Data-driven cross-domain intelligent asset knowledge reasoning and value assessment method and system

Abstract

The data-driven cross-domain smart asset knowledge reasoning and value assessment method and system of the present invention relates to the field of smart asset technology, including extracting multi-domain data features of smart assets based on deep learning models, building a cross-domain smart asset knowledge graph; parsing user evaluation goals based on pre-trained language models, retrieving related assets in the knowledge graph, and using attention mechanisms and graph reasoning algorithms for knowledge reasoning; based on the value assessment model trained by reinforcement learning, combined with a time series prediction algorithm, generating a final assessment value and an assessment report containing a confidence interval. The present invention can effectively integrate multi-source heterogeneous data, realize knowledge reasoning and value assessment of cross-domain smart assets, and improve assessment efficiency and accuracy.

Description

Technical Field

[0001] The present invention relates to intelligent asset technology, and in particular to a data-driven cross-domain intelligent asset knowledge reasoning and value assessment method and system. Background Art

[0002] With the rapid development of artificial intelligence and Internet of Things technologies, more and more smart assets are emerging, such as smart robots, smart homes, smart factory equipment, etc. These smart assets are usually cross-domain, multimodal, and high-value, and effective knowledge reasoning and value assessment are essential. Traditional asset assessment methods mainly rely on manual experience and simple statistical models, which are difficult to adapt to the complexity and dynamics of smart assets. How to effectively use cross-domain data for knowledge reasoning and value assessment of smart assets has become an urgent problem to be solved.

[0003] First, traditional evaluation methods have difficulty in dealing with the complex relationships among cross-domain smart assets. Smart assets usually involve multiple fields, such as manufacturing, energy, and medical care, and their value is affected by multiple factors. Traditional evaluation methods lack the ability to integrate and reason about cross-domain knowledge, making it difficult to accurately evaluate the true value of smart assets.

[0004] Secondly, the existing evaluation models lack effective capture of the dynamic characteristics of smart assets. The operating status and market environment of smart assets will continue to change over time, which will have a significant impact on their value. Traditional static evaluation models are difficult to adapt to such dynamic changes, resulting in a decrease in the accuracy and reliability of evaluation results.

[0005] Finally, existing evaluation methods have difficulty in providing interpretable evaluation results. Traditional evaluation models are usually black boxes, and it is difficult to understand their internal reasoning process and value judgment basis. This makes it difficult for users to trust and accept the evaluation results, and also limits the promotion and application of evaluation methods. Summary of the invention

[0006] The embodiments of the present invention provide a data-driven cross-domain intelligent asset knowledge reasoning and value assessment method and system, which can solve the problems in the prior art.

[0007] According to a first aspect of the embodiments of the present invention,

[0008] Provide data-driven cross-domain intelligent asset knowledge reasoning and value assessment methods, including:

[0009] Acquire smart asset data in multiple fields, the smart asset data including asset attribute information, operation status information, historical transaction information and market evaluation information; extract features from the smart asset data and construct asset feature vectors based on a deep learning model; map the asset feature vectors to a preset knowledge graph ontology structure and establish associations between smart asset nodes; calculate weights for the associations based on a graph neural network to form a cross-domain smart asset knowledge graph;

[0010] Receive the smart asset evaluation target input by the user, and parse the semantic features of the evaluation target based on the pre-trained language model; perform path retrieval in the cross-domain smart asset knowledge graph according to the semantic features to obtain a set of asset nodes related to the evaluation target; apply the attention mechanism to the asset node set to perform correlation calculation and determine the importance weight; perform multi-hop reasoning on nodes whose importance weight is greater than a preset weight threshold based on the graph reasoning algorithm to form an asset knowledge reasoning path; match the asset knowledge reasoning path with a preset reasoning rule template to generate a knowledge reasoning result;

[0011] The value influencing factors of the smart assets are extracted according to the knowledge reasoning results, and the value influencing factors are input into a pre-trained value assessment model; the value assessment model is trained based on historical transaction data and market assessment data, and the assessment parameters are continuously optimized using a reinforcement learning method; the initial assessment results of the smart assets are obtained; the initial assessment results are dynamically adjusted based on a time series prediction algorithm to generate a final assessment value including a confidence interval; the final assessment value is associated with the knowledge reasoning results to form an assessment report.

[0012] Mapping the asset feature vector to a preset knowledge graph ontology structure to establish associations between smart asset nodes; and weighting the associations based on a graph neural network to form a cross-domain smart asset knowledge graph, including:

[0013] Mapping the asset feature vector to a preset knowledge graph ontology structure through a feature mapping model, wherein the feature mapping model calculates the semantic distance between the asset feature vector and the ontology concept vector based on cosine similarity, assigns weight values ​​to different dimensions of the asset feature vector through an attention mechanism, and selects the ontology concept with the highest semantic similarity as the mapping target node;

[0014] A multi-dimensional association analysis model is established for the mapping target node, wherein the multi-dimensional association analysis model constructs association relationships between nodes based on the attribute information, historical data information, spatial location information, and business process information of the smart asset, and the association relationships are formed into an initial knowledge graph; the nodes of the initial knowledge graph contain characteristic attributes of the smart asset, and the edges between the nodes represent the association types between the assets;

[0015] Inputting the initial knowledge graph into a graph neural network for weight calculation, the graph neural network comprises multiple graph attention layers, each graph attention layer is provided with multiple attention heads for learning the association pattern between nodes; calculating the hidden layer representation of the node based on the output of the attention head, and calculating the feature similarity, structural similarity and temporal similarity between the node pairs according to the hidden layer representation;

[0016] The feature similarity, the structural similarity and the temporal similarity are input into a multi-task learning model, which simultaneously optimizes the node classification task and the edge prediction task, and uses a graph-structure-based attention mechanism to adaptively adjust the weight coefficients of different types of associations, and finally outputs the association weights between smart asset nodes to form a cross-domain smart asset knowledge graph containing weight information.

[0017] Receiving the smart asset evaluation target input by the user, parsing the semantic features of the evaluation target based on the pre-trained language model; performing path retrieval in the cross-domain smart asset knowledge graph according to the semantic features, and obtaining the asset node set related to the evaluation target includes:

[0018] Receive the evaluation target input by the user, and normalize the evaluation target based on the evaluation target semantic template, wherein the evaluation target semantic template includes evaluation object information, evaluation dimension information, and evaluation constraint condition information; input the normalized evaluation target into the pre-trained language model, and the pre-trained language model performs semantic analysis on the evaluation target through the extended smart asset domain vocabulary, and adopts a multi-head cross-attention mechanism to respectively construct an evaluation factor attention head, a constraint condition attention head, and a context association attention head, and the multi-head cross-attention mechanism outputs word-level semantic features, phrase-level semantic features, and sentence-level semantic features;

[0019] Decomposing the word-level semantic features, the phrase-level semantic features, and the sentence-level semantic features into multiple query atoms, matching the query atoms with a preset query template library based on semantic similarity calculation, and generating a query statement including node matching conditions, path constraint conditions, multi-hop relationship conditions, and attribute filtering conditions according to the matching results;

[0020] Based on the query statement, a bidirectional extended search is initiated from both the evaluation target node and the candidate asset node in the cross-domain smart asset knowledge graph, the priority of the search direction is determined by the node connectivity, and the search pruning is performed using the semantic relevance; the retrieval task is decomposed into multiple parallel subtasks, and an asynchronous update mechanism is used to maintain the global optimal path; the retrieved path is comprehensively scored, and the comprehensive score is calculated based on the semantic relevance, path credibility, path novelty, and path completeness of the path to obtain a path quality score;

[0021] The correlation score between the node and the evaluation target is calculated according to the path quality score, and the correlation score is obtained by weighted calculation of direct semantic correlation, indirect association strength, temporal correlation, and business correlation; based on the correlation score, the candidate asset nodes are screened and dynamically updated to obtain a set of asset nodes related to the evaluation target.

[0022] Based on the query statement, a bidirectional extended search is initiated from both the evaluation target node and the candidate asset node in the cross-domain smart asset knowledge graph, the priority of the search direction is determined by the node connectivity, and the search pruning is performed using the semantic relevance; the retrieval task is decomposed into multiple parallel subtasks, and an asynchronous update mechanism is used to maintain the global optimal path; the retrieved path is comprehensively scored, and the comprehensive score is calculated based on the semantic relevance, path credibility, path novelty, and path completeness to obtain a path quality score including:

[0023] Extracting an evaluation target node and a candidate asset node from the query statement; calculating node connectivity for the evaluation target node and the candidate asset node respectively, and generating a node importance vector based on the node connectivity; inputting the query statement into a pre-trained language model to obtain a query semantic vector, and setting a search pruning threshold based on the query semantic vector;

[0024] Taking the evaluation target node and the candidate asset node as starting points, a bidirectional expansion search is simultaneously initiated in the cross-domain smart asset knowledge graph, wherein the bidirectional expansion search determines the priority of the search direction based on the node importance vector; during the search process, the similarity between the path to be expanded and the query semantic vector is calculated, and when the similarity is lower than the search pruning threshold, the expansion in this direction is stopped;

[0025] The search space is divided into multiple sub-areas, and independent parallel sub-tasks are constructed in each of the sub-areas to perform the bidirectional extended search; each parallel sub-task maintains its own local optimal path, and the local optimal paths are aggregated to obtain a global optimal path through an asynchronous update mechanism; the asynchronous update mechanism determines the update timing based on the path quality to avoid the interference of low-quality paths on the global optimal path;

[0026] The retrieved paths are scored in multiple dimensions: the similarity between the path node sequence and the query semantic vector is calculated to obtain the path semantic relevance, the path credibility is calculated based on the reliability of the path relationship and the authority of the knowledge source, the path novelty is calculated based on the rarity of the path pattern in the knowledge graph, and the path completeness is calculated based on the key node coverage and semantic completeness; an adaptive weight method is used to weightedly fuse the path semantic relevance, the path credibility, the path novelty, and the path completeness to obtain a path quality score.

