Large model science and technology problem generation method fusing external knowledge

Abstract

The invention provides a large model scientific and technological problem generation method fusing external knowledge, which comprises the following steps: dynamically solving an optimal knowledge fusion strength parameter in a current environment by adopting a gradient descent algorithm through designing and optimizing an objective function according to association representation and difference representation between knowledge units, the knowledge unit balancing module is used for balancing correlation and complementarity among knowledge units; on the basis of real-time updating of knowledge fusion strength parameters, heuristic search is carried out in the knowledge unit association map by utilizing an ant colony optimization algorithm, and candidate fusion knowledge units which achieve target correlation and target complementarity with the current knowledge unit are selected by combining semantic similarity, association degree and difference among the knowledge units; a knowledge mapping ontology is used for guiding fusion of cross-domain knowledge, an ontology inference engine based on description logic is adopted to infer a semantic consistency relationship between fused knowledge units, and potential semantic conflicts are identified and eliminated through a consistency check algorithm.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for generating large-model scientific and technological problems integrating external knowledge. Background Art

[0002] In the process of cross-domain knowledge fusion, there are a large number of knowledge units in multiple scientific and technological fields in external knowledge. External knowledge units in different scientific and technological fields can be integrated across domains to generate innovative technical solutions. The dynamic changes of external knowledge units need to be perceived and analyzed in real time to determine the correlation and difference between knowledge units and dynamically calculate the optimal knowledge fusion strength. However, knowledge in different fields differs in terms, concepts and methods, and there are inconsistencies and conflicts in the fusion process of cross-domain knowledge. These differences will affect the consistency and conflict-free nature of the fusion process. How to calculate the optimal knowledge fusion strength in real time in a dynamically changing knowledge unit interaction environment, and intelligently match and call appropriate knowledge units for cross-domain fusion according to the strength, and establish an effective mapping mechanism to eliminate the impact of knowledge differences in different fields on the fusion process, and ensure that the generated scientific and technological problems are both forward-looking and practical, is a technical problem that needs to be solved urgently. This requires in-depth analysis of the dynamic change laws of external knowledge units and the characteristics of knowledge unit correlation differences, research on the dynamic adjustment mechanism of knowledge fusion strength and cross-domain knowledge mapping mechanism, and exploration of a new cross-domain knowledge fusion method that takes into account the consistency and conflict-free nature of knowledge fusion and the forward-looking and practical nature of generating scientific and technological problems. Summary of the invention

[0003] The present invention provides a method for generating a large-scale scientific and technological problem by integrating external knowledge, which mainly includes:

[0004] Extract knowledge units from an external knowledge base, calculate the similarity between knowledge units using a semantic matching algorithm, generate an association strength matrix, analyze the dynamic changes of the association strength matrix in real time, perceive the changes in the association between knowledge units, determine the difference, analyze the attribute characteristics of knowledge units that have changed significantly, and adjust and update the association strength matrix;

[0005] Using graph neural networks for deep embedding learning, we can identify the association strength and difference between knowledge units, extract the implicit association patterns and subtle difference features between knowledge units by combining the association strength and difference information between knowledge units, and encode the target association representation and target difference representation in a low-dimensional dense vector space.

[0006] According to the association representation and difference representation between knowledge units, the optimal knowledge fusion intensity parameter under the current environment is dynamically solved by designing the optimization objective function and using the gradient descent algorithm to balance the correlation and complementarity between knowledge units.

[0007] Based on the real-time update of knowledge fusion strength parameters, the ant colony optimization algorithm is used to perform heuristic search in the knowledge unit association graph, and the candidate fusion knowledge units that achieve target relevance and target complementarity with the current knowledge units are selected by combining the semantic similarity, relevance and difference between the knowledge units.

[0008] Based on association representation and difference representation, potential contradictions, conflicts and incompatibilities between knowledge units are identified, and technical problems caused by contradictions are inferred. A combination of rule-based reasoning and case-based reasoning is used to match the characteristics of current knowledge units according to the historical technical problem case library to generate technical problem hypotheses.

[0009] By introducing technical problem hypotheses, identifying key terms and concepts related to the problem, highlighting the semantic associations in the problem space, and using multi-task learning and contrastive learning algorithms to build a domain ontology, a mapping relationship between technical problem hypotheses and knowledge units is established in the domain ontology, and the deep semantic associations in the problem space are identified and strengthened, thereby building a cross-domain problem-oriented knowledge mapping ontology;

[0010] The knowledge mapping ontology is used to guide the fusion of cross-domain knowledge. The ontology reasoning engine based on description logic is used to infer the semantic consistency relationship between fused knowledge units. The consistency checking algorithm is used to identify and eliminate potential semantic conflicts.

[0011] Conduct quality assessment on the fused knowledge units, design a comprehensive evaluation index system, evaluate the relevance, complementarity, consistency and non-conflict of fused knowledge, use the analytic hierarchy process to calculate the index weights of fused knowledge, and use the weighted summation method to obtain the knowledge fusion quality score. According to the evaluation results, iteratively optimize the fusion process, adjust the knowledge fusion intensity and mapping strategy, and finally generate cross-domain fusion knowledge.

[0012] Based on the technical problem hypothesis generated by the fusion of knowledge, an intelligent optimization algorithm is used to solve the solution. The objective function of technical feasibility and economic benefit is defined as the optimization criterion. The candidate solutions are generated and evaluated through algorithm iteration until convergence or the stopping condition is met. The candidate solutions are compared through simulation and analysis, and the target solution is selected to form a complete description of the technical problem and solution.

[0013] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0014] The present invention discloses a method for generating large-scale scientific and technological problems by integrating external knowledge. The present invention extracts knowledge units from an external knowledge base, calculates similarity through semantic matching, and generates an association strength matrix. A graph neural network is used for deep embedding learning to identify the association strength and difference between knowledge units and extract implicit association patterns. According to the association representation and difference representation, the knowledge fusion strength parameters are dynamically optimized to balance relevance and complementarity. An ant colony algorithm is used to select candidate fusion knowledge units and identify potential contradictions and conflicts. A cross-domain problem-oriented knowledge mapping ontology is constructed based on technical problem assumptions to guide knowledge fusion and eliminate semantic conflicts. The fused knowledge is quality evaluated and iteratively optimized to generate high-quality cross-domain fusion knowledge. Finally, based on the technical problem assumptions generated by the fused knowledge, an intelligent optimization algorithm is used to solve the solution. The present invention realizes the efficient fusion and innovative application of cross-domain knowledge, and provides new methods and ideas for solving complex technical problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of a method for generating large-model scientific and technological problems by integrating external knowledge according to the present invention.

[0016] Figure 2 A schematic diagram of a large-model scientific and technological problem generation method integrating external knowledge according to the present invention.

[0017] Figure 3 This is another schematic diagram of a large-model scientific and technological problem generation method that integrates external knowledge according to the present invention. DETAILED DESCRIPTION

[0018] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than 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.

[0019] like Figure 1-3 In this embodiment, a large-model scientific and technological problem generation method integrating external knowledge may specifically include:

[0020] Step S101, extract knowledge units from the external knowledge base, use a semantic matching algorithm to calculate the similarity between knowledge units, generate an association strength matrix, analyze the dynamic changes of the association strength matrix in real time, perceive the changes in the association between knowledge units, and determine the difference, perform attribute feature analysis on knowledge units that have changed significantly, and adjust and update the association strength matrix.

[0021] The word segmentation process is performed according to the content of the knowledge unit to obtain a feature vector, and the weighted average algorithm is used to calculate the semantic similarity between the knowledge units to construct an initial association strength matrix; a time window is set for the initial association strength matrix for dynamic sampling, and if the difference in association strength in adjacent time windows exceeds a preset threshold, the position of the knowledge unit pair is marked as 1 in the difference matrix; the keyword frequency vector, part-of-speech distribution vector, and semantic dependency relationship vector are extracted from the corresponding knowledge unit marked as 1 in the difference matrix to form a feature triple, and a normalization process is performed to obtain an attribute similarity matrix; a weighted fusion is performed based on the initial association strength matrix and the attribute similarity matrix, and the fused matrix is ​​decomposed and reconstructed by a non-negative matrix decomposition algorithm to obtain an updated knowledge unit association network.

