to construct a comprehensive difference feature vector; then, these difference features are concatenated with global state information (such as domain similarity, adaptation difficulty, and current adaptation progress) to form a complete input feature; next, these features are nonlinearly transformed by the first layer of fully connected neural network and the ReLU activation function is applied to extract key patterns; finally, after further processing by the second layer of neural network, a scalar value is output as the importance score of the current level. The introduction of global state information enables the system to dynamically adjust the importance of each level according to the overall adaptation progress, realizing more flexible and adaptive hierarchical weight distribution.

[0118] In specific application scenarios, when migrating from the field of construction engineering to the field of road engineering, the system finds that the two fields have many common points in the underlying terms (such as concrete strength, steel bar specifications, etc.), but there are differences in the risk logic framework. Therefore, adaptive weight calculation reduces the underlying weight, increases the top and middle layer weight, and focuses on the adaptation of the risk framework and logical structure. For example, when mapping "support structure uneven stress leads to local deformation" to the road field, the system will retain the underlying "uneven stress" and "local deformation" terms, but adjust the middle layer risk causal logic to map it to "uneven compaction of subgrade leads to local deformation of pavement", accurately capturing the essential similarity and expression difference of risk patterns in the two fields.

[0119] Optionally, in some embodiments, the loss function can be further extended to a multi-objective optimization form, considering the three objectives of semantic similarity, structure preservation and domain feature retention, and finding the best balance point between the three through weighted combination or Pareto front optimization method.

[0120] In addition, the semantic distance metric function d can be flexibly selected according to specific application scenarios, such as cosine distance, Wasserstein distance or customized distance metric function combined with domain knowledge.

[0121] As shown in Figure 5 The weight distribution of each semantic level in the migration of different fields is shown. For similar fields, the underlying weight is higher (0.4), and more attention is paid to specific term mapping; and for fields with large differences, the top layer weight is higher (0.45), and more attention is paid to abstract concept mapping. This reflects that the method can dynamically adjust the weight of different semantic levels according to the characteristics of field differences, and adopt different weight allocation strategies.

[0122] Step 2.4, gradient path optimization;

[0123] According to the embodiments of the present application, the gradient path optimization method is implemented, the flow path of the gradient between different levels is analyzed, the parameter sensitivity is identified, and targeted optimization is performed. This method identifies the parameter subset that has the greatest impact on the adaptation ability by tracking the key conversion points in the gradient flow process, and optimizes these parameters first to accelerate convergence.

[0124] The gradient path optimization method provided by the present application has the following implementation details:

[0125] The gradient path optimization method is composed of a gradient flow analysis unit, a parameter sensitivity evaluation unit, and a selective update unit. The gradient flow analysis unit tracks the propagation path and intensity change of the gradient in the network through visualization techniques and numerical analysis methods. Specifically, a gradient flow graph (GFH) is used to represent the propagation relationship of the gradient between network layers, with each node representing a network layer or parameter group and the weight of the edge representing the gradient propagation intensity.

[0126] 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 on each domain adaptation task are calculated for the original parameters and the perturbed parameters respectively; next, 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 and averaged to obtain the overall sensitivity score of the parameter. In the weighted averaging process, the system assigns weights according to the representativeness and importance of different tasks to ensure that the evaluation results reflect the influence of the parameter 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.

[0127] The selective update unit implements differentiated optimization strategies based on parameter sensitivity: high-sensitivity parameters use a larger learning rate and an accurate optimization algorithm (such as Adam); medium-sensitivity parameters use a standard learning rate and SGD; and low-sensitivity parameters use a smaller learning rate or are completely frozen. This differentiated update strategy improves optimization efficiency.

[0128] In addition, the system also implements a hierarchical optimization scheduling algorithm that dynamically adjusts optimization resource allocation based on the convergence speed of different semantic levels. For example, in the early stages of optimization, the system will prioritize allocating resources to optimize the mapping parameters of the top-level abstract semantics to establish the framework correspondence between domains; as the top-level parameters gradually converge, the system will gradually shift the focus of optimization to the middle and bottom layers to fine-tune the relationship mapping and term mapping.