[0027] Apply the attention mechanism to the asset node set to perform correlation calculation and determine the importance weight; perform multi-hop reasoning on the nodes whose importance weight is greater than the preset weight threshold based on the graph reasoning algorithm to form an asset knowledge reasoning path; match the asset knowledge reasoning path with the preset reasoning rule template to generate knowledge reasoning results including:

[0028] Perform multi-level attention calculation on the asset node set, calculate the structural importance score based on the node degree centrality and the eigenvector centrality, calculate the semantic importance score based on the semantic association between the node attribute and the evaluation target, and calculate the temporal importance score based on the time decay factor and the state transition frequency; perform weighted combination of the structural importance score, the semantic importance score, and the temporal importance score to obtain the node comprehensive importance weight;

[0029] Based on the comprehensive importance weight of the node, a seed node with a weight greater than a preset weight threshold is selected, a directed reasoning graph is constructed in the cross-domain smart asset knowledge graph, and a maximum reasoning depth is determined according to the comprehensive importance weight of the seed node; path extension is performed on the directed reasoning graph, and the extension direction is constrained by the semantic association between nodes; a reliability score is calculated for the extended path based on the comprehensive importance weight of the path node, and a coherence score is calculated based on the semantic continuity of the path, and a path with both the reliability score and the coherence score higher than the preset score threshold is selected to form an asset knowledge reasoning path;

[0030] The paths in the asset knowledge reasoning path are structurally matched and semantically matched with the preset reasoning rule templates, and supporting evidence is extracted based on the matching results. The supporting evidence includes key nodes and relationships on the path; the supporting evidence is verified for rule consistency and logical integrity, and based on the verification results, trusted paths with scores higher than the preset score threshold are screened and conflicts are resolved to generate the final knowledge reasoning results.

[0031] Obtaining the initial evaluation results of the smart asset; dynamically adjusting the initial evaluation results based on the time series prediction algorithm to generate a final evaluation value including a confidence interval; associating the final evaluation value with the knowledge reasoning results to form an evaluation report including:

[0032] Obtain the initial evaluation result of the smart asset, form a time series based on the initial evaluation result and historical evaluation data, decompose the time series to obtain trend component, periodic component and random component; input the trend component into the long short-term memory network to extract the long-term dependency feature vector, input the periodic component into the Prophet model to extract the seasonal feature vector, and input the random component into the autoregressive model to extract the volatility feature vector; construct a Stacking model, input the long-term dependency feature vector, the seasonal feature vector and the volatility feature vector into the Stacking model to obtain a fusion feature vector;

[0033] Inputting the fused feature vector into the correction model to dynamically correct the initial evaluation result to obtain a corrected evaluation value, using the quantile regression method to calculate the upper and lower limits of the confidence interval of the corrected evaluation value, and obtaining a dynamic evaluation result including a confidence interval;

[0034] In a relational database, based on the dynamic evaluation results, the technical dependent nodes of the evaluation object are screened to build an association network, the transmission coefficient between the nodes in the network is calculated to obtain a transmission feature vector, and the market linkage feature vector is obtained according to the correlation coefficient between the market index and the transmission feature vector; the transmission feature vector, the market linkage feature vector and the dynamic evaluation results are input into a principal component analysis model, and features with a value contribution greater than 0.7 are extracted as core value element vectors, features with a volatility greater than 0.3 are extracted as risk element vectors, and features with a time series autocorrelation coefficient greater than 0.6 are extracted as trend element vectors;

[0035] The main framework of the evaluation report is constructed based on the core value element vector, and the risk element vector and the trend element vector are mapped to corresponding chapters; the newly added data is input into the value evaluation model to update the dynamic evaluation results, and the corresponding chapters in the evaluation report are dynamically updated according to the updated dynamic evaluation results; wherein the newly added data includes updated data of the conduction characteristic vector and the market linkage characteristic vector.

[0036] The fused feature vector is input into the correction model to dynamically correct the initial evaluation result to obtain a corrected evaluation value, and the upper and lower limits of the confidence interval of the corrected evaluation value are calculated using the quantile regression method. The dynamic evaluation result including the confidence interval is obtained, including:

[0037] Input the fused feature vector into a feature mapping network to obtain a mapping feature; construct an attention weight calculation module to calculate a feature attention score based on the correlation between the mapping feature and the historical correction effect; and weight the feature attention score and the mapping feature to obtain a correction feature;

[0038] Based on the correction feature, a dynamic correction model is constructed, and the initial evaluation result is input into the dynamic correction model to obtain a correction evaluation value; an exponential sliding average method is used to track the correction amplitude, and the correction step length is adaptively adjusted according to the correction amplitude; a historical record of the correction evaluation value is established for dynamic update of subsequent correction parameters;

[0039] Constructing a quantile regression model group, inputting the modified evaluation value and the modified feature into the quantile regression model group to obtain prediction values ​​of different quantiles; constructing a distributed loss function based on the predicted value, wherein the distributed loss function considers both prediction accuracy and distribution rationality;

[0040] The quantile regression model group is trained by adopting an alternating optimization strategy: first, the distribution parameters are fixed to optimize the quantile prediction model to obtain the regression coefficient of each quantile; then, the regression coefficient is fixed to optimize the distribution parameters to ensure the continuity of the prediction distribution; the prediction error of each quantile is calculated based on the verification data, and the optimal model parameters are selected;

[0041] The modified evaluation value is input into the trained quantile regression model group to calculate the upper limit value of the confidence interval and the lower limit value of the confidence interval; the modified evaluation value is combined with the upper limit value of the confidence interval and the lower limit value of the confidence interval to generate a dynamic evaluation result including the confidence interval.

[0042] According to a second aspect of the embodiments of the present invention,

[0043] Provides a data-driven cross-domain intelligent asset knowledge reasoning and value assessment system, including:

[0044] The first unit is used to obtain smart asset data in multiple fields, wherein the smart asset data includes asset attribute information, operation status information, historical transaction information and market evaluation information; perform feature extraction on the smart asset data, and construct an asset feature vector based on a deep learning model; map the asset feature vector to a preset knowledge graph ontology structure, and establish an association relationship between smart asset nodes; perform weight calculation on the association relationship based on a graph neural network to form a cross-domain smart asset knowledge graph;

[0045] The second unit is used to receive the smart asset evaluation target input by the user, parse the semantic features of the evaluation target based on the pre-trained language model; perform path retrieval in the cross-domain smart asset knowledge graph according to the semantic features to obtain a set of asset nodes related to the evaluation target; apply the attention mechanism to the asset node set to perform correlation calculation and determine the importance weight; perform multi-hop reasoning on the nodes whose importance weight is greater than the preset weight threshold based on the graph reasoning algorithm to form an asset knowledge reasoning path; match the asset knowledge reasoning path with a preset reasoning rule template to generate a knowledge reasoning result;

[0046] The third unit is used to extract the value influencing factors of the intelligent assets according to the knowledge reasoning results, and input the value influencing factors into a pre-trained value assessment model; the value assessment model is trained based on historical transaction data and market assessment data, and the assessment parameters are continuously optimized by a reinforcement learning method; the initial assessment results of the intelligent assets are obtained; the initial assessment results are dynamically adjusted based on a time series prediction algorithm to generate a final assessment value including a confidence interval; the final assessment value is associated with the knowledge reasoning results to form an assessment report.

[0047] According to a third aspect of the embodiments of the present invention,

[0048] An electronic device is provided, comprising:

[0049] processor;

[0050] a memory for storing processor-executable instructions;

[0051] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0052] According to a fourth aspect of the embodiments of the present invention,

[0053] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0054] The beneficial effects of this application are as follows:

[0055] 1. Improve the comprehensiveness and accuracy of the evaluation: This method builds a cross-domain intelligent asset knowledge graph and combines it with graph neural networks and attention mechanisms to more comprehensively consider the relationship between assets in different fields and accurately identify key assets related to the evaluation targets, thereby improving the accuracy and reliability of the evaluation results.

[0056] 2. Realize intelligent knowledge reasoning and value assessment: This method uses pre-trained language models, graph reasoning algorithms and preset reasoning rule templates to automatically perform knowledge reasoning, and combines reinforcement learning to continuously optimize the value assessment model to achieve intelligent value assessment, improve assessment efficiency and reduce manual intervention.

