Large model compression and real-time inference method and system based on knowledge distillation

Abstract

The application discloses a large model compression and real-time inference method and system based on knowledge distillation, relates to the technical field of data analysis, and comprises the following steps: extracting general and industry model keywords, calculating the correlation coefficient of the two, and constructing a general-industry model set and updating the keywords if the standard is met; forming an enterprise model set in combination with enterprise demand based on the industry model of the set, and constructing a general-industry-enterprise model or a sub-model set and updating the keywords according to the correlation of the enterprise model with the general and industry models; screening out a first model set from the various models and sub-sets that have been constructed in combination with enterprise demand keywords, defining a missing rate, calculating a matching value in combination with the past success rate of the model, the highest success rate of a candidate model and demand coverage, extracting defects and advantage features in comparison with an ideal model, calculating a recommendation score, sorting and pushing an optimal model, and the application realizes large model compression and real-time inference.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data analysis, in particular to a large model compression and real-time inference method and system based on knowledge distillation. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, large models are increasingly widely used in data analysis, business decision-making and other fields. Enterprises have an increasingly urgent demand for large models to support business operations and improve decision-making efficiency. Large model compression and real-time inference have become the key to enterprise application landing. As a core means connecting the general capabilities of large models and the demand for scenario-based lightweight, knowledge distillation technology plays an irreplaceable role in realizing "lightweight deployment + precise business support".

[0003] However, existing large model compression and inference solutions often face the following problems when dealing with industry individualization and enterprise customization needs. First, the industry adaptability of general large models is insufficient. Although existing general large models have cross-domain basic capabilities, they lack industry-specific knowledge. When directly applied to specific industries, the output results are inconsistent with the actual business scenarios of the industry, and the industry data characteristics are not optimized. Second, the enterprise-level model screening and evaluation system is not accurate. Existing model recommendations are mostly based on keyword matching or single historical success rate, without quantifying the gap between the model and the demand, leading to the problem that the selected model is not matched with the function and does not meet the standard when landing. SUMMARY

[0004] The purpose of the present application is to provide a large model compression and real-time inference method and system based on knowledge distillation to solve the problems in the prior art.

[0005] To achieve the above purpose, the present application provides the following technical solution: a large model compression and real-time inference method based on knowledge distillation, the method comprising the following steps:

[0006] Extracting the keywords of the general model and the industry model to form corresponding keyword sets;

[0007] Calculating the correlation coefficient of the general model and the industry model. When the correlation coefficient is not less than a preset threshold, it is determined that the two are highly correlated, and they are combined to form a general-industry model set and update the keywords;

[0008] Based on the industry model in the general-industry model set, the enterprise model set is formed combined with the business demand of the target enterprise, and the keywords of each enterprise model are extracted to construct the corresponding set; when a certain enterprise model is highly correlated with the general model and the industry model in the general-industry model set, it is added to form a general-industry-enterprise model set and update the keywords; when it is only highly correlated with one type of model, a sub-model set is constructed with the model and the keywords of the sub-model set are updated; wherein, the sub-model set represents the general-enterprise model set or the industry-enterprise model set;

[0009] Based on the general, industry, general-industry, general-industry-enterprise and sub-model sets that have been constructed, the first model set is integrated by screening the models combined with the demand keywords of the target enterprise, the keyword missing rate is defined, and the matching value of the model and the enterprise demand is calculated combined with the past similar project success rate of the model, the highest success rate of all candidate models and the demand coverage;

[0010] Based on the comparison between the models in the first model set and the ideal model, the defect and advantage feature sets are extracted, the similarity of each feature and the corresponding feature library is calculated and the highest value is taken, and then the defect comprehensive influence coefficient and the advantage comprehensive gain coefficient are calculated, wherein, the ideal model represents a virtual benchmark model that matches all demand keywords of the target enterprise and has the highest historical success rate;

[0011] The recommendation score is calculated based on the matching value, the defect comprehensive influence coefficient and the advantage comprehensive gain coefficient, the models are sorted from high to low according to the score, and the optimal model is pushed to the target enterprise.