[0022] Specifically, the knowledge unit content is segmented according to the Chinese word segmenter, the word vector is extracted through the feature vector generator, the semantic vector of the knowledge unit is synthesized based on the weighted average algorithm, and the semantic similarity between knowledge units is calculated using the cosine similarity calculator to construct the initial association strength matrix. For the initial association strength matrix, the time window is set to t time units for dynamic sampling, and the difference in association strength between each pair of knowledge units in adjacent time windows is calculated. If the difference exceeds the preset threshold θ, the position of the knowledge unit pair is marked as 1 in the difference matrix, otherwise it is marked as 0. For the knowledge unit marked as 1 in the difference matrix, its keyword frequency vector, part-of-speech distribution vector, and semantic dependency vector are extracted to form a feature triple, and the standardized feature vector is obtained by normalization. The Euclidean distance is calculated based on the standardized feature vector to obtain the attribute difference matrix between knowledge units, and the difference is mapped to the interval [0,1] through the sigmoid function to obtain the attribute similarity matrix. The initial association strength matrix and the attribute similarity matrix are merged using the weighted fusion method, and the fusion matrix is ​​decomposed and reconstructed based on the non-negative matrix decomposition algorithm to obtain the updated knowledge unit association network. For the reconstructed association network, the PageRank algorithm is used to calculate the importance of knowledge units, and the network is pruned in combination with the association strength of knowledge units, retaining the knowledge unit pairs with association strength greater than the threshold δ and the top k importance rankings. The semantic vector of the knowledge unit is obtained by weighted average of the word vectors. For a knowledge unit containing n words, the vector dimension of each word is d, and the weight is determined according to the word frequency and the importance of the part of speech. For example, the weight of nouns is 0.4, the value of verbs is 0.3, the value of adjectives is 0.2, and the value of other parts of speech is 0.1, and finally a d-dimensional semantic vector is obtained. Taking the knowledge unit "machine learning algorithm performance evaluation" as an example, after word segmentation, four words "machine learning / algorithm / performance / evaluation" are obtained, and the corresponding parts of speech are noun, noun, noun, and verb, respectively. The 300-dimensional vector of each word is obtained through the pre-trained word vector model, and the weighted average is calculated in combination with weights 0.4, 0.4, 0.4, and 0.3. When the association strength changes during dynamic sampling within the time window t, the window size is set to 30 days and the sliding step is set to 1 day. If the difference in the association strength between two knowledge units in adjacent time windows is greater than the threshold of 0.3, the knowledge unit pair is marked as having changed significantly. When extracting features for the marked knowledge units, the dimension of the keyword frequency vector is the same as the size of the vocabulary. If a knowledge unit contains the i-th word in the vocabulary, the dimension is the number of times the word appears in the unit. The dimension of the part-of-speech distribution vector is the number of part-of-speech types, and the j-th dimension is the proportion of the number of times the part-of-speech appears in the unit. The semantic dependency vector represents the syntactic relationship between words. The difference of feature vectors is calculated by Euclidean distance. For two m-dimensional feature vectors a and b, the difference is sqrt(Σ(ai-bi)2), where i ranges from 1 to m. The calculated difference is mapped to the interval [0,1] by the sigmo id function f(x)=1 / (1+e^(-x)) to obtain the attribute similarity.

[0023] Step S102, using graph neural network for deep embedding learning, identifying the association strength and difference between knowledge units, combining the association strength and difference information between knowledge units, extracting implicit association patterns and subtle difference features between knowledge units, and encoding target association representation and target difference representation in a low-dimensional dense vector space.

[0024] An adjacency matrix is ​​constructed according to the cosine similarity between knowledge units, and the features of directly connected nodes are extracted from the adjacency matrix. The features are transferred and aggregated between nodes through an information communicator to obtain an initial node representation vector. For the initial node representation vector, a multi-head attention calculator is used to obtain the attention score of the knowledge unit in the feature subspace, and the association strength value and the weak association label matrix are obtained through normalization processing. A knowledge unit feature matrix is ​​constructed according to the association strength value and the weak association label matrix, and the feature matrix is ​​compressed by an autoencoder to obtain a compressed representation vector. A cosine distance calculator is used to calculate the similarity value between the compressed representation vectors, and a knowledge graph is constructed according to the similarity value. The knowledge units are grouped by a clustering algorithm. If the ratio of the average distance within a group to the minimum distance between groups is less than a preset threshold, the mapping position of the knowledge unit in the low-dimensional dense space is updated.

[0025] Specifically, the graph structure of the knowledge unit is constructed according to the graph convolution algorithm, and the adjacency matrix is ​​constructed by the cosine similarity calculator. It is set that there is a connection edge between the knowledge unit pairs with similarity greater than the preset threshold. For each knowledge unit node, the features of the directly connected nodes are extracted, and the feature information is transmitted and aggregated between nodes through the information propagator to obtain the initial node representation vector. The multi-head attention calculator is used to perform weighted aggregation on the initial node representation vector, and the attention score of the knowledge unit in different feature subspaces is calculated. The association strength value is obtained through normalization processing, and a weak association labeling matrix is ​​generated for the knowledge unit pairs with association strength values ​​lower than the preset threshold. The Euclidean distance is calculated according to the initial node representation vector to obtain the knowledge unit difference matrix, and the knowledge unit feature matrix is ​​constructed by combining the association strength value and the weak association labeling matrix. The feature matrix is ​​compressed by the encoding layer in the autoencoder. The compressed feature matrix is ​​decoded and reconstructed, and the autoencoder parameters are optimized by the reconstruction error, and the knowledge unit compressed representation vector is obtained in the encoding layer. The cosine distance calculator is used to calculate the similarity between the compressed representation vectors, and the knowledge graph is constructed according to the similarity. The node sequence is sampled in the graph through the random walk algorithm, and the target association representation vector and target difference representation vector are generated by the graph embedding encoder. For the target representation vector, a clustering algorithm is used to group the knowledge units. The compactness of the knowledge unit cluster is judged by the ratio of the average distance within the group to the minimum distance between the groups, and the mapping position of the knowledge unit in the low-dimensional dense space is updated. In the process of graph structure construction, the degree of association between knowledge units is calculated by cosine similarity. Taking the two knowledge units of "deep learning foundation" and "neural network structure" as examples, the cosine similarity calculated after extracting the keyword vector is 0.85, which is higher than the preset threshold of 0.6, so edge connections are constructed in the graph. The similarity of all knowledge units is calculated pairwise to obtain the adjacency matrix. The value of 1 in the matrix indicates that there is a connection edge between the corresponding knowledge unit pair, and the value of 0 indicates that there is no connection. In the attention calculation stage, the node representation vector is mapped to 8 different feature subspaces, each with a dimension of 64. The vector is linearly transformed to obtain the query vector, key vector and value vector. The attention score is calculated by the dot product of the query vector and the key vector. The attention score of the knowledge unit "machine learning algorithm" for "data preprocessing" is 0.72, and the score for "model evaluation" is 0.68, which reflects the strength of association between knowledge units. The knowledge unit pairs with an association strength value lower than 0.3 are marked as 1 in the corresponding position of the weak association labeling matrix. The encoding layer of the autoencoder adopts a three-layer fully connected network structure. The input layer dimension is the same as the feature matrix dimension, which is 1024, and the intermediate hidden layer dimensions are 512, 256, and 128 respectively. The feature matrix is ​​compressed and encoded to obtain a 128-dimensional compressed representation vector, and then the original feature matrix is ​​reconstructed through a symmetrical decoding layer. The reconstruction error is calculated using the mean square error, and the optimization is stopped when the error is less than 0.01. The compressed representation vector can retain the key feature information of the knowledge unit.When performing random walks on the knowledge graph, a walk sequence of length 10 is generated from each node, and the next node is selected based on the weight of the edge each time. For the node "deep learning", the selection probabilities of its adjacent nodes "neural network", "optimization algorithm", and "loss function" are 0.4, 0.35, and 0.25, respectively. After collecting all walk sequences, the low-dimensional dense representation of the nodes is learned by the Skip-gram algorithm. In the knowledge unit clustering, the K-means algorithm is used to divide the knowledge units into 5 clusters, and the ratio of the average Euclidean distance between samples in the cluster to the minimum distance between the center points of the clusters is calculated. The smaller the ratio, the better the clustering effect. For the knowledge unit cluster related to "machine learning", its internal average distance is 0.25, the distance to the nearest cluster is 0.8, and the ratio is 0.31, indicating that the knowledge units of this cluster have a strong correlation. Based on the clustering results, the position of the knowledge unit in the low-dimensional space is fine-tuned to make the mapping positions of similar knowledge units closer.