[0129] In actual application scenarios, when migrating from the financial risk domain to the insurance risk domain, the gradient path optimization method can identify the core parameter groups: the top-level concept mapping matrix and the risk correlation representation parameters have the greatest impact on adaptation performance. The system will prioritize optimizing these parameters to achieve rapid adaptation. For example, when processing the mapping task of "asset impairment due to derivative price fluctuations" to the insurance domain, the system can find the ideal parameter configuration in a small number of iterations and accurately map it to "insufficient reserve due to fluctuations in the value of underwriting assets".

[0130] In other embodiments, second-order derivative information can be introduced to construct a Hessian matrix to guide the optimization direction and avoid getting trapped in local optima. Furthermore, a meta-optimizer can be employed to automatically learn the optimal parameter update strategy, further improving optimization efficiency.

[0131] A semantic hierarchical prototype network building module receives the output of a cross-domain mapping function and maps complex risk structures.

[0132] Furthermore, this application provides a semantic-level prototype network construction method, specifically designed for efficient mapping of complex risk structures. It should be noted that this network is an important supplement to the aforementioned meta-learning framework, particularly offering significant advantages in handling complex nested risk descriptions in the target domain. Therefore, this module includes the following sub-steps:

[0133] Step 3.1, constructing a multi-level prototype representation;

[0134] According to embodiments of this application, multi-level prototype representations are constructed for both the source and target domains. For each semantic level... Representative risk patterns are extracted from historical data using clustering methods as prototypes. Each prototype represents a typical risk pattern. The prototypes are represented as follows:

[0135]

[0136] in This represents the set of prototypes for the l2nd semantic level. They represent the 1st, 2nd, and 3rd elements of the 12th layer, respectively. One prototype vector; l2 represents the semantic hierarchy index; This indicates the number of prototypes in the l2 layer.

[0137] The semantic hierarchical prototype network provided in this application has the following structure and implementation:

[0138] The Semantic Hierarchical Prototype Network is a multi-layered, dual-branch network structure consisting of three parts: a prototype representation layer, a cross-layer association layer, and an adaptation mapping layer. Its innovation lies in combining prototype learning with hierarchical representation learning, forming a knowledge representation system capable of operating simultaneously at multiple levels of abstraction.

[0139] The prototype representation layer is for each semantic level Maintain a set of prototype vectors Each prototype vector represents a typical semantic pattern at this level. Unlike traditional prototype networks, this network uses a hierarchical-sensitive clustering algorithm in the prototype extraction stage. This algorithm takes into account the hierarchical position information of the nodes in the semantic tree, allowing 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, safety accident risk, etc.), middle-level prototypes capture risk forms (such as control system failure, operational error, etc.), and bottom-level prototypes capture specific risk instances (such as specific failure modes of specific equipment).

[0140] The cross-level association layer establishes the association between prototypes at different levels. It uses a graph attention network (GAN) structure to organize prototypes at different levels into a directed graph structure. The edges in the graph represent the association strength between prototypes at different levels, which is automatically learned through end-to-end training. This graph structure allows the network to model the hierarchical dependency between prototypes. For example, the "equipment failure risk" prototype at the top level may have a strong association with the "control system failure" prototype at the middle level, but a weak association with the "human operational error" prototype.

[0141] The adaptation mapping layer is responsible for mapping prototypes from the source domain to the target domain. It uses a bidirectional alignment algorithm to simultaneously optimize the mapping consistency from the source domain to the target domain and from the target domain back to the source domain. This layer consists of a series of nonlinear transformation units, each of which performs mapping operations for prototypes at a specific level. To handle domain-specific concepts, the network introduces a residual connection structure, allowing certain features to be directly passed through, avoiding information loss caused by forced mapping.

[0142] In practical application scenarios, when dealing with a risk description such as "the abnormal movement trajectory of the robot on the intelligent manufacturing production line may cause insufficient workpiece processing precision", the semantic hierarchical prototype network can activate corresponding prototypes at different levels of abstraction: bottom-level prototypes related to entities such as "robot", "movement trajectory", and "workpiece processing"; middle-level prototypes related to relationship patterns such as "motion control anomaly" and "precision reduction"; and top-level prototypes related to abstract risk categories such as "device function anomaly" and "quality risk". Through this multi-level prototype activation mode, the network can capture the complete semantic structure of the risk description and find the mapping relationship with the corresponding concepts in the traditional manufacturing domain.