[0012] The keywords of the general model and the industry model are extracted to form the corresponding keyword set, and the specific steps include:

[0013] Select a mainstream open source large model in the industry as the basic carrier, initialize all parameters, complete the basic training through general corpus, form a general model set with cross-domain general service capability; extract the keywords of the i-th general model L(0,i), construct the corresponding general model keyword set, denoted as {ki(0,1), ki(0,2),..., ki(0,K0)}; wherein, ki(0,1), ki(0,2),..., ki(0,K0) represent the 1st, 2nd,..., K0th keywords of the i-th general model, K0 represents the number of keywords of the i-th general model, L(0,i) represents the i-th general model, i=1, 2,..., I, I represents the number of general models formed;

[0014] Based on the general model L(0, i), the professional corpus of the target industry is combined to carry out industry data training, the general knowledge of the general model is migrated to the model architecture adapted to the industry scene through the knowledge distillation technology, and an industry model set for the industry knowledge field application is formed; keywords of the jth industry model L(1, j) are extracted, and a corresponding industry model keyword set is constructed, denoted as {kj(1, 1), kj(1, 2), …, kj(1, K1)}; wherein kj(1, 1), kj(1, 2), …, kj(1, K1) represent the 1st, 2nd, …, K1th keywords of the jth industry model respectively, K1 represents the number of keywords of the jth industry model, L(1, j) represents the jth industry model, j = 1, 2, …, J, and J represents the number of formed industry models.

[0015] The correlation coefficient of the general model and the industry model is calculated, when the correlation coefficient is not less than the preset threshold, it is determined that the two are highly correlated, and they are constituted into a general-industry model set and the keywords are updated, and the specific steps include:

[0016] The correlation coefficient between the ith general model and the jth industry model is calculated, which is defined as follows: R(0, 1) = s0 / s1; wherein s0 represents the number of intersection elements of the ith general model keyword set and the jth industry model keyword set, and s1 represents the number of union elements of the ith general model keyword set and the jth industry model keyword set;

[0017] If the correlation coefficient between the ith general model and the jth industry model is greater than or equal to the preset correlation threshold, at this time, it is judged that the ith general model and the jth industry model have high correlation, the ith general model and the jth industry model are constituted into a general-industry model set, and the keywords of the general-industry model set are updated;

[0018] If the correlation coefficient between the i-th general model and the j-th industry model is less than the preset correlation threshold, it is determined that the i-th general model and the j-th industry model do not have high correlation. At this time, the reasons for causing the correlation coefficient between the i-th general model and the j-th industry model to be less than the preset correlation threshold are analyzed, and the specific investigation directions include that the general model is not suitable for the target industry, the quality of the target industry professional corpus is not up to standard, there is an operation or logic error in the knowledge distillation process, and the model training target deviates from the actual demand of the industry. After secondary adjustment, if the correlation coefficient between the i-th general model and the j-th industry model is greater than or equal to the preset correlation threshold, it is determined that the i-th general model and the j-th industry model have high correlation. If the correlation coefficient between the i-th general model and the j-th industry model is less than the preset correlation threshold after secondary adjustment, the characteristics, parameter size and knowledge distillation characteristics of the i-th general model pre-training corpus are marked.

[0019] Based on the industry models in the general-industry model set, an enterprise model set is formed in combination with the business demand of the target enterprise, and the keywords of each enterprise model are extracted to construct a corresponding set; when a certain enterprise model is highly correlated with both the general models and the industry models in the general-industry model set, it is added to form a general-industry-enterprise model set and the keywords are updated; when it is only highly correlated with one type of model, a sub-model set is constructed with the type of model and the keywords of the sub-model set are updated; wherein the sub-model set represents a general-enterprise model set or an industry-enterprise model set, and the specific steps include:

[0020] Based on the industry models in the general-industry model set, an enterprise model set is formed in combination with the business demand of the target enterprise, and the keywords of each enterprise model are extracted to construct a corresponding set; when a certain enterprise model is highly correlated with both the general models and the industry models in the general-industry model set, it is added to form a general-industry-enterprise model set and the keywords are updated; when it is only highly correlated with one type of model, a sub-model set is constructed with the type of model and the keywords of the sub-model set are updated; wherein the sub-model set represents a general-enterprise model set or an industry-enterprise model set, and the specific steps include:

[0021] If the k-th enterprise model has high correlation with the i-th general model in the general-industry model set and has high correlation with the j-th industry model in the general-industry model set, the k-th enterprise model is added to the general-industry model set to form a general-industry-enterprise model set, and the keywords of the general-industry-enterprise model set are updated.

[0022] If the kth enterprise model has high correlation with the ith general model in the general-industry model set or with the jth industry model in the general-industry model set, a sub-model set is constructed with the kth enterprise model and the relevant model, and the keywords of the sub-model set are updated, wherein the relevant model represents the model with high correlation between the ith general model in the general-industry model set or the jth industry model in the general-industry model set and the kth enterprise model, and the sub-model set represents the general-enterprise model set or the industry-enterprise model set.

[0023] Based on the constructed general, industry, general-industry, general-industry-enterprise and sub-model sets, the target enterprise demand keywords are combined to screen the models into the first model set, the keyword missing rate is defined, the model and enterprise demand matching value is calculated in combination with the model past similar project success rate, the highest success rate of all candidate models and the demand coverage, the defect and advantage feature set is extracted based on the comparison between the models in the first model set and the ideal model, the highest value is taken by calculating the similarity of each feature and the corresponding feature library, and the defect comprehensive influence coefficient and the advantage comprehensive gain coefficient are calculated, wherein the ideal model represents a virtual benchmark model that matches all target enterprise demand keywords and has the highest historical success rate; the recommendation score is calculated based on the matching value, the defect comprehensive influence coefficient and the advantage comprehensive gain coefficient, the optimal model is pushed to the target enterprise in descending order of the score, and the specific steps include:

[0024] Based on the constructed general model set, industry model set, general-industry model set, general-industry-enterprise model set and sub-model set, the demand keywords of the target enterprise are matched, the models meeting the demand keyword requirements of the target enterprise are screened out from the corresponding model set, and the first model set is obtained by integration, wherein the zth model in the first model set is denoted as Lz(q), wherein Lz(q) represents the model meeting the demand keyword requirements of the target enterprise, and the number of keywords matching the demand keywords of the target enterprise is q, q=1, 2,..., Q, Q represents the number of demand keywords of the target enterprise, z=1, 2,..., Z, and Z represents the number of models meeting the demand keyword requirements of the target enterprise;

[0025] When the model Lz(q) is applied to the target enterprise, the matching effect will be affected by the unmatched keywords, and the keyword missing rate of the model is defined as M z =(Q-q) / Q;

[0026] The matching value of the model Lz(q) and the demand of the target enterprise is calculated according to the formula, and is defined as follows: F z =S z / S max ×(1-M z ); wherein S zrepresents the historical success rate of the model Lz(q) in the past similar enterprise projects, that is, the proportion of the number of projects successfully landed and achieved the expected effect to the total number of projects when the model is applied to the historical similar projects of the target enterprise; S max represents the highest historical success rate among all candidate models, that is, the maximum value of the proportion of projects successfully landed and achieved the expected effect from the historical performance of all candidate models;

[0027] Feature information is extracted from the model Lz(q) and the ideal model L(Q) respectively: the feature set D(z) of the model Lz(q) is extracted, including demand keyword semantic features, industry scene features and technical capability features, and the feature set D(Q) of the ideal model L(Q) is extracted, including demand keyword semantic features, industry scene features and technical capability features, as a benchmark feature set, wherein the ideal model L(Q) is defined as a virtual benchmark model matching Q target enterprise demand keywords and having the highest historical success rate;

[0028] The distinguishing feature information set F0(z)=D(Q)-D(Q)∩D(z) of the model Lz(q) is extracted;