[0057] 3. Provide more interpretable and dynamic evaluation results: This method associates the knowledge reasoning results with the final evaluation value to form an evaluation report, which enhances the interpretability of the evaluation results. At the same time, the time series prediction algorithm is used to dynamically adjust the evaluation results and provide confidence intervals, making the evaluation results closer to the actual market situation and having a higher reference value. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A schematic diagram of a flow chart of a data-driven cross-domain intelligent asset knowledge reasoning and value assessment method according to an embodiment of the present invention;

[0059] Figure 2 Schematic diagram of the structure of the data-driven cross-domain intelligent asset knowledge reasoning and value assessment system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0062] Figure 1 Schematic diagram of the process of the data-driven cross-domain intelligent asset knowledge reasoning and value assessment method according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0063] S11. Acquire smart asset data in multiple fields, the smart asset data including asset attribute information, operation status information, historical transaction information and market evaluation information; extract features from the smart asset data, and construct asset feature vectors based on a deep learning model; map the asset feature vectors to a preset knowledge graph ontology structure, and establish associations between smart asset nodes; calculate weights for the associations based on a graph neural network to form a cross-domain smart asset knowledge graph;

[0064] S12. Receive the smart asset evaluation target input by the user, and parse the semantic features of the evaluation target based on the pre-trained language model; perform path retrieval in the cross-domain smart asset knowledge graph according to the semantic features to obtain a set of asset nodes related to the evaluation target; apply the attention mechanism to the asset node set to perform correlation calculation and determine the importance weight; perform multi-hop reasoning on nodes whose importance weight is greater than a preset weight threshold based on the graph reasoning algorithm to form an asset knowledge reasoning path; match the asset knowledge reasoning path with a preset reasoning rule template to generate a knowledge reasoning result;

[0065] S13. Extract the value impact factors of the smart assets based on the knowledge reasoning results, and input the value impact factors into a pre-trained value assessment model; train the value assessment model based on historical transaction data and market assessment data, and use reinforcement learning methods to continuously optimize assessment parameters; obtain initial assessment results of the smart assets; dynamically adjust the initial assessment results based on a time series prediction algorithm to generate a final assessment value including a confidence interval; associate the final assessment value with the knowledge reasoning results to form an assessment report.

[0066] In an optional implementation, the asset feature vector is mapped to a preset knowledge graph ontology structure to establish an association relationship between smart asset nodes; and the association relationship is weighted based on a graph neural network to form a cross-domain smart asset knowledge graph, including:

[0067] Mapping the asset feature vector to a preset knowledge graph ontology structure through a feature mapping model, wherein the feature mapping model calculates the semantic distance between the asset feature vector and the ontology concept vector based on cosine similarity, assigns weight values ​​to different dimensions of the asset feature vector through an attention mechanism, and selects the ontology concept with the highest semantic similarity as the mapping target node;

[0068] A multi-dimensional association analysis model is established for the mapping target node, wherein the multi-dimensional association analysis model constructs association relationships between nodes based on the attribute information, historical data information, spatial location information, and business process information of the smart asset, and the association relationships are formed into an initial knowledge graph; the nodes of the initial knowledge graph contain characteristic attributes of the smart asset, and the edges between the nodes represent the association types between the assets;

[0069] Inputting the initial knowledge graph into a graph neural network for weight calculation, the graph neural network comprises multiple graph attention layers, each graph attention layer is provided with multiple attention heads for learning the association pattern between nodes; calculating the hidden layer representation of the node based on the output of the attention head, and calculating the feature similarity, structural similarity and temporal similarity between the node pairs according to the hidden layer representation;

[0070] The feature similarity, the structural similarity and the temporal similarity are input into a multi-task learning model, which simultaneously optimizes the node classification task and the edge prediction task, and uses a graph-structure-based attention mechanism to adaptively adjust the weight coefficients of different types of associations, and finally outputs the association weights between smart asset nodes to form a cross-domain smart asset knowledge graph containing weight information.

[0071] The method of mapping asset feature vectors to knowledge graph ontology and constructing cross-domain intelligent asset knowledge graph includes the following steps:

[0072] First, preset the knowledge graph ontology structure. This ontology structure is pre-defined according to the specific application scenario, and contains important concepts, entities and the relationships between them in the field. For example, in the field of smart grids, the ontology can include concepts such as power plants, substations, transmission lines, and electrical equipment, as well as the relationships between them such as "connection", "power supply", and "control". Each concept has a corresponding vector representation, called the ontology concept vector, which is used to characterize the semantic information of the concept. The construction of the ontology concept vector can use word embedding technology, such as Word2Vec or BERT, and is obtained by training on large-scale text data.

[0073] Then, the asset feature vector is mapped to the knowledge graph ontology. Each asset has a feature vector that describes the attributes and status of the asset. For example, the feature vector of a substation can contain information such as its voltage level, geographical location, and operating status. In order to map the asset feature vector to the ontology, the semantic distance between the asset feature vector and each ontology concept vector is first calculated. Here, cosine similarity is used to measure the semantic distance, and the larger the value, the more similar the semantics. In order to highlight the importance of different dimensions in the feature vector, the attention mechanism is introduced. The attention mechanism assigns different weight values ​​to different dimensions of the feature vector according to the semantic relevance between the feature vector and the ontology concept vector. For example, if the feature vector of a substation contains voltage level and geographical location information, the voltage level may have a higher weight when mapping it to the concept of "substation". Finally, the ontology concept with the highest semantic distance is selected as the mapping target node, and the asset is mapped to this concept.

[0074] Next, a multi-dimensional association analysis model is constructed to establish the association relationship between the mapping target nodes. The construction of the association relationship is based on the multi-dimensional information of the smart assets, including attribute information, historical data information, spatial location information, and business process information. For example, two substations can be associated through a "connection" relationship if there is a physical connection between them; they can also be associated through a "power supply" relationship if one substation supplies power to another substation. These association relationships constitute the initial knowledge graph, in which nodes represent smart assets and edges represent the types of associations between assets. For example, a node can represent a specific substation, including its voltage level, geographical location and other attribute information; an edge can represent a "connection" relationship between two substations.

[0075] After that, the initial knowledge graph is input into the graph neural network for weight calculation. The graph neural network contains multiple layers of graph attention layers, each with multiple attention heads. Each attention head focuses on different neighbors and features of the node, and learns different association patterns between nodes. For example, one attention head can focus on the direct neighbors of the node, and another attention head can focus on the indirect neighbors of the node. Based on the output of the attention head, the hidden layer representation of the node is calculated, and the feature similarity, structural similarity, and temporal similarity between node pairs are further calculated. Feature similarity measures the similarity of node attributes, structural similarity measures the similarity of node positions in the graph structure, and temporal similarity measures the similarity of node historical data change trends.

[0076] Finally, feature similarity, structural similarity, and temporal similarity are input into the multi-task learning model. The model optimizes both the node classification task and the edge prediction task. The goal of the node classification task is to predict the category to which the node belongs, and the goal of the edge prediction task is to predict whether there is an edge between nodes and the type of edge. The multi-task learning model uses an attention mechanism based on the graph structure to adaptively adjust the weight coefficients of different types of associations. For example, the weight of the "connection" relationship may be higher than the weight of the "power supply" relationship. Finally, the model outputs the association weights between smart asset nodes to form a cross-domain smart asset knowledge graph containing weight information. For example, the weight of the "connection" relationship between two substations can represent the strength or importance of the connection.

[0077] The solution of this application can:

[0078] Improve the efficiency of knowledge fusion: By mapping smart assets in different fields into a unified knowledge graph ontology, cross-domain data and knowledge can be effectively integrated, breaking down data silos and enabling knowledge sharing and reuse. Enhance the accuracy of association relationships: By using multi-dimensional association analysis models and graph neural networks, deeper and more complex association relationships between smart assets can be mined to improve the accuracy and completeness of association relationships. For example, some implicit association relationships can be identified, such as two devices that seem to have no direct association, but may actually be associated through a complex business process. Improve the application value of knowledge graphs: By calculating the weights of association relationships through multi-task learning models, a more sophisticated and richer knowledge graph can be constructed to provide stronger knowledge support for various intelligent applications. For example, more accurate asset risk assessment, fault prediction, and optimization decisions can be made based on association weights.

[0079] In an optional implementation, receiving a smart asset evaluation target input by a user, parsing semantic features of the evaluation target based on a pre-trained language model; performing path retrieval in the cross-domain smart asset knowledge graph according to the semantic features, and obtaining a set of asset nodes related to the evaluation target includes:

[0080] Receive the evaluation target input by the user, and normalize the evaluation target based on the evaluation target semantic template, wherein the evaluation target semantic template includes evaluation object information, evaluation dimension information, and evaluation constraint condition information; input the normalized evaluation target into the pre-trained language model, and the pre-trained language model performs semantic analysis on the evaluation target through the extended smart asset domain vocabulary, and adopts a multi-head cross-attention mechanism to respectively construct an evaluation factor attention head, a constraint condition attention head, and a context association attention head, and the multi-head cross-attention mechanism outputs word-level semantic features, phrase-level semantic features, and sentence-level semantic features;

[0081] Decomposing the word-level semantic features, the phrase-level semantic features, and the sentence-level semantic features into multiple query atoms, matching the query atoms with a preset query template library based on semantic similarity calculation, and generating a query statement including node matching conditions, path constraint conditions, multi-hop relationship conditions, and attribute filtering conditions according to the matching results;

[0082] Based on the query statement, a bidirectional extended search is initiated from both the evaluation target node and the candidate asset node in the cross-domain smart asset knowledge graph, the priority of the search direction is determined by the node connectivity, and the search pruning is performed using the semantic relevance; the retrieval task is decomposed into multiple parallel subtasks, and an asynchronous update mechanism is used to maintain the global optimal path; the retrieved path is comprehensively scored, and the comprehensive score is calculated based on the semantic relevance, path credibility, path novelty, and path completeness of the path to obtain a path quality score;

[0083] The correlation score between the node and the evaluation target is calculated according to the path quality score, and the correlation score is obtained by weighted calculation of direct semantic correlation, indirect association strength, temporal correlation, and business correlation; based on the correlation score, the candidate asset nodes are screened and dynamically updated to obtain a set of asset nodes related to the evaluation target.