[0143] Optionally, in some embodiments, the prototype representation can adopt a dynamic prototype strategy, dynamically adjusting the prototype representation according to context information to better adapt to specific scenarios. In addition, a hierarchical prototype generation network can be introduced, which constructs hierarchical prototype representations from bottom to top by stacking multiple levels of generation modules.

[0144] Step 3.2, hierarchical attention method implementation;

[0145] According to an embodiment of the present application, a hierarchical attention method is constructed to dynamically focus on relevant prototypes of different semantic levels in the mapping process. For the input risk description, the similarity between it and the prototypes of each level is calculated, and the attention weight is generated based on the similarity distribution. By weighted combination of the prototype mapping results of different levels, the final mapping output is formed. The calculation of the attention weight uses a multi-head attention module, which can simultaneously focus on multiple related prototypes.

[0146] In addition, the present application also provides an enhanced hierarchical attention method, which introduces a gating update unit to control the transmission and fusion mode of information of different semantic levels, so that the model can more flexibly adjust the attention allocation strategy according to the input features.

[0147] Step 3.3, prototype updating and maintenance method;

[0148] According to an embodiment of the present application, an online prototype updating system is implemented to make the prototype set adapt to new data and field changes. A prototype freshness evaluation index is established to regularly evaluate the representativeness of the prototype, and the prototype set is updated by an incremental clustering method. For the newly emerging risk patterns, if the distance from the existing prototypes exceeds the preset threshold, they are added as new prototypes; for the rarely activated prototypes, their necessity is evaluated and removal is considered.

[0149] Optionally, in some embodiments, an uncertainty-based active learning strategy can be used to actively identify samples that need expert annotation, effectively improving the efficiency and quality of prototype updating. In addition, a memory enhancement module can be introduced to specially preserve historical rare but important prototypes, ensuring the model's processing ability for rare risk types.

[0150] Step 3.4, cross-level relationship modeling;

[0151] According to an embodiment of the present application, a cross-level relationship model between prototypes is established to capture the dependency relationship between different abstraction levels. By using a conditional random field (CRF) model, the transition probability between upper and lower level prototypes is modeled to realize collaborative inference between levels. This enables the model to handle complex nested risk descriptions while maintaining the integrity of the hierarchical structure of the risk description when performing cross-domain mapping.

[0152] In other embodiments, a recurrent neural network or a graph neural network can be used to construct a cross-level relationship model to capture more complex nonlinear dependency relationships. In addition, an external knowledge graph can be introduced to assist the model in understanding and reasoning the semantic association between levels, further improving the reasoning ability of the model.

[0153] The sample rapid adaptation implementation module performs efficient knowledge transfer based on the output of the semantic hierarchy prototype network construction module.

[0154] In addition, the application also proposes a few-sample rapid adaptation implementation method for efficient knowledge transfer using a small amount of samples in a new field. It should be understood that this module comprehensively uses all the components constructed in the foregoing, and is a key innovative point of the technical solution. Through this module, the system can realize effective cross-domain adaptation under the condition of only 5 to 10 samples. Specifically, the following sub-steps are included:

[0155] Step 4.1, small sample learning strategy;

[0156] According to an embodiment of the application, based on the meta-learning framework and the semantic hierarchy prototype network constructed in the foregoing, a small sample learning strategy is realized. When entering a new field, only 5 to 10 representative risk description samples need to be collected, which are converted into hierarchical semantic representations by the recurrent neural parser, and are paired with the standard representations of these samples in the source field to form a small number of mapping examples.

[0157] It should be noted that in the sample selection stage, the application adopts a representative sample screening algorithm to ensure that the small number of samples selected can cover the main risk types and expression methods of the target field through clustering and diversity analysis. In addition, an active learning strategy can also be combined to select the sample with the largest amount of information from the initial sample set for labeling, further improving the efficiency of small sample learning.

[0158] As shown in Figure 6 the application method and the traditional method in the new field adaptation required sample quantity comparison is shown, the application method only needs 10 samples to complete the new field adaptation, and the traditional method needs more than 1000 samples, the sample demand reduces 98%.

[0159] Step 4.2, parameter efficient fine-tuning;

[0160] According to an embodiment of the application, a parameter efficient fine-tuning method is constructed, which only adjusts the parameter subset in the mapping function that is most sensitive to the characteristics of the new field, and keeps most of the pre-trained parameters unchanged. The low-rank adapter (Low-Rank Adapters) technology is adopted to insert a small number of adapter layers between the original network layers to realize the adaptation to the characteristics of a specific field while maintaining the universality of the model. This method reduces the number of parameters that need to be updated, thereby reducing the risk of overfitting.