[0029] The advantage feature set F1(z)=D(z)-D(Q)∩D(z) of the model Lz(q) relative to the ideal model is extracted;

[0030] The similarity of each feature information in the distinguishing feature information set F0(z) with each feature in the model defect feature library is calculated, and the highest similarity value is taken:

[0031] Suppose F0(z) contains B defect features: θ1, θ2,..., θ B ;

[0032] Each feature θ b calculates the similarity with the defect feature library and takes the highest value: max s (θ b );

[0033] The defect comprehensive influence coefficient of the model Lz(q) is calculated, which is defined as shown below:

[0034] ;

[0035] Wherein, α b is the weight coefficient of feature θ b ;

[0036] The similarity of each feature information in the advantage feature set F1(z) with each feature in the model advantage feature library is calculated, and the highest similarity value is taken:

[0037] Suppose F1(z) contains V advantage features: φ1, φ2,..., φV ;

[0038] Each feature φ v Calculate the similarity with the advantage feature library, and take the maximum value: max s (φ v );

[0039] Calculate the advantage comprehensive gain coefficient of the model Lz(q), defined as shown below:

[0040] ;

[0041] Where, β v is the weight coefficient of feature φ v ;

[0042] Determine the recommendation score, defined as shown below: G z =F z ×x z / y z ;

[0043] According to the recommendation score, sort the models that meet the demand keywords of the target enterprise according to the recommendation score from large to small, and preferentially select the model that meets the demand keywords of the target enterprise corresponding to the highest recommendation score to push to the target enterprise.

[0044] A large model compression and real-time inference system based on knowledge distillation, the system comprises a model construction module, a model integration module, a construction module and an evaluation module; the model construction module is used to extract general and industry model keywords; the model integration module is used to calculate the correlation coefficient of general and industry models, and if the standard is met, a general-industry model set is constructed and the keywords are updated; the construction module is used to form an enterprise model set based on the industry model of the set and the enterprise demand, and construct a general-industry- enterprise model or a sub-model set according to the correlation of the enterprise model with the general and industry models and update the keywords; the evaluation module is used to filter out a first model set from the constructed various models and sub-sets, define a missing rate, calculate a matching value in combination with the past success rate of the model, the highest success rate of the candidate model and the demand coverage, extract defects and advantage features by comparing with an ideal model, calculate a recommendation score, and sort and push the optimal model.

[0045] The model construction module comprises a general model construction unit and an industry model construction unit, which selects mainstream open source large models in the industry as a basic carrier and performs full parameter initialization; basic training is completed through general corpus to form a general model set with cross-domain general service capability; the keywords of each general model are extracted to construct a corresponding general model keyword set, and the general model construction unit migrates general knowledge through knowledge distillation based on the general model adapting to industry scenarios to extract industry model keywords.

[0046] The model integration module comprises a correlation coefficient calculation unit and a keyword updating unit, the correlation coefficient calculation unit is used for quantifying the correlation degree of the general model and the industry model, and the keyword updating unit is used for updating the keywords of the general-industry model set after integrating the high-correlation model.

[0047] The construction module comprises an enterprise model fine-tuning unit, an enterprise model correlation determination unit, a sub-model set construction unit and a model set keyword updating unit, the enterprise model fine-tuning unit is used for fine-tuning the general-industry model set in the general-industry model set in combination with the business demand of the target enterprise, forming an enterprise model set, extracting the keywords of each enterprise model, and constructing an enterprise model keyword set, the enterprise model correlation determination unit is used for determining the correlation of the enterprise model with the general model and the industry model in the general-industry model set, the sub-model set construction unit is used for constructing a general-enterprise model set when the enterprise model is highly correlated with the general model, and constructing an industry-enterprise model set when the enterprise model is highly correlated with the industry model, and the model set keyword updating unit updates the keywords of the general-industry-enterprise model set and the sub-set.