[0026] Step S103, based on the association representation and difference representation between knowledge units, by designing an optimization objective function, a gradient descent algorithm is used to dynamically solve the optimal knowledge fusion strength parameter in the current environment to balance the correlation and complementarity between knowledge units.

[0027] The cosine distance matrix is ​​calculated according to the knowledge unit association representation vector and the difference representation vector, and the correlation matrix and the complementarity matrix are obtained by maximum and minimum value normalization processing; the cross entropy loss function is used to construct the optimization objective function, and the optimization objective function includes the weighted sum of the correlation matrix and the complementarity matrix, and the gradient value of the fusion strength parameter to the objective function is obtained by the gradient calculator; the fusion strength parameter is updated by the momentum optimizer according to the gradient value, and the rate of change of the loss function value is calculated by the sliding window calculator, and the optimization is stopped if the rate of change obtained by continuous calculation is less than the preset threshold; a weight matrix is ​​constructed for the optimized fusion strength parameter, and a weighted summer is used to perform a fusion operation on the knowledge unit association representation vector and the difference representation vector to obtain a real-valued matrix representing the degree of fusion of the knowledge units.

[0028] Specifically, the distance matrix between the knowledge unit association representation vector and the difference representation vector is calculated according to the cosine distance calculator, and the knowledge unit correlation matrix and complementarity matrix are obtained through maximum and minimum value normalization. The initial parameters of fusion strength in the interval [0,1] are set for the matrix elements. The cross entropy loss function is used to construct the optimization objective function. The correlation term weight coefficient α and the complementarity term weight coefficient β are set according to the historical fusion effect of the knowledge unit to meet the constraint condition of α+β=1. The first-order gradient value of the fusion strength parameter to the objective function is calculated by the gradient calculator. The learning rate η is set according to the exponential decay method, with an initial value of 0.1 and a decay factor of 0.95 for each iteration. The parameter gradient is accumulated through the momentum optimizer, and the momentum coefficient is set to 0.9. The loss function value is recorded during the iteration process. For the loss function value of each round of iteration, the sliding window calculator is used to calculate the loss value change rate of the last 10 iterations. If the change rate is less than the preset threshold for 5 consecutive times or the maximum number of iterations is reached, the optimization is stopped. The optimized fusion strength parameters are constructed into a weight matrix, and the weighted summer is used to perform weighted fusion on the knowledge unit association representation vector and the difference representation vector. The real-valued matrix representing the degree of fusion of the knowledge unit is obtained through matrix multiplication. The knowledge units are sorted according to the size of the elements in the real-valued matrix, and the adaptive threshold selector is set to select the first k knowledge unit pairs with the highest fusion strength values ​​to generate the optimal fusion solution for the knowledge units. In the process of calculating the knowledge fusion strength, the cosine distance between the knowledge unit vectors is first calculated. Taking the two knowledge units of "machine learning" and "deep neural network" as examples, their association representation vectors are [0.8, 0.6, 0.4] and

[0029] [0.7,0.5,0.3], the cosine distance calculation result is 0.89, and the cosine distance of the difference representation vector is 0.25. Through the maximum and minimum value normalization processing, all distance values ​​are mapped to the interval [0,1], and the correlation matrix element 0.85 and the complementarity matrix element 0.32 are obtained. In the objective function construction, the correlation weight coefficient α=0.6 and the complementarity weight coefficient β=0.4 are set to reflect the balance between the correlation and complementarity of knowledge units. The cross entropy loss function is in the form of L=-Σ(y i l og(p i )), where y i is the true value, p i is the predicted value. For the fusion strength parameter θ of the knowledge unit pair, its gradient calculation formula is Where L_r and L_c represent the correlation loss and complementary loss, respectively. During the parameter optimization process, the learning rate η starts from 0.1, and is multiplied by the attenuation factor 0.95 for each iteration, and drops to 0.06 after 10 iterations. The momentum optimizer accumulates historical gradient information, and the update formula is vt=0.9v(t-1)+ηgt, θt=θ(t-1)-vt, where vt is the current momentum, gt is the current gradient, and θt is the current fusion strength parameter. Taking a certain iteration as an example, the current gradient is 0.05, and the momentum at the previous moment is 0.03. The momentum after the update is 0.077, and the parameter is reduced by 0.077. The rate of change of the loss value is calculated by a sliding window with a window size of 10. For the sequence of loss values ​​obtained by continuous iterations

[0030] [0.52, 0.48, 0.45, 0.43, 0.42, 0.41, 0.40, 0.39, 0.385, 0.382], the average value of the difference between adjacent loss values ​​is calculated to be 0.015. When the change rate is less than the threshold value of 0.02 for 5 consecutive times, the optimization process is considered to converge. In the final fusion strength matrix, the fusion strength value of the knowledge unit pair "machine learning-deep learning" is 0.82, "deep learning-neural network" is 0.78, and "machine learning-data mining" is 0.65. Through the adaptive threshold selector, the k value is set to 20% of the total number of knowledge unit pairs, and the knowledge unit pair with the highest fusion strength value is selected to construct a knowledge fusion scheme to achieve the optimal combination of knowledge units.

[0031] Step S104, based on the real-time update of the knowledge fusion strength parameters, the ant colony optimization algorithm is used to perform a heuristic search in the knowledge unit association graph, and the semantic similarity, association and difference between the knowledge units are combined to select the candidate fusion knowledge units that achieve the target relevance and target complementarity with the current knowledge unit.

[0032] According to the knowledge unit association graph, the semantic similarity vector and the association vector between adjacent knowledge units are calculated, and the initial pheromone concentration matrix is ​​obtained by weighted summation; the initial pheromone concentration matrix is ​​used to perform ant colony optimization search on the graph path, and the ant colony optimization search determines the next hop node according to the transfer probability distribution; for the path obtained by the ant colony optimization search, the pheromone concentration value on the path is calculated using the pheromone update function, wherein the pheromone concentration value is determined by the product of the pheromone volatility coefficient and the original concentration value; the target correlation score between the knowledge units on the path is calculated using the pheromone concentration value, and if the target correlation score is higher than a preset threshold, the path is determined to be the optimal fusion path.

[0033] Specifically, a node transfer probability matrix is ​​constructed based on the knowledge unit association graph. The semantic similarity vector, association vector and difference vector are calculated for each pair of adjacent knowledge units in the graph. The initial pheromone concentration matrix is ​​obtained by weighted summation, and the transfer probability distribution is generated based on the initial matrix. The ant colony optimization algorithm is used to search on the graph path, and the next node is selected according to the current transfer probability distribution. The current path score is calculated by the path evaluation function, and the pheromone increment matrix is ​​calculated based on the score value and the fusion strength parameter. For each search path, the pheromone update function is used to update the pheromone concentration on the path. The update formula is τ(t+1)=(1-ρ)τ(t)+Δτ, where ρ is the pheromone volatility coefficient, τ(t) is the current concentration value, and Δτ is the pheromone increment. The roulette selector is used to screen the search path, and the target relevance score between knowledge units on the path is calculated by the semantic similarity calculator, and the target complementarity score is calculated based on the vector cosine distance. The fitness evaluation matrix is ​​constructed based on the target relevance score and the target complementarity score. The path fitness value is obtained through normalization processing, and the path with the highest fitness value is selected as the local optimal solution. The global pheromone concentration matrix is ​​updated based on the local optimal path, and the pheromone is volatilized and updated through the exponential decay function. The candidate fusion set is constructed for the knowledge units whose concentration values ​​exceed the preset threshold after the update. In the knowledge unit association map, the transition probability matrix is ​​constructed with "machine learning" as the central node, and the adjacent nodes include "deep learning", "statistical methods" and "data mining". When calculating the semantic similarity, the word vector cosine similarity method is used. The similarity between "machine learning" and "deep learning" is 0.85, the similarity with "statistical methods" is 0.65, and the similarity with "data mining" is 0.75. The correlation is calculated by co-occurrence frequency, and the difference is calculated by the Euclidean distance of the feature vector. The weights of the three indicators are 0.4, 0.3, and 0.3 respectively, and the initial pheromone concentration value is obtained by weighting. In the path search process, starting from the "machine learning" node, the next node is selected based on the transition probability {deep learning: 0.45, statistical method: 0.25, data mining: 0.3}. The path evaluation function comprehensively considers the path length and node relevance. For the path "machine learning → deep learning → neural network", its score is 0.82. The current fusion strength parameter is 0.7, and the pheromone increment calculation formula is Δτ = Q score value fusion strength, where Q is a constant of 0.1, and the pheromone increment is 0.0574. In the pheromone update process, the volatility coefficient ρ is set to 0.1, and for the initial concentration τ(t) = 0.5, the increment Δτ = 0.0574, and the updated concentration value τ(t+1) = 0.45 + 0.0574 = 0.5074. When selecting a path through the roulette selector, the selection probability of each path is proportional to its pheromone concentration. The target relevance is obtained by calculating the geometric mean of the semantic similarity of adjacent nodes on the path. The relevance score of the path "machine learning → deep learning → neural network" is 0.83.The target complementarity is calculated based on the difference degree of node feature vectors, and the score is 0.45. When evaluating fitness, the relevance weight is 0.6 and the complementarity weight is 0.4. For the above path, the fitness value is 0.83×0.6+0.45×0.4=0.678. The global pheromone update adopts the exponential decay function f(t)=e^(-0.05t), where t is the number of iterations. The pheromone concentration threshold is set to 0.4. When the pheromone concentration on a certain path exceeds the threshold, the knowledge unit on the path is selected into the candidate fusion set. After 10 rounds of iterations, the pheromone concentration of the path "machine learning→deep learning→neural network→deep reinforcement learning" reaches 0.45, and the knowledge unit on this path becomes a candidate fusion object.