[0084] Receive user input for smart asset assessment goals, such as "assess the core technologies in the new energy vehicle industry chain with high growth potential in the next three years."

[0085] First, the evaluation target is standardized based on the evaluation target semantic template. The semantic template contains evaluation object information (core technology in the new energy vehicle industry chain), evaluation dimension information (high growth potential), and evaluation constraint information (next three years). For example, the evaluation target input by the user is parsed into a structured representation: {evaluation object: "core technology in the new energy vehicle industry chain", evaluation dimension: "high growth potential", constraint: "next three years"}.

[0086] Next, the normalized evaluation target is input into the pre-trained language model. The model uses an extended vocabulary of smart asset domains (including professional vocabulary such as "new energy", "automobile", "industry chain", "core technology", and "growth potential") for semantic analysis. The model uses a multi-head cross-attention mechanism to construct evaluation factor attention heads (focusing on "core technology"), constraint attention heads (focusing on "the next three years"), and contextual association attention heads (focusing on "new energy vehicle industry chain"). For example, the model may identify "new energy vehicle industry chain" as an overall concept and associate it with "core technology" to understand that the user wants to evaluate the core technology in the context of the new energy vehicle industry chain. The multi-head cross-attention mechanism outputs word-level semantic features (for example, the word vector of "new energy"), phrase-level semantic features (for example, the phrase vector of "core technology"), and sentence-level semantic features (for example, the vector representation of the entire evaluation target sentence). For example, the word-level features of "core technology" may include its relevance to words such as "innovation", "research and development", and "patent"; the phrase-level features may include its relevance to phrases such as "battery technology", "motor technology", and "autonomous driving technology"; and the sentence-level features include its importance and role in the entire evaluation target.

[0087] Then, the semantic features at the word level, phrase level, and sentence level are decomposed into multiple query atoms. For example, "core technology in the new energy vehicle industry chain" is decomposed into query atoms such as "new energy", "automobile", "industry chain", and "core technology". Based on the semantic similarity calculation, these query atoms are matched with the preset query template library. The query template library contains various types of query templates, such as: finding assets of a specific type, finding assets related to a specific entity, finding assets that meet specific conditions, etc. Assuming that "core technology" matches the template of "finding assets of a specific type", a query statement containing node matching conditions (for example, the node type is "technology"), path constraint conditions (for example, the path length is not less than 1), multi-hop relationship conditions (for example, there is an association relationship with the "new energy vehicle industry chain"), and attribute filtering conditions (for example, having the "high growth potential" attribute) is generated. For example, the generated query statement may be: find nodes associated with the "new energy vehicle industry chain", of type "technology", and with the "high growth potential" attribute.

[0088] Next, based on the query statement, a bidirectional expansion search is initiated from both the evaluation target node (e.g., "new energy vehicle industry chain") and the candidate asset node (e.g., "battery technology", "motor technology") in the cross-domain smart asset knowledge graph. The priority of the search direction is determined by the node connectivity. For example, if "battery technology" is connected to more other nodes, the search direction of "battery technology" is preferentially expanded. Search pruning is performed using semantic relevance. For example, if a node has a low semantic relevance to the evaluation target, the node is stopped from being expanded. The retrieval task is decomposed into multiple parallel subtasks. For example, paths related to "battery technology" and "motor technology" are searched simultaneously. An asynchronous update mechanism is used to maintain the global optimal path. For example, if a better path is found, the global optimal path is updated. The retrieved paths are comprehensively scored. The comprehensive score is calculated based on the path semantic relevance (e.g., the semantic relevance of the nodes on the path to the evaluation target), path credibility (e.g., the credibility of the relationship on the path), path novelty (e.g., whether the path contains new information), and path completeness (e.g., whether the path fully describes the relationship between the evaluation target and the candidate asset). For example, if a path connecting “new energy vehicle industry chain” and “battery technology” contains nodes and relationships such as “R&D investment”, “market share”, and “technical barriers”, its semantic relevance, credibility, and completeness are all high, so the path quality score is also high.

[0089] Finally, the relevance score between the node and the evaluation target is calculated based on the path quality score. The relevance score is obtained by weighted calculation of direct semantic relevance (for example, the semantic similarity between "battery technology" and "core technology"), indirect association strength (for example, the association strength between "battery technology" and "core technology" through "new energy vehicle industry chain"), temporal relevance (for example, the development trend of "battery technology" in the next three years), and business relevance (for example, the impact of "battery technology" on the new energy vehicle industry chain). For example, if the semantic similarity between "battery technology" and "core technology" is high and has high growth potential in the next three years, its relevance score is high. Based on the relevance score, the candidate asset nodes are screened and dynamically updated to obtain a set of asset nodes related to the evaluation target. For example, several asset nodes with the highest relevance scores are screened as the final evaluation result.

[0090] The solution of this application can:

[0091] Improve evaluation efficiency and accuracy: Through semantic parsing, knowledge graph retrieval, and multi-dimensional correlation analysis, assets related to the target can be automatically identified and evaluated, reducing manual intervention and improving evaluation efficiency and accuracy. Expand the scope and depth of evaluation: The cross-domain intelligent asset knowledge graph can integrate data and knowledge from different fields, expand the scope of evaluation, and mine deeper associations through multi-hop relationships to provide more comprehensive evaluation results. Support dynamic update and optimization: This method can dynamically update the knowledge graph and evaluation model based on new data and knowledge, and optimize through user feedback to continuously improve the effectiveness and reliability of the evaluation.

[0092] In an optional implementation, based on the query statement, a bidirectional extended search is initiated from both the evaluation target node and the candidate asset node in the cross-domain smart asset knowledge graph, the priority of the search direction is determined by the node connectivity, and the search pruning is performed using the semantic relevance; the retrieval task is decomposed into multiple parallel subtasks, and an asynchronous update mechanism is used to maintain the global optimal path; the retrieved path is comprehensively scored, and the comprehensive score is calculated based on the semantic relevance of the path, the credibility of the path, the novelty of the path, and the completeness of the path to obtain a path quality score including:

[0093] Extracting an evaluation target node and a candidate asset node from the query statement; calculating node connectivity for the evaluation target node and the candidate asset node respectively, and generating a node importance vector based on the node connectivity; inputting the query statement into a pre-trained language model to obtain a query semantic vector, and setting a search pruning threshold based on the query semantic vector;

[0094] Taking the evaluation target node and the candidate asset node as starting points, a bidirectional expansion search is simultaneously initiated in the cross-domain smart asset knowledge graph, wherein the bidirectional expansion search determines the priority of the search direction based on the node importance vector; during the search process, the similarity between the path to be expanded and the query semantic vector is calculated, and when the similarity is lower than the search pruning threshold, the expansion in this direction is stopped;

[0095] The search space is divided into multiple sub-areas, and independent parallel sub-tasks are constructed in each of the sub-areas to perform the bidirectional extended search; each parallel sub-task maintains its own local optimal path, and the local optimal paths are aggregated to obtain a global optimal path through an asynchronous update mechanism; the asynchronous update mechanism determines the update timing based on the path quality to avoid the interference of low-quality paths on the global optimal path;

[0096] The retrieved paths are scored in multiple dimensions: the similarity between the path node sequence and the query semantic vector is calculated to obtain the path semantic relevance, the path credibility is calculated based on the reliability of the path relationship and the authority of the knowledge source, the path novelty is calculated based on the rarity of the path pattern in the knowledge graph, and the path completeness is calculated based on the key node coverage and semantic completeness; an adaptive weight method is used to weightedly fuse the path semantic relevance, the path credibility, the path novelty, and the path completeness to obtain a path quality score.

[0097] A cross-domain intelligent asset evaluation method based on knowledge graph aims to efficiently and accurately retrieve intelligent assets related to the evaluation target from massive information and conduct a comprehensive evaluation on them. This method achieves accurate evaluation of cross-domain intelligent assets through technical means such as bidirectional extended search, parallel computing and multi-dimensional scoring.

[0098] First, identify the evaluation target and candidate assets from the query statement provided by the user. For example, if the user queries "Evaluate the application value of artificial intelligence in the medical field", then "Application of artificial intelligence in the medical field" is the evaluation target, and "Medical imaging analysis software", "AI-assisted diagnosis system", etc. are candidate assets. The connectivity of the evaluation target and candidate assets in the knowledge graph is calculated respectively. The higher the connectivity, the higher the node importance, so as to generate a node importance vector. At the same time, the query statement is input into a pre-trained language model, such as BERT, to obtain the semantic vector of the query. According to the query semantic vector, the search pruning threshold is set for pruning operations in the subsequent search process.

[0099] Next, starting from the evaluation target and candidate asset nodes, a bidirectional expansion search is started simultaneously in the cross-domain smart asset knowledge graph. The priority of the search direction is determined by the node importance vector. The higher the importance of the node, the higher the priority of its expansion direction. During the search process, the similarity between the path to be expanded and the query semantic vector is calculated. If the similarity is lower than the preset search pruning threshold, the expansion in that direction is stopped to avoid invalid search and improve search efficiency.