[0161] In some embodiments, a parameter freezing strategy can be optionally adopted, according to the parameter sensitivity analysis results, the parameter layer with low sensitivity is completely frozen, and only the parameter layer with high sensitivity is updated, so as to further improve the fine-tuning efficiency. In addition, the application also proposes a mixed precision training strategy, which adopts different calculation precisions for parameters with different importance, thereby reducing the calculation resource demand while maintaining the performance.

[0162] Step 4.3, hierarchical progressive fine-tuning;

[0163] According to an embodiment of the application, a hierarchical progressive fine-tuning strategy is implemented, starting from high-level semantics and gradually fine-tuning to low-level semantics. First, the mapping parameters of the abstract layer are adjusted to establish the correspondence of the inter-domain risk framework; then, the intermediate layer parameters are adjusted to adapt to the domain-specific risk relationship structure; finally, the bottom layer parameters are adjusted to handle the mapping relationship of specific terms and entities. This top-down fine-tuning method can efficiently adapt to the differences of different levels while maintaining consistency.

[0164] It should be understood that in other embodiments, a bidirectional fine-tuning strategy can also be adopted, that is, starting from the top layer and the bottom layer at the same time, and converging to the intermediate layer. This strategy can be more effective when dealing with domains where the differences between semantic frameworks and specific terms are equally important.

[0165] Step 4.4, zero-shot generalization ability enhancement;

[0166] According to an embodiment of the application, a zero-shot generalization enhancement module is constructed, so that the system can still make inferences in the case of partial lack of samples. Based on the known mapping relationship, the module uses transitivity reasoning and semantic similarity propagation to predict the mapping of unseen risk concepts. By modeling the relationship network between concepts using a graph neural network, and propagating mapping information on the graph using a message passing algorithm, the mapping inference of unseen concepts is realized.

[0167] Optionally, the application also proposes a hybrid reasoning framework that combines the advantages of symbolic reasoning and neural reasoning to enhance the interpretability and generalization ability of the model. The symbolic reasoning part uses the concept relationships in the risk ontology and domain knowledge base for explicit reasoning, and the neural reasoning part handles fuzzy matching and implicit associations, and the combination of the two forms a powerful zero-shot generalization ability.

[0168] As shown in Figure 7 The application method completes the new domain adaptation in only 24 hours, while the traditional method takes 336 hours (about 2 weeks), and the adaptation speed is increased by more than 90%.

[0169] The standardized risk knowledge base construction and application module receives the output of the sample rapid adaptation implementation module and realizes risk standardization and application;

[0170] Finally, the technical achievements of the foregoing modules are applied to actual risk standardization scenarios, a unified standardized risk knowledge base is constructed, and specific applications in project management are implemented. Notably, this module converts theoretical innovations into practical applications, embodying the practical value of the technical solution. Specifically, the following sub-steps are included:

[0171] Step 5.1, construction of a bidirectional mapping relationship library;

[0172] According to an embodiment of the present application, based on the semantic level adaptation learning result, a bidirectional mapping relationship library between the source domain and the target domain is constructed. This relationship library not only contains mapping at the concept level, but also contains conversion rules at the structure level, and can support bidirectional conversion of complex risk descriptions. The relationship library is stored in a graph database, which facilitates the representation and efficient query of complex relationships.

[0173] It should be noted that the bidirectional mapping relationship library of the present application adopts a hierarchical triple structure, which is different from traditional flat knowledge graphs. Each mapping relationship contains rich meta-information such as hierarchical identifier, relationship type, and confidence score, supporting accurate cross-level relationship query and reasoning. In addition, the present application also provides a mapping consistency verification method, which evaluates the quality of the mapping relationship through a cycle conversion test (source domain→target domain→source domain) to ensure the consistency of bidirectional conversion.

[0174] Step 5.2, risk automatic standardization processing flow;

[0175] According to an embodiment of the present application, a standardization processing pipeline is constructed to realize automatic standardization processing of risk descriptions. The pipeline includes text preprocessing, recursive semantic parsing, hierarchical representation extraction, cross-domain mapping application, result generation and verification, etc. The system automatically judges the domain to which the input risk description belongs, and applies the corresponding mapping rules to convert it into a standardized format while preserving the original structure and semantic integrity.