[0034] Step S105, based on the association representation and difference representation, identify the potential contradictions, conflicts and incompatibilities between knowledge units, infer the technical problems caused by the contradictions, adopt a combination of rule-based reasoning and case-based reasoning, match the characteristics of the current knowledge unit according to the historical technical problem case library, and generate technical problem hypotheses.

[0035] A knowledge unit feature matrix is ​​constructed according to the association representation vector and the difference representation vector, and a cosine distance calculator is used to obtain the conflict degree value between the knowledge units. A normalization function is used to determine whether the conflict degree value is higher than a preset threshold. If the conflict degree value is higher than the preset threshold, a rule description vector corresponding to the knowledge unit is obtained from the rule base, and a text similarity calculator is used to score the matching degree of the rule description vector. A rule dependency graph is constructed according to the matching degree score of the rule description vector, and forward reasoning is performed on the rule dependency graph through a graph traversal algorithm. The state transfer matrix is ​​recorded at the reasoning node of the rule dependency graph to obtain the problem feature vector. A vector distance calculator is used to calculate the Euclidean distance value between the problem feature vector and the historical case feature vector. If the Euclidean distance value is lower than the preset similarity threshold, the problem description vector and the solution vector of the historical case are extracted to construct a case feature matrix.

[0036] Specifically, the knowledge unit feature matrix is ​​constructed based on the association representation vector and the difference representation vector. The cosine distance calculator is used to calculate the conflict value between knowledge units. The conflict degree is mapped to the [0,1] interval through the normalization function. When the conflict degree is higher than the preset threshold, the knowledge unit pair is marked in the contradiction feature matrix. For the knowledge unit pairs marked in the contradiction feature matrix, the rule matcher is used to extract the corresponding rule description vector from the rule base, and the text similarity calculator is used to score the matching degree of the rule description, and the rules are sorted in descending order based on the score value. The rule dependency graph is constructed based on the sorted rule sequence, and the rules are forward reasoned through the graph traversal algorithm. The intermediate state is recorded at each reasoning node, and the problem feature vector is generated based on the state transition matrix. The vector distance calculator is used to calculate the Euclidean distance between the problem feature vector and the historical case feature vector, and the historical cases are sorted by similarity according to the distance value. The cases with similarity exceeding the preset threshold are selected as the matching case set. For the matching case set, the problem description vector, solution vector and result evaluation vector contained in the case are extracted, and the case feature matrix is ​​constructed through vector splicing operation. The weighted fusion is used to fuse the rule reasoning result vector and the case feature matrix. The fused features are clustered and grouped by the feature clustering algorithm, and the problem hypothesis set of the current knowledge unit is constructed based on the cluster center. Taking the knowledge units "deep learning" and "traditional machine learning" as examples, their association representation vectors are [0.8, 0.6, 0.4] and [0.5, 0.7, 0.3] respectively, and the difference representation vectors are [0.3, 0.5, 0.7] and [0.5, 0.7, 0.3] respectively.

[0037] [0.6, 0.4, 0.2]. The conflict degree between two knowledge units is calculated, and the cosine distance is used to obtain 0.85. The normalized value is 0.92, which exceeds the preset threshold of 0.8. The corresponding position in the contradiction feature matrix is ​​marked as 1. The rule description extracted from the rule base contains "computational complexity limit", "data scale requirement", "hardware resource constraint" and other contents. Through text similarity calculation, the matching degree between rule 1 and the current knowledge unit is 0.78, rule 2 is 0.65, and rule 3 is 0.82. The rule dependency graph is constructed after sorting based on the matching degree, and the transition probability between nodes is calculated according to the rule matching degree. In the forward reasoning process, starting from the rule "computational complexity limit", the "GPU insufficient memory" problem is derived, and the state transition matrix records the state changes on the reasoning path. The problem feature vector contains three dimensions: "resource limitation", "performance bottleneck", and "optimization space", with values ​​of [0.9, 0.7, 0.8] respectively. When searching for similar cases in the historical case library, the Euclidean distance is calculated to obtain a distance of 0.25 for case 1, 0.42 for case 2, and 0.31 for case 3. Case 1 and case 3 whose distance is less than the threshold of 0.35 are selected as matching cases. Problem description vectors [Insufficient memory, 0.9], [Computational delay, 0.8], solution vectors [Model compression, 0.85], [Distributed computing, 0.75], and result evaluation vectors [Performance improvement, 0.8], [Resource saving, 0.7] are extracted from the matching cases. The case feature matrix dimension is 6×2, and each row corresponds to a feature dimension. The rule reasoning result vector is fused with the case feature matrix, with the rule weight of 0.6 and the case weight of 0.4. The K-means algorithm is used to cluster the fused features, and the number of clusters is set to 3. The first type of center represents the "model structure optimization" problem, with the eigenvalues ​​of [0.85, 0.78, 0.82]; the second type of center represents the "computing resource allocation" problem, with the eigenvalues ​​of

[0038] [0.75, 0.82, 0.68]; the third type of center represents the "training strategy adjustment" problem, and the eigenvalue is

[0039] [0.72, 0.65, 0.88]. The problem assumptions generated based on cluster centers contain potential conflicts and solution directions of knowledge units at different levels.

[0040] Step S106, by introducing technical problem hypotheses, identifying key terms and concepts related to the problem, highlighting the semantic associations in the problem space, using multi-task learning and contrastive learning algorithms to build a domain ontology, establishing a mapping relationship between technical problem hypotheses and knowledge units in the domain ontology, identifying and strengthening the deep semantic associations in the problem space, and thus building a cross-domain problem-oriented knowledge mapping ontology.

[0041] According to the problem hypothesis, term phrases are obtained, and the term phrases are segmented and labeled by a word segmenter and a named entity recognizer to obtain an annotated term vector; a co-occurrence matrix is ​​constructed for the annotated term vector, and the co-occurrence matrix calculates the correlation strength between nodes by a node connection weight calculator, and if the correlation strength is higher than a preset threshold, a concept map is obtained; a task model is constructed for the concept map using a multi-task learner, and the task model calculates the classification error and the correlation prediction error by a cross entropy loss function to obtain an optimized concept map; a concept node with the highest degree centrality is selected from the optimized concept map as an anchor concept, and the anchor concept calculates the semantic similarity between nodes by a cosine similarity calculator to obtain a sample feature representation.