[0048] The evaluation module comprises a first model set screening unit, a keyword missing rate calculation unit, a matching value calculation unit, a comprehensive coefficient calculation unit and a model recommendation unit, the first model set screening unit is used for matching the general, industry, general-industry, general-industry-enterprise and sub-model set with the target enterprise demand keywords, screening out the required models and integrating, the keyword missing rate calculation unit is used for quantifying the enterprise demand gap not covered by the model, the matching value calculation unit is used for calculating the model matching value in combination with the historical performance and demand coverage, the comprehensive coefficient calculation unit is used for extracting the characteristics of the model and the ideal model, calculating the similarity with the feature library, calculating the defect comprehensive influence coefficient and the advantage comprehensive gain coefficient, and the model recommendation unit calculates the recommendation score based on the evaluation result, and pushes the optimal model to the enterprise after sorting.

[0049] Compared with the prior art, the present application has the following advantages:

[0050] 1、The present application constructs a quantitative model screening and evaluation system, quantifies the demand gap not covered by the model through the keyword missing rate, calculates the basic matching value in combination with the past success rate of the model in similar projects and the highest success rate of all candidate models, introduces the defect comprehensive influence coefficient and the advantage comprehensive gain coefficient to optimize the recommendation score, and compares the model characteristics with the ideal model; different from the prior art which only depends on the surface matching of keywords or single historical success rate and lacks gap quantitative analysis, the present application can evaluate the fit degree of the model and the enterprise demand, and avoid the problem that the selected model does not match the function and does not meet the standard after landing.

[0051] 2. This invention uses knowledge distillation technology to transfer cross-domain knowledge from a general model to an industry-specific model architecture. It combines this knowledge with target industry corpora for training and constructs a hierarchical model system of "general-industry-enterprise". The model's adaptability is optimized for industry data characteristics (such as professional terminology and business logic). Unlike existing technologies where general models lack industry-specific knowledge and are easily disconnected from industry scenarios when applied directly, this invention enables models to deeply integrate general capabilities with industry characteristics, meeting the personalized business needs of different industries. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the large model compression and real-time reasoning method based on knowledge distillation of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] like Figure 1 As shown, this invention provides a technical solution: a large model compression and real-time inference method based on knowledge distillation. This method includes the following steps:

[0055] Extract keywords from the general model and industry model to form corresponding keyword sets;

[0056] Calculate the correlation coefficient between the general model and the industry model. When the correlation coefficient is not lower than the preset threshold, the two are determined to be highly correlated, and they are combined into a general-industry model set and the keywords are updated.

[0057] Based on the industry models in the general-industry model set, and combined with the business needs of the target enterprise, an enterprise model set is formed. Keywords for each enterprise model are extracted and corresponding sets are constructed. When an enterprise model is highly correlated with both the general model and the industry model in the general-industry model set, it is added to form a general-industry-enterprise model set and its keywords are updated. When it is highly correlated with only one type of model, a sub-model set is constructed with that type of model, and the keywords of the sub-model set are updated. Here, the sub-model set represents either the general-enterprise model set or the industry-enterprise model set.

[0058] Based on the constructed general, industry, general-industry, general-industry-enterprise and sub-model sets, combined with the target enterprise demand keywords, the models are screened and integrated into the first model set, the keyword missing rate is defined, combined with the model past similar project success rate, the highest success rate of all candidate models and the demand coverage, the matching value of the model and the enterprise demand is calculated;

[0059] Based on the comparison between the models in the first model set and the ideal model, the defect and advantage feature set is extracted, the similarity of each feature and the corresponding feature library is calculated and the highest value is taken, and the defect comprehensive influence coefficient and the advantage comprehensive gain coefficient are calculated, wherein the ideal model represents a virtual benchmark model that matches all target enterprise demand keywords and has the highest historical success rate;

[0060] Based on the matching value, the defect comprehensive influence coefficient and the advantage comprehensive gain coefficient, the recommendation score is calculated, the scores are sorted from high to low, and the optimal model is pushed to the target enterprise.