[0176] In some embodiments, a multi-path processing strategy can be optionally introduced, with multiple different mapping paths being used to process the same risk description, and then the optimal result is selected through voting or fusion algorithms to improve the accuracy and robustness of the standardization processing. In addition, the present application also provides incremental processing capability, which can effectively process newly added risk description fragments without reprocessing the entire document.

[0177] Step 5.3, human-machine collaborative verification and correction system;

[0178] According to the embodiments of the present application, a man-machine collaborative verification and correction system is implemented to further improve the standardization quality in combination with expert knowledge. The system calculates a confidence score for each automatic standardization result, and submits the results with low confidence to domain experts for verification and correction. The feedback of the experts is fed back to the system through an active learning method to continuously improve the mapping model and the semantic hierarchical prototype network.

[0179] It should be understood that in the man-machine collaborative verification process, the present application provides a hierarchical review strategy to allocate tasks to experts of different levels according to the importance and uncertainty of the results, optimizing the efficiency of human resource utilization. In addition, the system also has self-learning ability, which can summarize the correction mode from the correction behavior of the experts, and gradually reduce the need for manual intervention.

[0180] As shown in Figure 8 , the comparison of the man-machine collaborative method and the pure automation method of the present application in the accuracy of standardization results is shown. The accuracy of the standardization results of the man-machine collaborative method of the present application reaches 92%, while the pure automation method is 78%, which is improved by 18%.

[0181] Step 5.4, diversified project management application interface;

[0182] According to the embodiments of the present application, a diversified project management application interface is constructed, so that the standardized risk knowledge can be seamlessly integrated into the existing project management system. The interface includes functional modules such as risk identification, risk classification, risk correlation analysis, and cross-domain risk comparison. Services are provided through API or microservice mode to support the calling and integration of different project management platforms.

[0183] As can be seen, the diversified application interface proposed in the present application not only supports standard RESTAPI calling, but also provides an event-driven interface that can respond to real-time risk change events in the project. In addition, the interface also supports batch processing mode and stream processing mode to adapt to different business scenario requirements.

[0184] A computer storage medium includes a memory and one or more processors, the memory having stored executable code, and the one or more processors executing the executable code to implement the above-mentioned project intelligent management analysis device based on security risk standardization.

[0185] The embodiments of the present application are described above, but the embodiments are not limited to the specific implementation described above, which is only illustrative and not limiting. Those skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which are all within the protection of the embodiments.