[0042] Specifically, key term phrases are extracted according to the problem hypothesis, and the terms are segmented and annotated using a word segmenter and a named entity recognizer. The annotated terms are converted into semantic vectors through a word vector generator, and a problem concept node set is constructed based on a vector distance matrix. For the concept node set, a co-occurrence matrix is ​​constructed based on word frequency statistics, the node connection weight calculator is used to calculate the association strength between nodes, and a threshold filter is used to retain node pairs with an association strength higher than a preset threshold to generate a problem space concept map. For the concept map, a multi-task learner is used to construct concept classification tasks and association prediction tasks, the classification error is calculated using a cross entropy loss function, the association prediction error is calculated using a mean square error function, and a comprehensive loss function is obtained based on weighted summation. The gradient descent optimizer is used to iteratively optimize the loss function, the shared feature layer parameters are updated through a back-propagation algorithm, and the optimal feature representation is selected based on the validation set evaluation index. The concept node with the highest degree centrality is selected from the optimized concept map as the anchor concept, the semantic similarity between concept nodes is calculated through a cosine similarity calculator, nodes with similarity higher than the positive sample threshold are selected to construct positive sample pairs, and nodes with similarity lower than the negative sample threshold are selected to construct negative sample pairs. The contrast loss function is used to train the sample feature representation. The loss function is in the form of L=max(0,m+d(a,p)-d(a,n)), where m is the boundary parameter, d is the Euclidean distance, a is the anchor sample, p is the positive sample, and n is the negative sample. The domain ontology hierarchy is constructed based on the trained feature representation, the same-level concept relationship is established through the concept similarity matrix, and the hierarchical relationship identifier is used to construct the concept relationship to generate the ontology tree structure. Based on the ontology tree structure, the semantic mapper is used to establish the correspondence between the problem hypothesis node and the knowledge unit node. The indirect association between the nodes is expanded through the relationship propagation algorithm to identify the cross-domain semantic mapping relationship. Starting from the problem hypothesis "deep neural network training optimization", key terms including "gradient vanishing", "overfitting", "parameter tuning" are extracted and converted into 300-dimensional word vectors after word segmentation and annotation. For the node co-occurrence matrix, the co-occurrence frequency of "gradient vanishing" and "activation function" is 85, and the co-occurrence frequency of "overfitting" and "regularization" is 92. The node association strength is calculated based on the co-occurrence frequency. The association strength threshold is set to 0.6, and the concept graph is constructed including the edge weights between nodes. In multi-task learning, the classification task divides the concept nodes into two categories: "problem class" and "method class", and the association prediction task predicts the probability of association between node pairs. The cross entropy loss value of the classification task is 0.42, and the mean square error of the association prediction task is 0.35. The task weights are set to 0.6 and 0.4 respectively, and the comprehensive loss value is 0.392. By back-propagation to update the parameters, the classification accuracy on the validation set reaches 0.85, and the association prediction F1 value reaches 0.78. The concept node "neural network optimization" with the highest degree centrality is selected as the anchor concept, and the cosine similarity with other nodes is calculated.The similarity of "parameter initialization" is 0.82, and the similarity of "batch normalization" is 0.78, both of which are higher than the positive sample threshold of 0.7, and are selected as positive sample pairs. The similarity of "data cleaning" is 0.25, and the similarity of "feature selection" is 0.28, which is lower than the negative sample threshold of 0.3, and is selected as negative sample pairs. In contrastive learning, the boundary parameter m=0.5 is set, and the Euclidean distance between the anchor sample and the positive and negative samples is calculated. Taking a set of training samples as an example, d(a,p)=0.3, d(a,n)=0.9, substituting into the loss function, we get L=max(0,0.5+0.3-0.9)=0, indicating that the feature representation of this sample pair has satisfied the boundary constraint. When constructing the ontology hierarchy, the same-level relationship is determined by the concept similarity, such as the similarity of "gradient disappearance" and "gradient explosion" is 0.88, and the same-level association is constructed. The identification of the upper and lower relationships is based on the word inclusion relationship and contextual semantics, such as "optimization strategy" is the superordinate concept of "parameter tuning". The final ontology tree contains a 4-layer structure, with the top node being "deep learning optimization" and the second layer containing three branches: "model structure", "training process", and "evaluation index". In the semantic mapping, the problem hypothesis node "gradient vanishing" is directly associated with the knowledge unit "activation function selection", with a correlation of 0.85. The indirect association with "parameter initialization" is further identified through relationship propagation, with a correlation of 0.72 after propagation. The entire mapping network reflects the multi-level semantic association between the problem and the solution.

[0043] Step S107, using the knowledge mapping ontology to guide the fusion of cross-domain knowledge, using an ontology reasoning engine based on description logic to infer the semantic consistency relationship between fused knowledge units, and identifying and eliminating potential semantic conflicts through a consistency checking algorithm.

[0044] The concept, relationship and attribute constraint information in the knowledge unit is obtained through a semantic parser, and a description logic representor is used to obtain a formal expression containing concept types, role relationships and attribute constraints; a set of inference rules is called according to the formal expression to perform forward reasoning operations, and the inference rule set contains concept subtypes, equivalence relations, and mutually exclusive relations, so as to judge the semantic consistency scores between similar concepts; the semantic consistency scores are processed by a constraint propagation function to obtain propagation path information, and the target nodes with inconsistent propagation in the node subset are located by a connected subgraph identifier; the semantic structure contradiction degree of the target node is judged, and if the contradiction degree exceeds a preset threshold, a conflict resolver is used to adjust the node semantic constraints to obtain new semantic constraint rules.

[0045] Specifically, a set of semantic descriptions of knowledge units is constructed based on the mapping ontology. Concepts, relations and attribute constraints are extracted through a semantic parser. The description logic representor is used to convert the semantic content into a formal expression containing concept type C(x), role relationship R(x,y) and attribute constraint P(x). For the formal expression, forward reasoning is performed based on a set of inference rules containing concept subtypes, equivalence relations and mutually exclusive relations. The semantic consistency scores between similar concepts are calculated based on the rule matching degree. The propagation path of the semantic consistency score is calculated using the constraint propagation function. The node subsets with inconsistent propagation are located through the interconnected subgraph identifier. The semantic relationship graph of the knowledge unit is constructed based on the node semantic constraints. Mutually exclusive concept pairs, contradictory attribute pairs and conflicting relationship pairs are extracted from the semantic relationship graph to construct a conflict identification rule base. The subgraph matcher is used to determine the degree of contradiction in the semantic structure between knowledge units. The semantic structure with a degree of contradiction exceeding the preset threshold is traced back through the path tracker, the key nodes on the contradiction propagation path are recorded, and the conflicting nodes are sorted based on the node importance. For the conflicting nodes after sorting, the constraint satisfaction-based conflict resolver is used to adjust the node semantic constraints, and new semantic constraint rules are generated through constraint relaxation and constraint strengthening. According to the new semantic constraint rules, the semantic completion function is used to reconstruct the association relationship of the resolved knowledge units, and the rationality of the association relationship is evaluated by the relationship strength calculator. For the knowledge unit pairs with association strength lower than the preset threshold, their semantic descriptions are expanded based on the completion rule library, and the semantic integrity of the expanded knowledge units is verified by the semantic consistency verifier. Taking the knowledge units "deep neural network" and "traditional machine learning" as examples, the extracted semantic descriptions include the concept type "learning model (x)", the role relationship "use (x, data)", and the attribute constraint "computational complexity (x)> linear". It is converted into a formal expression through the description logic representor: C (deep neural network) ∧ R (deep neural network, data) ∧ P (deep neural network). Semantic reasoning is performed based on a set of inference rules, including "if x is a deep neural network, then x is a machine learning model" (concept subtype rule), "neural network optimization ≡ parameter tuning" (equivalence relationship rule), and "online learning ⊥ offline learning" (mutually exclusive relationship rule). When calculating the rule matching degree, the semantic consistency score of "deep neural network" and "machine learning model" is 0.85. When propagating constraints, starting from the initial node "deep learning", the semantic consistency score is propagated through the edge weight. For the path "deep learning→neural network→parameter optimization", the score after propagation is 0.85×0.78=0.663. Nodes with scores lower than the threshold of 0.6 in the connected subgraph are identified, and a semantic relationship graph is constructed to reflect the inconsistent relationships between nodes. In conflict identification, mutually exclusive concept pairs such as "batch learning" and "incremental learning", contradictory attribute pairs such as "high-dimensional features" and "feature dimensionality reduction", and conflicting relationship pairs such as "model compression" and "model expansion" are extracted.The structural similarity is calculated through subgraph pattern matching to determine the degree of conflict. If the number of mutually exclusive concepts contained in two knowledge units exceeds the threshold of 2, they are marked as strongly conflicting structures. When tracing the path, the structure with a conflict degree of 0.75 is backtracked and the importance values ​​of the nodes on the propagation path are recorded. The importance of the node "parameter optimization" is 0.82, and the importance of "model compression" is 0.65. The conflicting nodes are sorted based on the importance. When adjusting the constraints, the constraints of high-importance nodes are relaxed, such as changing the computational complexity constraint of "parameter optimization" from "greater than" to "greater than or equal to". In the semantic completion stage, the association of knowledge units is rebuilt based on the new constraint rules. The completion rules include "concept hierarchy completion", "attribute inheritance completion" and "relationship transfer completion". The association strength after completion is calculated. The association strength of "deep learning" and "model compression" is 0.72, which exceeds the preset threshold of 0.7, verifying the rationality of the association. If the association strength is lower than the threshold, the semantic description is expanded based on the completion rules, such as adding the intermediate concept "model structure optimization" to establish an indirect association.