[0061] The keywords of the general model and the industry model are extracted to form the corresponding keyword set, and the specific steps include:

[0062] A mainstream open source large model in the industry is selected as a basic carrier, and full parameter initialization is performed on it. The basic training is completed through general corpus to form a general model set with cross-domain general service capability. The keywords of the i-th general model L(0,i) are extracted to construct the corresponding general model keyword set, denoted as {ki(0,1), ki(0,2),..., ki(0,K0)}. Wherein, ki(0,1), ki(0,2),..., ki(0,K0) represent the 1st, 2nd,..., K0th keywords of the i-th general model, K0 represents the number of keywords of the i-th general model, L(0,i) represents the i-th general model, i=1, 2,..., I, I represents the number of formed general models;

[0063] Based on the general model L(0,i), combined with the professional corpus of the target industry, the industry data training is carried out, the general knowledge of the general model is transferred to the model architecture suitable for the industry scene through the knowledge distillation technology, and the industry model set for the industry knowledge field application is formed. The keywords of the j-th industry model L(1,j) are extracted to construct the corresponding industry model keyword set, denoted as {kj(1,1), kj(1,2),..., kj(1,K1)}. Wherein, kj(1,1), kj(1,2),..., kj(1,K1) represent the 1st, 2nd,..., K1th keywords of the j-th industry model, K1 represents the number of keywords of the j-th industry model, L(1,j) represents the j-th industry model, j=1, 2,..., J, J represents the number of formed industry models.

[0064] Correlation coefficients of the general model and the industry model are calculated, and when the correlation coefficient is not less than a preset threshold, it is determined that the two are highly correlated, and a general-industry model set is formed and the keywords are updated, including the following steps:

[0065] The correlation coefficient between the i-th general model and the j-th industry model is calculated, which is defined as follows: R(0, 1) = s0 / s1; wherein s0 represents the number of intersection elements of the i-th general model keyword set and the j-th industry model keyword set, and s1 represents the number of union elements of the i-th general model keyword set and the j-th industry model keyword set;

[0066] If the correlation coefficient between the i-th general model and the j-th industry model is greater than or equal to the preset correlation threshold, it is determined that the i-th general model and the j-th industry model have high correlation, and the i-th general model and the j-th industry model are formed into a general-industry model set, and the keywords of the general-industry model set are updated;

[0067] If the correlation coefficient between the i-th general model and the j-th industry model is less than the preset correlation threshold, it is determined that the i-th general model and the j-th industry model do not have high correlation, and the reason why the correlation coefficient between the i-th general model and the j-th industry model is less than the preset correlation threshold is analyzed, including the following specific investigation directions: the adaptability of the general model to the target industry is insufficient, the quality of the target industry professional corpus is not up to standard, there are operation or logic errors in the knowledge distillation process, and the model training target and the actual demand of the industry exist deviation, and secondary adjustment is carried out, if after the secondary adjustment, the correlation coefficient between the i-th general model and the j-th industry model is greater than or equal to the preset correlation threshold, it is determined that the i-th general model and the j-th industry model have high correlation; if after the secondary adjustment, the correlation coefficient between the i-th general model and the j-th industry model is less than the preset correlation threshold, the characteristics, parameter size and knowledge distillation characteristics of the i-th general model pre-training corpus are marked.

[0068] Based on the industry model in the general-industry model set, an enterprise model set is formed combining the target enterprise business demand, and the keywords of each enterprise model are extracted to construct a corresponding set; when a certain enterprise model is highly correlated with the general model and the industry model in the general-industry model set, it is added to form a general-industry-enterprise model set and the keywords are updated; when it is only highly correlated with one type of model, a sub-model set is constructed with the type of model and the keywords of the sub-model set are updated; wherein the sub-model set represents a general-enterprise model set or an industry-enterprise model set, including the following steps:

[0069] Based on the industry model in the general-industry model set, the business demand of the target enterprise is combined for directional fine-tuning to form an enterprise model set; the keywords of the kth enterprise model L(2, k) are extracted to construct the corresponding enterprise model keyword set, denoted as {kk(2, 1), kk(2, 2), …, kk(2, K2)}; wherein kk(2, 1), kk(2, 2), …, kk(2, K2) represent the 1st, 2nd, …, K2th keywords of the kth enterprise model respectively, K2 represents the number of keywords of the kth enterprise model, L(2, k) represents the kth enterprise model, k = 1, 2, …, K, and K represents the number of enterprise models formed;

[0070] If the kth enterprise model has high correlation with the ith general model in the general-industry model set and has high correlation with the jth industry model in the general-industry model set, the kth enterprise model is added to the general-industry model set to form a general-industry-enterprise model set, and the keywords of the general-industry-enterprise model set are updated;

[0071] If the kth enterprise model has high correlation with one of the ith general model in the general-industry model set or the jth industry model in the general-industry model set, a sub-model set is constructed with the kth enterprise model and the related model, and the keywords of the sub-model set are updated, wherein the related model represents the model that has high correlation with the kth enterprise model among the ith general model in the general-industry model set or the jth industry model in the general-industry model set, and the sub-model set represents the general-enterprise model set or the industry-enterprise model set.

[0072] Based on the general, industry, general-industry, general-industry-enterprise and sub-model sets that have been constructed, the target enterprise demand keywords are combined to screen and integrate the models into a first model set, the keyword missing rate is defined, the model and enterprise demand matching value is calculated in combination with the past similar project success rate of the model, the highest success rate of all candidate models and the demand coverage, the defect and advantage feature sets are extracted based on the comparison between the models in the first model set and the ideal model, the similarity of each feature and the corresponding feature library is calculated and the highest value is taken, and the defect comprehensive influence coefficient and the advantage comprehensive gain coefficient are calculated, wherein the ideal model represents a virtual benchmark model that matches all target enterprise demand keywords and has the highest historical success rate; the recommendation score is calculated based on the matching value, the defect comprehensive influence coefficient and the advantage comprehensive gain coefficient, the optimal model is pushed to the target enterprise in descending order of the score, and the specific steps include:

[0073] Based on the constructed general model set, industry model set, general-industry model set, general-industry-enterprise model set, and sub-model set, the model is matched with the target enterprise's demand keywords. Models that meet the target enterprise's demand keyword requirements are selected from the corresponding model sets and integrated to obtain the first model set. The z-th model in the first model set is denoted as Lz(q), where Lz(q) represents the z-th model that meets the target enterprise's demand keyword requirements, and the number of keywords that this model matches with the target enterprise's demand keywords is q, q=1,2,...,Q, where Q represents the number of target enterprise's demand keywords, and z=1,2,...,Z, where Z represents the number of models that meet the target enterprise's demand keyword requirements.

[0074] When the model Lz(q) is applied to the target enterprise, its matching effect will be affected by unmatched keywords. The keyword missing rate of the model is defined as M. z =(Qq) / Q;

[0075] The matching value between the model Lz(q) and the target company's needs is calculated using the formula, defined as follows: F z =S z / S max ×(1-M z ); where S z The Lz(q) model represents the historical success rate of similar enterprise projects, that is, the proportion of projects in which the model was successfully implemented and achieved the expected results when applied to similar projects in the target enterprise's past; S max This represents the highest historical success rate among all candidate models, that is, the maximum percentage of projects that have been successfully implemented and achieved the expected results based on the historical performance of all candidate models.

[0076] Feature information is extracted from model Lz(q) and ideal model L(Q) respectively: the feature set D(z) of model Lz(q) is extracted, including semantic features of demand keywords, industry scenario features and technical capability features; the feature set D(Q) of ideal model L(Q) is extracted, including semantic features of demand keywords, industry scenario features and technical capability features, as the benchmark feature set. The ideal model L(Q) is defined as the virtual benchmark model that matches the demand keywords of Q target enterprises and has the highest historical success rate.