Claims

1. A project intelligent management and analysis device based on security risk standardization, characterized in that, The application comprises the following steps: A hierarchical semantic space representation model construction module is used to construct a multi-level semantic representation and output a hierarchical semantic tree. The hierarchical semantic space representation model construction module comprises a semantic hierarchy definition, a recursive neural parser, a hierarchical feature extraction algorithm, and a semantic space pre-training. The recursive neural parser is based on a tree structure long short-term memory network and is used to parse risk text into a hierarchical semantic tree structure. The recursive neural parser comprises an input layer, a tree structure encoding layer, an attention layer, and an output layer. The input layer tokenizes and performs word embedding processing on the text, and maps each word to a dense vector representation. The tree structure encoding layer uses an improved Child-SumTree-LSTM unit to explicitly model the dependency relationship in the syntax tree. Each Tree-LSTM unit comprises an input gate, an output gate, a forget gate, a memory unit, and a hidden state. The Tree-LSTM is equipped with a separate forget gate for each child node. The recursive neural parser integrates a semantic role labeling module that is specifically designed for risk expressions. The module identifies event triggers, risk objects, consequences, conditions, and responsible subjects in risk descriptions and encodes this information into the corresponding nodes of the tree. The semantic hierarchy definition defines a multi-level semantic space structure: Wherein H represents a hierarchical semantic space; h1 represents the most abstract semantic layer, mainly containing field-independent general risk concepts; h2 represents a semantic layer more specific than h1; hL represents the most specific semantic layer, containing field-specific detailed risk descriptions; the intermediate layers h2 to represent transition layers from abstract to specific in turn, each layer carrying more refined semantic information, realizing the layer-by-layer refinement and decomposition of semantics; L1 represents the total number of semantic layers; The meta-learning driven cross-domain mapping function construction module receives the hierarchical semantic tree and constructs a cross-domain mapping function to realize cross-domain knowledge transfer. 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 double-layer optimization process. The loss function of the hierarchical adaptation loss function construction is: wherein, represents the hierarchical adaptation loss function of the mapping function with parameters θ2; ∑ represents a summation operator, l1 is the index of the semantic hierarchy, and L2 is the total number of semantic hierarchies; represents the weight of the l1th semantic hierarchy, which is adaptively adjusted according to the domain difference; is the semantic representation of the l1th layer in the source domain and the semantic representation of the target domain ; and and represent the semantic representation of the source domain and the target domain at the l1th layer, respectively. The distance measurement function d uses different calculation methods according to the characteristics of the semantic hierarchy. For the most abstract top-level semantics, the calculation method based on the Wasserstein distance is used, i.e., when l1 = 1: wherein represents the distance measurement of the 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; and W2 represents the 2-Wasserstein distance function. represents the probability distribution of the source domain; represents the probability distribution of the target domain; For intermediate semantic hierarchies, a structure-sensitive graph distance measurement is used, i.e., when 1 < l1 < L2, the form is: wherein represents the distance measurement of the intermediate-level semantics; represents the semantic representation of the l1th layer in the source domain; represents the semantic representation of the l1th layer in the target domain; represents the adjacency matrix of the source domain; represents the adjacency matrix of the target domain; ||·||F represents the Frobenius norm; and λ1 represents the balance coefficient. represents the feature matrix of the source domain; represents the feature matrix of the target domain; ||·||2 represents the L2 norm; and for the most specific bottom-level semantics, a context-enhanced cosine distance is used, i.e., when l1 = L2, it is represented as: wherein represents the distance measurement of the bottom-level semantics. represents the most specific underlying semantic representation of the source domain; represents the most specific underlying semantic representation of the target domain; cos represents the cosine similarity function; γ represents the context similarity adjustment factor function; represents the context information of the source domain; represents the context information of the target domain; A semantic hierarchical prototype network construction module receives the output of a cross-domain mapping function and maps complex risk structures. This module includes multi-level prototype representation construction, multi-level attention method implementation, prototype updating and maintenance, and cross-level relationship modeling. The multi-level prototype representation maintains a set of prototype vectors for each semantic level, with each prototype vector representing a typical semantic pattern. 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 uses a graph attention network structure to organize prototypes at each level into a directed graph structure, with edges representing the association strength between prototypes at different levels. The multi-level prototype representation, for each semantic level… Representative risk patterns are extracted from historical data using clustering methods as prototypes. Each prototype represents a typical risk pattern, and the prototypes are represented as follows: in This represents the set of prototypes for the l2nd semantic level. They represent the 1st, 2nd, and 3rd elements of the 12th layer, respectively. One prototype vector; l2 represents the semantic hierarchy index; Indicates the number of prototypes in the l2 layer; A sample rapid adaptation implementation module is used to perform efficient knowledge transfer based on the output of the semantic hierarchy prototype network construction module. The sample rapid adaptation implementation module comprises a small-scale sample learning strategy, a parameter efficient fine-tuning, a hierarchical progressive fine-tuning, and a zero-shot generalization capability enhancement. Efficient transfer of knowledge in a new domain is achieved through 5 to 10 risk description samples. The parameter efficient fine-tuning uses a low-rank adapter technology to adapt to specific domain features by inserting an adapter layer with a number of no more than 10% of the original network layers between the original network layers. The hierarchical progressive fine-tuning implements a top-down parameter adjustment method that starts from high-level semantics and adjusts parameters layer by layer to low-level semantics according to a preset hierarchical order. A standardized risk knowledge base construction and application module is used to receive the output of the sample rapid adaptation implementation module and implement risk standardization and application. The standardized risk knowledge base construction and application module comprises a bidirectional mapping relationship database construction, a risk automatic standardization processing flow, a man-machine collaborative verification and correction system, and a diversified project management application interface. The bidirectional mapping relationship database uses a hierarchical triple structure. Each mapping relationship comprises a hierarchical identifier, a relationship type, and a credibility score.

2. A computer storage medium, characterized in that, The application comprises a memory and one or more processors. The memory stores executable code. When the one or more processors execute the executable code, a project intelligent management analysis device based on security risk standardization is implemented.

Images