[0046] Step S108, conduct quality assessment on the fused knowledge units, design a comprehensive evaluation index system, evaluate the relevance, complementarity, consistency and conflict-free nature of the fused knowledge, use the analytic hierarchy process to calculate the index weights of the fused knowledge, and use the weighted summation method to obtain the knowledge fusion quality score. According to the evaluation results, iteratively optimize the fusion process, adjust the knowledge fusion intensity and mapping strategy, and finally generate cross-domain fusion knowledge.

[0047] According to the fused knowledge unit, a relevance index value is calculated by using vector cosine, a complementarity index value is calculated by using information entropy, a consistency index value is calculated by using semantic distance, and a conflict detection function is used to calculate a conflict-free index value, so as to obtain a four-dimensional evaluation vector; a judgment matrix is ​​constructed for the four-dimensional evaluation vector, and the maximum eigenvalue of the judgment matrix is ​​calculated by using the row vector geometric mean method, and the corresponding eigenvector is obtained as a weight vector; a weighted quality score is obtained by multiplying the evaluation vector and the weight vector, and a dual threshold comparator is used to determine whether the weighted quality score is within a preset quality range; an error feedback signal is constructed for the weighted quality score, and a gradient descent method is used to update a fusion strength parameter, and a knowledge unit mapping relationship is reconstructed by using the fusion strength parameter to determine whether the association strength of the mapping relationship exceeds a preset threshold.

[0048] Specifically, an evaluation index matrix is ​​constructed based on the fused knowledge units. The relevance index value is calculated using vector cosine, the complementarity index value is calculated using information entropy, the consistency index value is calculated based on semantic distance, and the conflict-free index value is calculated using the conflict detection function to generate a four-dimensional evaluation vector. The judgment matrix A is constructed by pairwise comparison. The matrix elements ai j represent the importance of index i relative to index j. The 1-9 scaling method is used to assign values, and the maximum eigenvalue of the judgment matrix is ​​calculated by the row vector geometric mean method. For the maximum eigenvalue λmax, the corresponding eigenvector w is calculated as the index weight vector, and the consistency ratio CR=CI / RI is used to evaluate the rationality of weight allocation, where CI=(λmax-n) / (n-1) is the consistency index and RI is the random consistency index. The weighted quality score s=v·w is calculated based on the evaluation vector v and the weight vector w, and the score sorter is used to generate the knowledge unit quality sorting table. The knowledge units with quality scores in the interval [θ1,θ2] are screened out by the double threshold comparator. For the screened knowledge units, an error feedback signal is constructed based on the quality score, the fusion strength parameter is updated by the gradient descent method, and the parameter update rate is adjusted by the adaptive step size controller. The mapping optimizer is used to reconstruct the knowledge unit mapping relationship based on the updated fusion strength parameters. The rationality of the mapping relationship is evaluated by the association strength calculator, and the mapping relationship with an association strength lower than the threshold is optimized and adjusted. The evaluation process is re-executed for the optimized knowledge unit, and the change curve of the index value is recorded by the evaluation index tracker. The convergence of the quality evaluation process is judged based on the convergence judge. The multi-round iterative optimizer is used to repeatedly evaluate and optimize the non-convergent knowledge units until the change rate of the quality score is less than the preset threshold or the maximum number of iterations is reached, and the final fusion knowledge is output. Taking the knowledge unit "deep learning" and "transfer learning" as examples, the constructed evaluation vector contains four dimensions: when calculating the relevance, 0.82 is obtained based on the cosine similarity of the word vector; complementarity is calculated by information entropy H=-Σp(x)log(p(x)) to obtain 0.75; consistency is calculated based on semantic distance to obtain 0.68; the result of non-conflict detection is 0.90, thus obtaining the evaluation vector v=[0.82,0.75,0.68,0.90]. When constructing the judgment matrix, the importance ratio of relevance to complementarity is set to 3, the ratio of relevance to consistency is set to 5, and the ratio of relevance to non-conflict is set to 2, and the judgment matrix is ​​obtained: A =

[1352] [1 / 3121 / 2][1 / 51 / 211 / 3][1 / 2231] The maximum eigenvalue λmax of the judgment matrix is ​​calculated to be 4.05, and the corresponding eigenvector

[0049] w=[0.45,0.15,0.08,0.32] as the weight vector. Calculate the consistency index CI=(4.05-4) / (4-1)=0.017, look up the table to get the random consistency index RI=0.89, calculate the consistency ratio

[0050] CR=0.017 / 0.89=0.019<0.1, indicating that the weight distribution is reasonable. The weighted sum is used to calculate the quality score: s=0.82×0.45+0.75×0.15+0.68×0.08+0.90×0.32=0.831. The quality score threshold interval is set to [0.7,0.9], and the score of this knowledge unit pair is within the threshold interval. The error signal e=0.9-0.831=0.069 is used to guide parameter update, and the fusion strength parameter is adjusted from the initial value of 0.6 in steps of 0.05. In the mapping optimization, the updated fusion strength parameter is 0.65, and the association strength between knowledge units is recalculated, and the association strength between "deep learning feature extraction" and "transfer learning target domain" is 0.76, which exceeds the preset threshold of 0.7. After three rounds of iterative optimization, the quality score increased from 0.831 to 0.852, 0.867, and 0.875. The change rate of the last two rounds of scores was 0.009, which was less than the convergence threshold of 0.01, and the optimization process ended. During the entire evaluation and optimization process, the change curve of the evaluation indicators showed that the relevance index rose steadily from 0.82 to 0.88, the complementarity index fluctuated slightly and remained at around 0.75, the consistency index increased significantly from 0.68 to 0.82, and the conflict-free index was basically stable at around 0.90, which comprehensively reflected the continuous improvement of the quality of knowledge fusion.

[0051] Step S109, based on the technical problem hypothesis generated by the fusion knowledge, an intelligent optimization algorithm is used to solve the solution, and the objective function of technical feasibility and economic benefit is defined as the optimization criterion. The candidate solutions are generated and evaluated through algorithm iteration until convergence or the stopping condition is met. The candidate solutions are compared through simulation and analysis, and the target solution is selected to form a complete description of the technical problem and the solution.

[0052] A binary code string is obtained according to the problem parameters, wherein the binary code string is obtained by converting continuous parameters by a variable encoder and performing one-hot encoding on discrete parameters by a constraint mapping function; a feasibility score and an economic score are calculated for the binary code string, wherein the feasibility score is generated by a resource consumption calculator and the economic score is generated by an input-output calculator; a particle swarm algorithm is used to iteratively optimize the feasibility score and the economic score, and a search direction is obtained through a speed update formula; an evaluation matrix is ​​constructed for the search direction, wherein the evaluation matrix is ​​generated by performance indicators and a comprehensive score of the solution is obtained through a fuzzy comprehensive evaluation method.