[0100] In order to further improve the search efficiency, the search space is divided into multiple sub-areas, and independent parallel sub-tasks are constructed in each sub-area to perform bidirectional extended search. Each parallel sub-task maintains its own local optimal path, and the local optimal paths are aggregated through the asynchronous update mechanism to finally obtain the global optimal path. The asynchronous update mechanism determines the update timing based on the path quality to avoid the interference of low-quality paths on the global optimal path and ensure the accuracy of the final result. For example, the update is only performed when the quality score of the new path exceeds a certain threshold of the quality score of the current global optimal path.

[0101] After the path is retrieved, it is necessary to perform a multi-dimensional scoring on the path. First, the similarity between the path node sequence and the query semantic vector is calculated to obtain the path semantic relevance. Then, the path credibility is calculated based on the reliability of the relationship in the path and the authority of the knowledge source. For example, information released by an authoritative organization has higher credibility. In addition, the path novelty is calculated based on the rarity of the path pattern in the knowledge graph, and rare paths usually contain more valuable information. Finally, the path completeness is calculated based on the key node coverage and semantic completeness. For example, if a path covers all the key aspects of the evaluation target, its completeness is high. The adaptive weight method is used to weightedly fuse the path semantic relevance, path credibility, path novelty and path completeness to obtain the final path quality score. For example, for a specific query, if novelty is more important, a higher weight is given to novelty.

[0102] Case: Assume that the user queries "Evaluate the application value of blockchain technology in the financial field". The evaluation target is "Application of blockchain technology in the financial field", and the candidate assets are "digital currency" and "decentralized trading platform". The knowledge graph contains the relationship "applied to" between "blockchain technology" and "financial field", and the relationship "supported by" between "blockchain technology" and "digital currency". Through bidirectional extended search, a path can be found: "blockchain technology" - "applied to" - "financial field" - "includes" - "digital currency". By calculating the similarity between the path and the query statement, the credibility of the relationship, the rarity of the path, and the coverage of key nodes, the quality score of the path can be obtained.

[0103] The solution of this application can:

[0104] Improve evaluation efficiency: Through bidirectional extended search and parallel computing, relevant information can be quickly retrieved from massive amounts of information, significantly improving evaluation efficiency. Improve evaluation accuracy: The multi-dimensional scoring mechanism comprehensively considers the semantic relevance, credibility, novelty, and completeness of the path, and can more comprehensively and accurately evaluate the value of smart assets. Enhance the comprehensiveness and depth of evaluation: Using the rich semantic information of the knowledge graph, the potential relationship between the evaluation target and the candidate assets can be discovered, providing a more in-depth and comprehensive evaluation result.

[0105] In an optional implementation, the attention mechanism is applied to the asset node set to perform correlation calculation to determine the importance weight; multi-hop reasoning is performed on the nodes whose importance weight is greater than a preset weight threshold based on a graph reasoning algorithm to form an asset knowledge reasoning path; the asset knowledge reasoning path is matched with a preset reasoning rule template to generate a knowledge reasoning result, including:

[0106] Perform multi-level attention calculation on the asset node set, calculate the structural importance score based on the node degree centrality and the eigenvector centrality, calculate the semantic importance score based on the semantic association between the node attribute and the evaluation target, and calculate the temporal importance score based on the time decay factor and the state transition frequency; perform weighted combination of the structural importance score, the semantic importance score, and the temporal importance score to obtain the node comprehensive importance weight;

[0107] Based on the comprehensive importance weight of the node, a seed node with a weight greater than a preset weight threshold is selected, a directed reasoning graph is constructed in the cross-domain smart asset knowledge graph, and a maximum reasoning depth is determined according to the comprehensive importance weight of the seed node; path extension is performed on the directed reasoning graph, and the extension direction is constrained by the semantic association between nodes; a reliability score is calculated for the extended path based on the comprehensive importance weight of the path node, and a coherence score is calculated based on the semantic continuity of the path, and a path with both the reliability score and the coherence score higher than the preset score threshold is selected to form an asset knowledge reasoning path;

[0108] The paths in the asset knowledge reasoning path are structurally matched and semantically matched with the preset reasoning rule templates, and supporting evidence is extracted based on the matching results. The supporting evidence includes key nodes and relationships on the path; the supporting evidence is verified for rule consistency and logical integrity, and based on the verification results, trusted paths with scores higher than the preset score threshold are screened and conflicts are resolved to generate the final knowledge reasoning results.

[0109] A cross-domain intelligent asset knowledge reasoning method is used to evaluate and analyze asset value. The core idea is to perform multi-level attention calculation and graph reasoning on the asset knowledge graph to finally generate knowledge reasoning results.

[0110] First, the importance weights of the asset node set are calculated. This step is divided into three sub-steps: calculating the structural importance score, calculating the semantic importance score, and calculating the temporal importance score. When calculating the structural importance score, the centrality of the node in the graph is considered, such as the number of nodes directly connected to the node (degree centrality) and the ability of the node to propagate information in the graph (eigenvector centrality). The higher the degree centrality, the higher the score. For example, if a node is connected to 10 other nodes and another node is only connected to 2, the former has a higher degree centrality score. When calculating the semantic importance score, the semantic association between the node attribute and the evaluation target is considered. The higher the association, the higher the score. For example, if the evaluation target is "investment potential", the node related to "yield" will score higher than the node related to "color". When calculating the temporal importance score, the time decay factor and state transition frequency are considered. The nodes with more recent time and more frequent state changes have higher scores. For example, an asset with frequent transactions in the past month will have a higher temporal importance score than an asset with no transaction records in a year. Finally, the structural importance score, semantic importance score and temporal importance score are weighted and combined to obtain the comprehensive importance weight of the node. For example, if the structural importance weight accounts for 40%, the semantic importance weight accounts for 30%, and the temporal importance weight accounts for 30%, then the comprehensive importance weight of node A = 0.4*node A's structural importance score + 0.3*node A's semantic importance score + 0.3*node A's temporal importance score.

[0111] Next, multi-hop reasoning is performed based on the comprehensive importance weight of the nodes to form an asset knowledge reasoning path. First, nodes with a comprehensive importance weight greater than the preset weight threshold are selected as seed nodes. For example, if the preset weight threshold is 0.8, all nodes with a comprehensive importance weight greater than or equal to 0.8 will be selected as seed nodes. Then, a directed reasoning graph is constructed in the cross-domain intelligent asset knowledge graph, and the maximum reasoning depth is determined based on the comprehensive importance weight of the seed nodes. For example, if the comprehensive importance weight of seed node A is 0.9, the maximum depth of the reasoning path starting from A can be set to 3 hops. After that, path extension is performed on the directed reasoning graph, and the extension direction is constrained by the semantic association between nodes. For example, if the current node is "real estate", it is preferentially extended to nodes with semantics related to "property tax" and "land use rights", rather than to nodes with semantics unrelated to "stocks" and "bonds". For the extended path, the reliability score is calculated based on the comprehensive importance weight of the path nodes, and the coherence score is calculated based on the semantic continuity of the path. For example, the reliability score of path ABC is obtained by weighted average of the comprehensive importance weights of nodes A, B, and C, and the coherence score depends on the semantic relevance of the relationship between AB and BC. Paths with reliability scores and coherence scores higher than the preset score threshold are selected to form asset knowledge reasoning paths. For example, if the preset reliability score threshold and coherence score threshold are both 0.7, only paths with reliability scores and coherence scores greater than or equal to 0.7 will be retained.

[0112] Finally, the asset knowledge reasoning path is matched with the preset reasoning rule template to generate the knowledge reasoning result. The reasoning path is structurally matched and semantically matched with the preset reasoning rule template. Supporting evidence is extracted based on the matching results, including key nodes and relationships on the path. The supporting evidence is verified for rule consistency and logical integrity. Based on the verification results, the trusted paths with scores higher than the preset score threshold are screened and conflicts are resolved, and finally the knowledge reasoning result is generated.

[0113] The solution of this application can:

[0114] Improve asset evaluation efficiency: Through automated reasoning, reduce the workload of manual analysis, thereby improving the efficiency of asset evaluation. Enhance the objectivity of asset evaluation: Through multi-level attention mechanism and graph reasoning algorithm, reduce the influence of subjective judgment and make the evaluation results more objective. Discover potential investment opportunities: Through deep mining of asset knowledge graph, potential investment opportunities can be discovered and decision support can be provided.