[0053] Specifically, the solution parameter vector is constructed according to the problem hypothesis, the variable encoder is used to convert the continuous parameters into binary strings according to the accuracy requirements, the discrete parameters are one-hot encoded through the constraint mapping function, and the candidate solution population is generated based on the encoding rules. For the candidate solution population, the resource consumption calculator is used to calculate the feasibility score f technology, the input-output calculator is used to calculate the economic score f benefit, the weight coefficients w1 and w2 are determined by the hierarchical analysis method, and the comprehensive objective function F = w1×f technology + w2×f benefit is constructed. The particle swarm algorithm is used to iteratively optimize the candidate solutions, the search direction is adjusted by the speed update formula, and the candidate solutions are updated based on the position update formula. For the updated candidate solutions, the fitness sorter is used to sort in descending order, the roulette selector is used to select the high-quality solutions, and the gene fragments are exchanged at the intersection position based on the crossover probability. For the crossover solution space, the mutation operator is used to flip the gene bit with the mutation probability, the local area is deeply searched by the hill climbing method, and the local optimal solution is selected based on the greedy criterion. The simulation environment is constructed according to the discrete event definition, the event scheduler is used to advance the simulation process in time sequence, the solution execution process is simulated by the state transition function, and the key performance indicators are recorded. An evaluation matrix is ​​constructed based on performance indicators, and the comprehensive score of the scheme is calculated by the fuzzy comprehensive evaluation method. The Pareto optimization criterion is used to screen the non-dominated solution set. For the non-dominated solution set, a multi-criteria decision maker is used to comprehensively evaluate the performance of the scheme under each objective. The final score is obtained through normalization and weighted summation, and the scheme with the highest score is selected as the optimal solution. For the knowledge fusion optimization problem, the "deep neural network structure optimization" is taken as an example to solve the scheme. The parameter vector contains three dimensions: the number of network layers, the number of neurons, and the type of activation function. For the continuous parameter "number of neurons", it is encoded with an accuracy of 0.1, and the value range [64,512] is mapped to a 10-bit binary string; for the discrete parameter "activation function type", 3-bit unique hot encoding is used to represent the three options of ReLU, Sigmoid and Tanh. When constructing the objective function, the f technology obtains the feasibility score by calculating resource consumption indicators such as GPU memory occupancy and computing time, and the f benefit calculates the economic score based on indicators such as training accuracy and inference speed. The weights were determined by analytic hierarchy process, and the judgment matrix was constructed by expert scoring. The calculated values ​​were w1=0.6 and w2=0.4. In particle swarm optimization, the population size was set to 50, the maximum number of iterations was 100, the inertia weight w=0.729, and the acceleration factor c1=c2=1.494. A candidate solution x=[10100|01000|100] in the initial population represents 384 neurons, and the sigmoid activation function was selected. After speed and position updates, the solution evolved to x=[10110|01000|100] at the 10th iteration, indicating that the number of neurons increased to 416. The population was sorted in descending order of fitness, and the roulette selector was used to select the top 80% of the solutions for crossover. The crossover probability was set to 0.8, and single-point crossover was performed at the 5th and 8th positions.The mutation probability is set to 0.1, and the gene bits are randomly flipped. For example,

[100] is mutated to

[110] to indicate that the Tanh activation function is used instead. In the local search, a search space with a neighborhood radius of 1 is constructed with the current solution as the center. A discrete event simulation environment is constructed, and four types of events, data loading, forward propagation, backpropagation, and parameter update, are defined. The simulation is advanced in time sequence through the event scheduler. Indicators such as loss value, accuracy, and resource occupancy rate during the training process are recorded to construct an evaluation matrix. The solution score is obtained by fuzzy comprehensive evaluation. For example, the membership of a solution in the three dimensions of computational efficiency, memory efficiency, and optimization effect is 0.85, 0.72, and 0.78, respectively. The Pareto optimization criterion is applied to select 15 non-dominated solutions from 100 candidate solutions to form the Pareto frontier. Multi-criteria decision evaluation is performed on these solutions, and the computational efficiency weight is set to 0.3, the memory efficiency weight is set to 0.3, and the optimization effect weight is set to 0.4. The final score is obtained by weighted summation. The solution with the highest score uses a 6-layer network structure with 384 neurons and Tanh activation function, achieving a good balance between resource consumption and optimization effect.

[0054] Obviously, those skilled in the art can make various changes and modifications to the embodiments of the present application without departing from the spirit and scope of the embodiments of the present application. Thus, if these modifications and variations of the embodiments of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for generating large-scale scientific and technological problems by integrating external knowledge, characterized in that: The method comprises: Extract knowledge units from an external knowledge base, calculate the similarity between knowledge units using a semantic matching algorithm, generate an association strength matrix, analyze the dynamic changes of the association strength matrix in real time, perceive the changes in the association between knowledge units, determine the difference, analyze the attribute characteristics of knowledge units that have changed significantly, and adjust and update the association strength matrix; Using graph neural networks for deep embedding learning, we can identify the association strength and difference between knowledge units, extract the implicit association patterns and subtle difference features between knowledge units by combining the association strength and difference information between knowledge units, and encode the target association representation and target difference representation in a low-dimensional dense vector space. According to the association representation and difference representation between knowledge units, the optimal knowledge fusion intensity parameter under the current environment is dynamically solved by designing the optimization objective function and using the gradient descent algorithm to balance the correlation and complementarity between knowledge units. Based on the real-time update of knowledge fusion strength parameters, the ant colony optimization algorithm is used to perform heuristic search in the knowledge unit association graph, and the candidate fusion knowledge units that achieve target relevance and target complementarity with the current knowledge units are selected by combining the semantic similarity, relevance and difference between the knowledge units. Based on association representation and difference representation, potential contradictions, conflicts and incompatibilities between knowledge units are identified, and technical problems caused by contradictions are inferred. A combination of rule-based reasoning and case-based reasoning is used to match the characteristics of current knowledge units according to the historical technical problem case library to generate technical problem hypotheses. By introducing technical problem hypotheses, identifying key terms and concepts related to the problem, highlighting the semantic associations in the problem space, and using multi-task learning and contrastive learning algorithms to build a domain ontology, a mapping relationship between technical problem hypotheses and knowledge units is established in the domain ontology, and the deep semantic associations in the problem space are identified and strengthened, thereby building a cross-domain problem-oriented knowledge mapping ontology; The knowledge mapping ontology is used to guide the fusion of cross-domain knowledge. The ontology reasoning engine based on description logic is used to infer the semantic consistency relationship between fused knowledge units. The consistency checking algorithm is used to identify and eliminate potential semantic conflicts. Conduct quality assessment on the fused knowledge units, design a comprehensive evaluation index system, evaluate the relevance, complementarity, consistency and non-conflict of fused knowledge, use the analytic hierarchy process to calculate the index weights of fused knowledge, and use the weighted summation method to obtain the knowledge fusion quality score. According to the evaluation results, iteratively optimize the fusion process, adjust the knowledge fusion intensity and mapping strategy, and finally generate cross-domain fusion knowledge. Based on the technical problem hypothesis generated by the fusion of knowledge, an intelligent optimization algorithm is used to solve the solution. The objective function of technical feasibility and economic benefit is defined as the optimization criterion. The candidate solutions are generated and evaluated through algorithm iteration until convergence or the stopping condition is met. The candidate solutions are compared through simulation and analysis, and the target solution is selected to form a complete description of the technical problem and solution.

2. The method according to claim 1, characterized in that The method extracts knowledge units from the external knowledge base, calculates the similarity between knowledge units based on semantic matching algorithm, generates an association strength matrix, analyzes the dynamic changes of the association strength matrix in real time, perceives the changes in the association between knowledge units, determines the difference, analyzes the attribute characteristics of knowledge units with significant changes, and adjusts and updates the association strength matrix, including: The word segmentation is performed according to the content of the knowledge unit to obtain the feature vector, and the weighted average algorithm is used to calculate the semantic similarity between the knowledge units to construct the initial association strength matrix; A time window is set for the initial association strength matrix to perform dynamic sampling, and if the difference in association strength in adjacent time windows exceeds a preset threshold, the position of the knowledge unit pair is marked as 1 in the difference matrix; Through the knowledge units marked as 1 in the difference matrix, the keyword frequency vector, the part-of-speech distribution vector and the semantic dependency vector are extracted to form a feature triple, and the attribute similarity matrix is ​​obtained by normalization processing; The initial association strength matrix and the attribute similarity matrix are weightedly fused, and the fused matrix is ​​decomposed and reconstructed by a non-negative matrix decomposition algorithm to obtain an updated knowledge unit association network.

3. The method according to claim 1, characterized in that The method uses graph neural network for deep embedding learning, identifies the association strength and difference between knowledge units, combines the association strength and difference information between knowledge units, extracts the implicit association pattern and slight difference features between knowledge units, and encodes the target association representation and target difference representation in a low-dimensional dense vector space, including: An adjacency matrix is ​​constructed according to the cosine similarity between knowledge units, features of directly connected nodes are extracted from the adjacency matrix, and initial node representation vectors are obtained by transferring and aggregating between nodes through information propagators; For the initial node representation vector, a multi-head attention calculator is used to obtain the attention score of the knowledge unit in the feature subspace, and the association strength value and the weak association label matrix are obtained through normalization processing; Constructing a knowledge unit feature matrix according to the association strength value and the weak association mark matrix, and compressing the feature matrix through an autoencoder to obtain a compressed representation vector; A cosine distance calculator is used to calculate the similarity values ​​between the compressed representation vectors, a knowledge graph is constructed according to the similarity values, and the knowledge units are grouped through a clustering algorithm. If the ratio of the average distance within a group to the minimum distance between groups is less than a preset threshold, the mapping position of the knowledge unit in the low-dimensional dense space is updated.