[0115] In an optional implementation, an initial evaluation result of a smart asset is obtained; the initial evaluation result is dynamically adjusted based on a time series prediction algorithm to generate a final evaluation value including a confidence interval; the final evaluation value is associated with a knowledge reasoning result to form an evaluation report including:

[0116] Obtain the initial evaluation result of the smart asset, form a time series based on the initial evaluation result and historical evaluation data, decompose the time series to obtain trend component, periodic component and random component; input the trend component into the long short-term memory network to extract the long-term dependency feature vector, input the periodic component into the Prophet model to extract the seasonal feature vector, and input the random component into the autoregressive model to extract the volatility feature vector; construct a Stacking model, input the long-term dependency feature vector, the seasonal feature vector and the volatility feature vector into the Stacking model to obtain a fusion feature vector;

[0117] Inputting the fused feature vector into the correction model to dynamically correct the initial evaluation result to obtain a corrected evaluation value, using the quantile regression method to calculate the upper and lower limits of the confidence interval of the corrected evaluation value, and obtaining a dynamic evaluation result including a confidence interval;

[0118] In a relational database, based on the dynamic evaluation results, the technical dependent nodes of the evaluation object are screened to build an association network, the transmission coefficient between the nodes in the network is calculated to obtain a transmission feature vector, and the market linkage feature vector is obtained according to the correlation coefficient between the market index and the transmission feature vector; the transmission feature vector, the market linkage feature vector and the dynamic evaluation results are input into a principal component analysis model, and features with a value contribution greater than 0.7 are extracted as core value element vectors, features with a volatility greater than 0.3 are extracted as risk element vectors, and features with a time series autocorrelation coefficient greater than 0.6 are extracted as trend element vectors;

[0119] The main framework of the evaluation report is constructed based on the core value element vector, and the risk element vector and the trend element vector are mapped to corresponding chapters; the newly added data is input into the value evaluation model to update the dynamic evaluation results, and the corresponding chapters in the evaluation report are dynamically updated according to the updated dynamic evaluation results; wherein the newly added data includes updated data of the conduction characteristic vector and the market linkage characteristic vector.

[0120] Obtain the initial evaluation results of smart assets, such as obtaining a preliminary value assessment through expert evaluation, market comparison or cost method, etc. For example, the initial evaluation value of a smart recommendation algorithm is 1 million yuan.

[0121] Construct a time series based on the initial evaluation results and historical evaluation data. Assume that the evaluation values ​​of the algorithm in the past three years were 800,000, 900,000, and 950,000 respectively. Combined with the initial evaluation value, a time series is formed: [80, 90, 95, 100].

[0122] Decompose the time series. Use time series decomposition techniques, such as the moving average method or the STL decomposition method, to decompose the time series into trend components, period components, and random components. Assume that the decomposition results are as follows: the trend component is [75, 85, 95, 105], the period component is [2, 3, 0, -5], and the random component is [3, 2, 0, 0].

[0123] Extract feature vectors. Input the trend component into the long short-term memory network (LSTM) model to extract the long-term dependency feature vector. Input the periodic component into the Prophet model to extract the seasonal feature vector. For example, the seasonal feature vector output by the Prophet model is [0.1, 0.15, 0.05, -0.1]. Input the random component into the autoregressive (AR) model to extract the volatility feature vector. For example, the volatility feature vector output by the AR model is [0.05, 0.03, 0, 0].

[0124] Build a Stacking model and obtain a fused feature vector. Input the long-term dependency feature vector output by the LSTM model, the seasonal feature vector output by the Prophet model, and the volatility feature vector output by the AR model into the Stacking model to obtain a fused feature vector.

[0125] Correct the initial evaluation result and calculate the confidence interval. Input the fused feature vector into the correction model (such as a linear regression model) to dynamically correct the initial evaluation result. Assume that the output of the correction model is 1.02 million. Use the quantile regression method to calculate the upper and lower limits of the confidence interval of the corrected evaluation value. For example, the 95% confidence interval is calculated to be [0.98 million, 1.06 million]. The final dynamic evaluation result including the confidence interval is 1.02 million, and the confidence interval is [0.98 million, 1.06 million].

[0126] Build an association network and calculate the eigenvector. In a relational database, filter the associated nodes of the evaluation object based on the dynamic evaluation results and build an association network. For example, filter the technology-dependent nodes related to the intelligent recommendation algorithm, such as data processing technology, machine learning platform, etc. Calculate the transmission coefficient between the nodes in the network to obtain the transmission eigenvector. For example, the transmission eigenvector is calculated to be [0.2, 0.3, 0.5]. The market linkage eigenvector is obtained based on the correlation coefficient between the market index and the transmission eigenvector. For example, the market linkage eigenvector is calculated to be [0.1, 0.2, 0.3].

[0127] Extract core value elements, risk elements, and trend elements. Input the conduction feature vector, market linkage feature vector, and dynamic evaluation results into the principal component analysis (PCA) model. Extract features with a value contribution greater than 0.7 as core value element vectors. For example, the extracted core value element vector is [data processing capability, model accuracy]. Extract features with a volatility greater than 0.3 as risk element vectors. For example, the extracted risk element vector is [market competition]. Extract features with a time series autocorrelation coefficient greater than 0.6 as trend element vectors. For example, the extracted trend element vector is [technology update speed].

[0128] Generate and dynamically update the evaluation report. Construct the main framework of the evaluation report based on the core value element vector. Map the risk element vector and trend element vector to the corresponding chapters. For example, in the "Risk Assessment" chapter, analyze the market competition risk in detail. In the "Future Trends" chapter, explore the impact of the speed of technology updates on the value of the algorithm. When there is new data, such as the update of the market linkage feature vector, recalculate the dynamic evaluation results, and dynamically update the corresponding chapters in the evaluation report based on the updated results.

[0129] The solution of this application can:

[0130] More accurate evaluation results: By combining time-series prediction algorithms, knowledge reasoning, and market dynamics data, the value of smart assets can be evaluated more accurately, reducing interference from human factors and improving the objectivity and reliability of the evaluation. The evaluation process is more dynamic: The evaluation results can be dynamically adjusted based on new data such as market changes and technological development, making the evaluation results closer to the actual situation and avoiding the lag caused by static evaluation. More comprehensive evaluation reports: The evaluation report not only includes the value evaluation results, but also covers the analysis of risk factors and trend factors, providing more comprehensive information to help decision makers better understand the value and potential risks of smart assets.

[0131] In an optional implementation, the fused feature vector is input into a correction model to dynamically correct the initial evaluation result to obtain a corrected evaluation value, and the upper and lower limits of the confidence interval of the corrected evaluation value are calculated using a quantile regression method, and the dynamic evaluation result including the confidence interval is obtained, including:

[0132] Input the fused feature vector into a feature mapping network to obtain a mapping feature; construct an attention weight calculation module to calculate a feature attention score based on the correlation between the mapping feature and the historical correction effect; and weight the feature attention score and the mapping feature to obtain a correction feature;

[0133] Based on the correction feature, a dynamic correction model is constructed, and the initial evaluation result is input into the dynamic correction model to obtain a correction evaluation value; an exponential sliding average method is used to track the correction amplitude, and the correction step length is adaptively adjusted according to the correction amplitude; a historical record of the correction evaluation value is established for dynamic update of subsequent correction parameters;

[0134] Constructing a quantile regression model group, inputting the modified evaluation value and the modified feature into the quantile regression model group to obtain prediction values ​​of different quantiles; constructing a distributed loss function based on the predicted value, wherein the distributed loss function considers both prediction accuracy and distribution rationality;

[0135] The quantile regression model group is trained by adopting an alternating optimization strategy: first, the distribution parameters are fixed to optimize the quantile prediction model to obtain the regression coefficient of each quantile; then, the regression coefficient is fixed to optimize the distribution parameters to ensure the continuity of the prediction distribution; the prediction error of each quantile is calculated based on the verification data, and the optimal model parameters are selected;

[0136] The modified evaluation value is input into the trained quantile regression model group to calculate the upper limit value of the confidence interval and the lower limit value of the confidence interval; the modified evaluation value is combined with the upper limit value of the confidence interval and the lower limit value of the confidence interval to generate a dynamic evaluation result including the confidence interval.

[0137] A dynamic evaluation method including confidence intervals, the specific implementation of which is as follows:

[0138] First, obtain the initial evaluation results of the object to be evaluated and the corresponding feature vector. For example, to evaluate a student's learning ability, the initial evaluation result may be the midterm exam score. The feature vector may include information such as the student's classroom performance, homework completion, and learning attitude, expressed in numerical form. For example, if the classroom performance is 80 points (out of 100 points), the homework completion is 90 points (out of 100 points), the learning attitude is positive, represented by 1, and negative is 0, then the student's feature vector is [80,90,1].

[0139] Next, the feature vector is input into the feature mapping network for processing. The feature mapping network can be a multi-layer perceptron network, which maps the input feature vector to a new feature space to obtain mapping features. For example, [80, 90, 1] is input into a fully connected network with one hidden layer, 10 hidden layer nodes, and 5 output layer nodes. After linear transformation and nonlinear activation function processing, a 5-dimensional mapping feature is finally obtained, such as [0.2, 0.5, 0.8, 0.1, 0.9].

[0140] Then, an attention weight calculation module is constructed. This module calculates the feature attention score based on the correlation between the mapping feature and the historical correction effect. For example, by analyzing historical data, it is found that students’ learning attitude has a greater impact on the improvement of final exam scores, so the feature dimension corresponding to the learning attitude will occupy a larger proportion in the attention score.

[0141] The feature attention score is weighted with the mapping feature to obtain the modified feature. For example, the corresponding elements of the mapping feature [0.2, 0.5, 0.8, 0.1, 0.9] and the attention score [0.1, 0.1, 0.3, 0.2, 0.3] are multiplied and added to obtain the modified feature 0.55.

[0142] A dynamic correction model is constructed based on the correction feature, and the initial evaluation result is input into the model to obtain a corrected evaluation value. The dynamic correction model can be a linear regression model, which uses the correction feature as input to correct the initial evaluation result. For example, assuming that a student's midterm exam score is 70 points, after calculation by the dynamic correction model, the corrected evaluation value is 75 points.

[0143] The exponential moving average method is used to track the correction amplitude, and the correction step size is adaptively adjusted according to the correction amplitude. For example, if the correction amplitude is large for several consecutive times, the correction step size is reduced to avoid over-correction; if the correction amplitude is small, the correction step size is increased to speed up the correction speed.

[0144] Establish a historical record of the modified evaluation value for dynamic updating of subsequent modified parameters. For example, save the modified evaluation value and its corresponding modified features calculated each time for subsequent model training and updating.

[0145] Construct a quantile regression model group, and input the modified evaluation value and modified features into the model group to obtain the prediction values ​​of different quantiles. For example, if the quantiles are set to 0.1, 0.5, and 0.9, three prediction values ​​can be obtained, corresponding to the 10% quantile, 50% quantile, and 90% quantile of the modified evaluation value. Assuming the modified evaluation value is 75 points, the corresponding three prediction values ​​are 70 points, 75 points, and 80 points.

[0146] The alternating optimization strategy is used to train the quantile regression model group. First, the distribution parameters are fixed to optimize the quantile prediction model and obtain the regression coefficient of each quantile. Then, the regression coefficient is fixed to optimize the distribution parameters to ensure the continuity of the predicted distribution. The prediction error of each quantile is calculated based on the validation data, and the optimal model parameters are selected.

[0147] Input the corrected evaluation value into the trained quantile regression model group and calculate the upper and lower limits of the confidence interval. For example, based on the three predicted values ​​of 70 points, 75 points, and 80 points calculated previously, 70 points can be used as the lower limit of the confidence interval and 80 points can be used as the upper limit of the confidence interval.

[0148] The corrected evaluation value is combined with the upper and lower limits of the confidence interval to generate a dynamic evaluation result including the confidence interval. For example, the final evaluation result can be expressed as 75 points [70, 80], which means that the corrected evaluation value is 75 points and the confidence interval is [70, 80].

[0149] The solution of this application can:

[0150] The evaluation results are more accurate: Through the dynamic correction model and feature attention mechanism, the true level of the object can be more accurately evaluated, and the accuracy of the evaluation results can be improved. The evaluation results are more reliable: by calculating the confidence interval through the quantile regression method, the uncertainty of the evaluation results can be quantified, making the evaluation results more reliable and more valuable for reference. The model has adaptive capabilities: through the exponential sliding average method and dynamic parameter update mechanism, the model can adaptively adjust the correction step size and model parameters according to the changes in the data, improving the generalization ability and robustness of the model.

[0151] Figure 2 FIG. 1 is a schematic diagram of the structure of the data-driven cross-domain intelligent asset knowledge reasoning and value assessment system according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0152] The first unit is used to obtain smart asset data in multiple fields, wherein the smart asset data includes asset attribute information, operation status information, historical transaction information and market evaluation information; perform feature extraction on the smart asset data, and construct an asset feature vector based on a deep learning model; map the asset feature vector to a preset knowledge graph ontology structure, and establish an association relationship between smart asset nodes; perform weight calculation on the association relationship based on a graph neural network to form a cross-domain smart asset knowledge graph;

[0153] The second unit is used to receive the smart asset evaluation target input by the user, parse the semantic features of the evaluation target based on the pre-trained language model; perform path retrieval in the cross-domain smart asset knowledge graph according to the semantic features to obtain a set of asset nodes related to the evaluation target; apply the attention mechanism to the asset node set to perform correlation calculation and determine the importance weight; perform multi-hop reasoning on the nodes whose importance weight is greater than the preset weight threshold based on the graph reasoning algorithm to form an asset knowledge reasoning path; match the asset knowledge reasoning path with a preset reasoning rule template to generate a knowledge reasoning result;

[0154] The third unit is used to extract the value influencing factors of the intelligent assets according to the knowledge reasoning results, and input the value influencing factors into a pre-trained value assessment model; the value assessment model is trained based on historical transaction data and market assessment data, and the assessment parameters are continuously optimized by a reinforcement learning method; the initial assessment results of the intelligent assets are obtained; the initial assessment results are dynamically adjusted based on a time series prediction algorithm to generate a final assessment value including a confidence interval; the final assessment value is associated with the knowledge reasoning results to form an assessment report.

[0155] According to a third aspect of the embodiments of the present invention,

[0156] An electronic device is provided, comprising:

[0157] processor;

[0158] a memory for storing processor-executable instructions;

[0159] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0160] According to a fourth aspect of the embodiments of the present invention,

[0161] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0162] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0163] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. Data-driven cross-domain intelligent asset knowledge reasoning and value assessment method, characterized by: include: Acquire smart asset data in multiple fields, the smart asset data including asset attribute information, operation status information, historical transaction information and market evaluation information; Extracting features from the smart asset data and constructing an asset feature vector based on a deep learning model; Mapping the asset feature vector to a preset knowledge graph ontology structure to establish associations between smart asset nodes; performing weight calculations on the associations based on a graph neural network to form a cross-domain smart asset knowledge graph; Receive the smart asset evaluation target input by the user, and parse the semantic features of the evaluation target based on the pre-trained language model; perform path retrieval in the cross-domain smart asset knowledge graph according to the semantic features to obtain a set of asset nodes related to the evaluation target; apply the attention mechanism to the asset node set to perform correlation calculation and determine the importance weight; Based on the graph reasoning algorithm, multi-hop reasoning is performed on nodes whose importance weight is greater than the preset weight threshold to form an asset knowledge reasoning path; Matching the asset knowledge reasoning path with a preset reasoning rule template to generate a knowledge reasoning result; Extracting the value impact factor of the smart asset according to the knowledge reasoning result, and inputting the value impact factor into a pre-trained value assessment model; the value assessment model is trained based on historical transaction data and market assessment data, and the assessment parameters are continuously optimized using a reinforcement learning method; Obtaining initial evaluation results of smart assets; dynamically adjusting the initial evaluation results based on a time series prediction algorithm to generate a final evaluation value including a confidence interval; associating the final evaluation value with knowledge reasoning results to form an evaluation report; Receive an evaluation target input by a user, and normalize the evaluation target based on an evaluation target semantic template, wherein the evaluation target semantic template includes evaluation object information, evaluation dimension information, and evaluation constraint condition information; The normalized evaluation target is input into a pre-trained language model, and the pre-trained language model performs semantic analysis on the evaluation target through an expanded smart asset domain vocabulary, and adopts a multi-head cross-attention mechanism to respectively construct an evaluation factor attention head, a constraint condition attention head, and a context association attention head, and the multi-head cross-attention mechanism outputs word-level semantic features, phrase-level semantic features, and sentence-level semantic features; Decomposing the word-level semantic features, the phrase-level semantic features, and the sentence-level semantic features into multiple query atoms, matching the query atoms with a preset query template library based on semantic similarity calculation, and generating a query statement including node matching conditions, path constraint conditions, multi-hop relationship conditions, and attribute filtering conditions according to the matching results; Based on the query statement, a bidirectional extended search is initiated from both the evaluation target node and the candidate asset node in the cross-domain smart asset knowledge graph, the priority of the search direction is determined by the node connectivity, and the search pruning is performed using the semantic relevance; the retrieval task is decomposed into multiple parallel subtasks, and an asynchronous update mechanism is used to maintain the global optimal path; the retrieved path is comprehensively scored, and the comprehensive score is calculated based on the semantic relevance, path credibility, path novelty, and path completeness of the path to obtain a path quality score; The correlation score between the node and the evaluation target is calculated according to the path quality score, and the correlation score is obtained by weighted calculation of direct semantic correlation, indirect association strength, temporal correlation, and business correlation; based on the correlation score, the candidate asset nodes are screened and dynamically updated to obtain a set of asset nodes related to the evaluation target.

2. The method according to claim 1, characterized in that Mapping the asset feature vector to a preset knowledge graph ontology structure to establish associations between smart asset nodes; and weighting the associations based on a graph neural network to form a cross-domain smart asset knowledge graph, including: Mapping the asset feature vector to a preset knowledge graph ontology structure through a feature mapping model, wherein the feature mapping model calculates the semantic distance between the asset feature vector and the ontology concept vector based on cosine similarity, assigns weight values ​​to different dimensions of the asset feature vector through an attention mechanism, and selects the ontology concept with the highest semantic similarity as the mapping target node; A multi-dimensional association analysis model is established for the mapping target node, wherein the multi-dimensional association analysis model constructs association relationships between nodes based on the attribute information, historical data information, spatial location information, and business process information of the smart asset, and the association relationships are formed into an initial knowledge graph; the nodes of the initial knowledge graph contain characteristic attributes of the smart asset, and the edges between the nodes represent the association types between the assets; Inputting the initial knowledge graph into a graph neural network for weight calculation, the graph neural network comprises multiple graph attention layers, each graph attention layer is provided with multiple attention heads for learning the association pattern between nodes; calculating the hidden layer representation of the node based on the output of the attention head, and calculating the feature similarity, structural similarity and temporal similarity between the node pairs according to the hidden layer representation; The feature similarity, the structural similarity and the temporal similarity are input into a multi-task learning model, which simultaneously optimizes the node classification task and the edge prediction task, and uses a graph-structure-based attention mechanism to adaptively adjust the weight coefficients of different types of associations, and finally outputs the association weights between smart asset nodes to form a cross-domain smart asset knowledge graph containing weight information.