4. The method according to claim 1, characterized in that The optimal knowledge fusion strength parameter under the current environment is dynamically solved by using a gradient descent algorithm based on the association representation and difference representation between knowledge units, which is used to balance the correlation and complementarity between knowledge units, including: The cosine distance matrix is ​​calculated according to the knowledge unit association representation vector and the difference representation vector, and the correlation matrix and the complementarity matrix are obtained through the maximum and minimum value normalization process; The optimization objective function is constructed by using a cross entropy loss function, wherein the optimization objective function includes a weighted sum of the correlation matrix and the complementarity matrix, and a gradient value of the fusion strength parameter to the objective function is obtained by using a gradient calculator; According to the gradient value, the momentum optimizer is used to update the fusion strength parameter, and the rate of change of the loss function value is calculated by a sliding window calculator, and if the rate of change obtained by continuous calculation is less than a preset threshold, the optimization is stopped; A weight matrix is ​​constructed for the optimized fusion strength parameters, and a weighted summer is used to perform a fusion operation on the knowledge unit association representation vector and the difference representation vector to obtain a real-valued matrix representing the degree of fusion of the knowledge units.

5. The method according to claim 1, characterized in that The real-time update of the knowledge fusion strength parameter is based on the use of the ant colony optimization algorithm to perform heuristic search in the knowledge unit association graph, combining the semantic similarity, association and difference between the knowledge units to select candidate fusion knowledge units that achieve target relevance and target complementarity with the current knowledge unit, including: Calculating semantic similarity vectors and association vectors between adjacent knowledge units according to the knowledge unit association graph, and obtaining an initial pheromone concentration matrix by weighted summation; An ant colony optimization search is performed on the graph path using the pheromone concentration initial matrix, wherein the ant colony optimization search determines the next hop node according to the transfer probability distribution; For the path obtained by the ant colony optimization search, a pheromone update function is used to calculate the pheromone concentration value on the path, wherein the pheromone concentration value is determined by multiplying the pheromone volatility coefficient by the original concentration value; The target correlation score between knowledge units on the path is calculated by the pheromone concentration value. If the target correlation score is higher than a preset threshold, the path is determined to be the optimal fusion path.

6. The method according to claim 1, characterized in that Based on the association representation and difference representation, the potential contradictions, conflicts and incompatibilities between knowledge units are identified, and the technical problems caused by the contradictions are inferred. A combination of rule-based reasoning and case-based reasoning is adopted to match the characteristics of the current knowledge units according to the historical technical problem case library to generate technical problem hypotheses, including: A knowledge unit feature matrix is ​​constructed according to the association representation vector and the difference representation vector, a cosine distance calculator is used to obtain the conflict degree value between the knowledge units, and a normalization function is used to determine whether the conflict degree value is higher than a preset threshold; If the conflict degree value is higher than a preset threshold, a rule description vector corresponding to the knowledge unit is obtained from a rule base, and a text similarity calculator is used to score the matching degree of the rule description vector; A rule dependency graph is constructed according to the matching score of the rule description vector, forward reasoning is performed on the rule dependency graph through a graph traversal algorithm, and a state transfer matrix is ​​recorded at the reasoning node of the rule dependency graph to obtain a problem feature vector; A vector distance calculator is used to calculate the Euclidean distance value between the problem feature vector and the historical case feature vector. If the Euclidean distance value is lower than a preset similarity threshold, the problem description vector and solution vector of the historical case are extracted to construct a case feature matrix.

7. The method according to claim 1, characterized in that The method introduces technical problem assumptions, identifies key terms and concepts related to the problem, highlights the semantic associations in the problem space, uses multi-task learning and contrastive learning algorithms to build a domain ontology, establishes a mapping relationship between technical problem assumptions and knowledge units in the domain ontology, identifies and strengthens the deep semantic associations in the problem space, and thus builds a cross-domain problem-oriented knowledge mapping ontology, including: Acquire a term phrase according to the problem hypothesis, segment and annotate the term phrase through a word segmenter and a named entity recognizer to obtain an annotated term vector; Constructing a co-occurrence matrix for the annotated term vectors, wherein the co-occurrence matrix calculates the strength of association between nodes through a node connection weight calculator, and if the strength of association is higher than a preset threshold, a concept map is obtained; A multi-task learner is used to construct a task model for the concept map, and the task model calculates the classification error and the associated prediction error through a cross entropy loss function to obtain an optimized concept map; A concept node with the highest degree centrality is selected from the optimized concept map as an anchor concept. The anchor concept calculates the semantic similarity between nodes through a cosine similarity calculator to obtain a sample feature representation.

8. The method according to claim 1, characterized in that The knowledge mapping ontology is used to guide the fusion of cross-domain knowledge, and an ontology reasoning engine based on description logic is used to infer the semantic consistency relationship between fused knowledge units, and potential semantic conflicts are identified and eliminated through a consistency checking algorithm, including: The concept, relationship and attribute constraint information in the knowledge unit are obtained through the semantic parser, and the formal expression containing the concept type, role relationship and attribute constraint is obtained by using the description logic expressor; Calling an inference rule set to perform forward inference operations according to the formal expression, the inference rule set includes concept subtypes, equivalence relations, and mutual exclusion relations, and determining semantic consistency scores between similar concepts; The constraint propagation function is used to process the semantic consistency score to obtain the propagation path information, and the target node with inconsistent propagation in the node subset is located by connecting the subgraph identifier; The semantic structure contradiction degree of the target node is judged. If the contradiction degree exceeds a preset threshold, a conflict resolver is used to adjust the node semantic constraint to obtain a new semantic constraint rule.

9. The method according to claim 1, characterized in that: The quality of the fused knowledge units is evaluated, a comprehensive evaluation index system is designed, the relevance, complementarity, consistency and non-conflict of the fused knowledge is evaluated, the index weights of the fused knowledge are calculated using the hierarchical analysis method, and the knowledge fusion quality score is obtained using the weighted summation method. According to the evaluation results, the fusion process is iteratively optimized, the knowledge fusion intensity and mapping strategy are adjusted, and finally cross-domain fusion knowledge is generated, including: According to the fusion knowledge unit, a relevance index value is calculated by using vector cosine, a complementarity index value is calculated by using information entropy, a consistency index value is calculated by using semantic distance, and a conflict detection function is used to calculate a conflict-free index value, thereby obtaining a four-dimensional evaluation vector; Constructing a judgment matrix for the four-dimensional evaluation vector, calculating the maximum eigenvalue of the judgment matrix by the row vector geometric mean method, and obtaining the corresponding eigenvector as a weight vector; A weighted quality score is obtained by multiplying the evaluation vector by the weight vector, and a dual-threshold comparator is used to determine whether the weighted quality score is within a preset quality range; An error feedback signal is constructed for the weighted quality score, a fusion strength parameter is updated using a gradient descent method, a knowledge unit mapping relationship is reconstructed using the fusion strength parameter, and it is determined whether the association strength of the mapping relationship exceeds a preset threshold.

10. The method according to claim 1, characterized in that The technical problem hypothesis generated based on the fusion knowledge is solved by using an intelligent optimization algorithm, and the objective function of technical feasibility and economic benefit is defined as the optimization criterion. The candidate solutions are generated and evaluated through algorithm iteration until convergence or the stop condition is met. The candidate solutions are compared through simulation and analysis, and the target solution is selected to form a complete description of the technical problem and solution, including: Obtain a binary code string according to the problem parameters, wherein the binary code string is obtained by converting the continuous parameters by a variable encoder and performing one-hot encoding on the discrete parameters by a constraint mapping function; Calculating a feasibility score and an economic score for the binary code string, wherein the feasibility score is generated by a resource consumption calculator and the economic score is generated by an input-output calculator; A particle swarm algorithm is used to iteratively optimize the feasibility score and the economic efficiency score, and a search direction is obtained through a speed update formula; An evaluation matrix is ​​constructed for the search direction, wherein the evaluation matrix is ​​generated by performance indicators and a comprehensive score of the scheme is obtained by a fuzzy comprehensive evaluation method.

Images