3. The method according to claim 1, characterized in that Based on the query statement, a bidirectional extended search is initiated from both the evaluation target node and the candidate asset node in the cross-domain smart asset knowledge graph, the priority of the search direction is determined by the node connectivity, and the search is pruned using the semantic relevance; the retrieval task is decomposed into multiple parallel subtasks, and an asynchronous update mechanism is used to maintain the global optimal path; The retrieved paths are comprehensively scored. The comprehensive score is calculated based on the path semantic relevance, path credibility, path novelty, and path completeness to obtain a path quality score including: Extracting an evaluation target node and a candidate asset node from the query statement; calculating node connectivity for the evaluation target node and the candidate asset node respectively, and generating a node importance vector based on the node connectivity; inputting the query statement into a pre-trained language model to obtain a query semantic vector, and setting a search pruning threshold based on the query semantic vector; Taking the evaluation target node and the candidate asset node as starting points, a bidirectional expansion search is simultaneously initiated in the cross-domain smart asset knowledge graph, wherein the bidirectional expansion search determines the priority of the search direction based on the node importance vector; during the search process, the similarity between the path to be expanded and the query semantic vector is calculated, and when the similarity is lower than the search pruning threshold, the expansion in this direction is stopped; The search space is divided into multiple sub-areas, and independent parallel sub-tasks are constructed in each of the sub-areas to perform the bidirectional extended search; each parallel sub-task maintains its own local optimal path, and the local optimal paths are aggregated to obtain a global optimal path through an asynchronous update mechanism; the asynchronous update mechanism determines the update timing based on the path quality to avoid the interference of low-quality paths on the global optimal path; The retrieved paths are scored in multiple dimensions: the similarity between the path node sequence and the query semantic vector is calculated to obtain the path semantic relevance, the path credibility is calculated based on the reliability of the path relationship and the authority of the knowledge source, the path novelty is calculated based on the rarity of the path pattern in the knowledge graph, and the path completeness is calculated based on the key node coverage and semantic completeness; an adaptive weight method is used to weightedly fuse the path semantic relevance, the path credibility, the path novelty, and the path completeness to obtain a path quality score.

4. The method according to claim 1, characterized in that: Applying an attention mechanism to the asset node set to perform relevance calculation and determine importance weights; Based on the graph reasoning algorithm, multi-hop reasoning is performed on nodes whose importance weight is greater than the preset weight threshold to form an asset knowledge reasoning path; Matching the asset knowledge reasoning path with a preset reasoning rule template to generate knowledge reasoning results includes: Perform multi-level attention calculation on the asset node set, calculate the structural importance score based on the node degree centrality and the eigenvector centrality, calculate the semantic importance score based on the semantic association between the node attribute and the evaluation target, and calculate the temporal importance score based on the time decay factor and the state transition frequency; perform weighted combination of the structural importance score, the semantic importance score, and the temporal importance score to obtain the node comprehensive importance weight; Based on the comprehensive importance weight of the node, a seed node with a weight greater than a preset weight threshold is selected, a directed reasoning graph is constructed in the cross-domain smart asset knowledge graph, and a maximum reasoning depth is determined according to the comprehensive importance weight of the seed node; path extension is performed on the directed reasoning graph, and the extension direction is constrained by the semantic association between nodes; a reliability score is calculated for the extended path based on the comprehensive importance weight of the path node, and a coherence score is calculated based on the semantic continuity of the path, and a path with both the reliability score and the coherence score higher than the preset score threshold is selected to form an asset knowledge reasoning path; The paths in the asset knowledge reasoning path are structurally matched and semantically matched with the preset reasoning rule templates, and supporting evidence is extracted based on the matching results. The supporting evidence includes key nodes and relationships on the path; the supporting evidence is verified for rule consistency and logical integrity, and based on the verification results, trusted paths with scores higher than the preset score threshold are screened and conflicts are resolved to generate the final knowledge reasoning results.

5. The method according to claim 1, characterized in that Obtaining an initial evaluation result of the smart asset; dynamically adjusting the initial evaluation result based on a time series prediction algorithm to generate a final evaluation value including a confidence interval; The final evaluation value is associated with the knowledge reasoning result to form an evaluation report including: Obtain the initial evaluation result of the smart asset, form a time series based on the initial evaluation result and historical evaluation data, decompose the time series to obtain trend component, periodic component and random component; input the trend component into the long short-term memory network to extract the long-term dependency feature vector, input the periodic component into the Prophet model to extract the seasonal feature vector, and input the random component into the autoregressive model to extract the volatility feature vector; construct a Stacking model, input the long-term dependency feature vector, the seasonal feature vector and the volatility feature vector into the Stacking model to obtain a fusion feature vector; Inputting the fused feature vector into the correction model to dynamically correct the initial evaluation result to obtain a corrected evaluation value, using the quantile regression method to calculate the upper and lower limits of the confidence interval of the corrected evaluation value, and obtaining a dynamic evaluation result including a confidence interval; In a relational database, based on the dynamic evaluation results, the technical dependent nodes of the evaluation object are screened to build an association network, the transmission coefficient between the nodes in the network is calculated to obtain a transmission feature vector, and the market linkage feature vector is obtained according to the correlation coefficient between the market index and the transmission feature vector; the transmission feature vector, the market linkage feature vector and the dynamic evaluation results are input into a principal component analysis model, and features with a value contribution greater than 0.7 are extracted as core value element vectors, features with a volatility greater than 0.3 are extracted as risk element vectors, and features with a time series autocorrelation coefficient greater than 0.6 are extracted as trend element vectors; The main framework of the evaluation report is constructed based on the core value element vector, and the risk element vector and the trend element vector are mapped to corresponding chapters; the newly added data is input into the value evaluation model to update the dynamic evaluation results, and the corresponding chapters in the evaluation report are dynamically updated according to the updated dynamic evaluation results; wherein the newly added data includes updated data of the conduction characteristic vector and the market linkage characteristic vector.

6. The method according to claim 5, characterized in that The fused feature vector is input into the correction model to dynamically correct the initial evaluation result to obtain a corrected evaluation value, and the upper and lower limits of the confidence interval of the corrected evaluation value are calculated using the quantile regression method. The dynamic evaluation result including the confidence interval is obtained, including: Input the fused feature vector into a feature mapping network to obtain a mapping feature; construct an attention weight calculation module to calculate a feature attention score based on the correlation between the mapping feature and the historical correction effect; and weight the feature attention score and the mapping feature to obtain a correction feature; Based on the correction feature, a dynamic correction model is constructed, and the initial evaluation result is input into the dynamic correction model to obtain a correction evaluation value; an exponential sliding average method is used to track the correction amplitude, and the correction step length is adaptively adjusted according to the correction amplitude; a historical record of the correction evaluation value is established for dynamic update of subsequent correction parameters; Constructing a quantile regression model group, inputting the modified evaluation value and the modified feature into the quantile regression model group to obtain prediction values ​​of different quantiles; constructing a distributed loss function based on the predicted value, wherein the distributed loss function considers both prediction accuracy and distribution rationality; The quantile regression model group is trained by adopting an alternating optimization strategy: first, the distribution parameters are fixed to optimize the quantile prediction model to obtain the regression coefficient of each quantile; then, the regression coefficient is fixed to optimize the distribution parameters to ensure the continuity of the prediction distribution; the prediction error of each quantile is calculated based on the verification data, and the optimal model parameters are selected; The modified evaluation value is input into the trained quantile regression model group to calculate the upper limit value of the confidence interval and the lower limit value of the confidence interval; the modified evaluation value is combined with the upper limit value of the confidence interval and the lower limit value of the confidence interval to generate a dynamic evaluation result including the confidence interval.

7. A data-driven cross-domain intelligent asset knowledge reasoning and value assessment system, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to obtain smart asset data in multiple fields, wherein the smart asset data includes asset attribute information, operation status information, historical transaction information and market evaluation information; Extracting features from the smart asset data and constructing an asset feature vector based on a deep learning model; Mapping the asset feature vector to a preset knowledge graph ontology structure to establish associations between smart asset nodes; performing weight calculations on the associations based on a graph neural network to form a cross-domain smart asset knowledge graph; The second unit is used to receive the smart asset evaluation target input by the user, parse the semantic features of the evaluation target based on the pre-trained language model; perform path retrieval in the cross-domain smart asset knowledge graph according to the semantic features to obtain a set of asset nodes related to the evaluation target; apply the attention mechanism to the asset node set to perform correlation calculation and determine the importance weight; Based on the graph reasoning algorithm, multi-hop reasoning is performed on nodes whose importance weight is greater than the preset weight threshold to form an asset knowledge reasoning path; Matching the asset knowledge reasoning path with a preset reasoning rule template to generate a knowledge reasoning result; The third unit is used to extract the value impact factor of the smart asset according to the knowledge reasoning result, and input the value impact factor into a pre-trained value evaluation model; the value evaluation model is trained based on historical transaction data and market evaluation data, and the evaluation parameters are continuously optimized using a reinforcement learning method; Obtaining initial evaluation results of smart assets; dynamically adjusting the initial evaluation results based on a time series prediction algorithm to generate a final evaluation value including a confidence interval; associating the final evaluation value with knowledge reasoning results to form an evaluation report.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images