[0077] Extract the set of discriminative feature information F0(z) = D(Q) - D(Q) ∩ D(z) from the model Lz(q);

[0078] Extract the set of advantageous features of model Lz(q) relative to the ideal model: F1(z) = D(z) - D(Q) ∩ D(z);

[0079] For each feature information in the distinguishing feature information set F0(z), the similarity with each feature in the model defect feature library is calculated, and the highest similarity value is taken:

[0080] Suppose that F0(z) contains B defect features: θ1, θ2,..., θ B ;

[0081] Each feature θ b calculates the similarity with the defect feature library, and takes the highest value: max s (θ b );

[0082] The defect comprehensive influence coefficient of the model Lz(q) is calculated, and is defined as shown below:

[0083] ;

[0084] Wherein, α b is the weight coefficient of feature θ b ;

[0085] For each feature information in the distinguishing feature information set F0(z), the similarity with each feature in the model defect feature library is calculated, and the highest similarity value is taken:

[0086] Suppose that F0(z) contains B defect features: θ1, θ2,..., θ V ;

[0087] Each feature φ v calculates the similarity with the defect feature library, and takes the highest value: max s (φ v );

[0088] The advantage comprehensive gain coefficient of the model Lz(q) is calculated, and is defined as shown below:

[0089] ;

[0090] Wherein, β v is the weight coefficient of feature φ v ;

[0091] Determine the recommended score, and is defined as shown below: G z =F z ×x z / y z ;

[0092] According to the recommended score, the models meeting the demand key word requirements of the target enterprise are sorted in descending order of the recommended score, and the model meeting the demand key word requirements of the target enterprise corresponding to the highest recommended score is preferentially selected and pushed to the target enterprise.

[0093] The system comprises a model construction module, a model integration module, a construction module and an evaluation module; the model construction module is used for extracting general and industry model keywords; the model integration module is used for calculating the correlation coefficient of general and industry models, and if the correlation coefficient meets the standard, a general-industry model set is constructed and the keywords are updated; the construction module is used for forming an enterprise model set based on the industry model set and combining enterprise demand, and constructing a general-industry-enterprise model or a sub-model set according to the correlation of the enterprise model with the general and industry models and updating the keywords; the evaluation module is used for screening a first model set from the constructed models and sub-sets, defining a missing rate, calculating a matching value in combination with the past success rate of the model, the highest success rate of the candidate model and the demand coverage, extracting defects and advantage features by comparing with an ideal model, calculating a recommendation score and sorting and pushing the optimal model.

[0094] The model construction module comprises a general model construction unit and an industry model construction unit, selects mainstream open source large models in the industry as a basic carrier, performs full parameter initialization, completes basic training through general corpus to form a general model set with cross-domain general service capability, extracts keywords of each general model to construct a corresponding general model keyword set, and the general model construction unit migrates general knowledge through knowledge distillation based on general model adaptation to industry scenarios to extract industry model keywords.

[0095] The model integration module comprises a correlation coefficient calculation unit and a keyword updating unit, the correlation coefficient calculation unit is used for quantifying the correlation degree of general models and industry models, and the keyword updating unit is used for updating the keywords of the general-industry model set after integrating high correlation models.

[0096] The construction module comprises an enterprise model fine-tuning unit, an enterprise model correlation determination unit, a sub-model set construction unit and a model set keyword updating unit, the enterprise model fine-tuning unit is used for fine-tuning the industry model in the general-industry model set in combination with the business demand of a target enterprise to form an enterprise model set, extracting keywords of each enterprise model to construct an enterprise model keyword set, the enterprise model correlation determination unit is used for determining the correlation of the enterprise model with the general model and the industry model in the general-industry model set, the sub-model set construction unit is used for constructing a general-enterprise model set when the enterprise model is highly correlated with the general model only, constructing an industry-enterprise model set when the enterprise model is highly correlated with the industry model only, and the model set keyword updating unit updates the keywords of the general-industry-enterprise model set and the sub-sets.

[0097] The evaluation module comprises a first model set screening unit, a keyword missing rate calculation unit, a matching value calculation unit, a comprehensive coefficient calculation unit and a model recommendation unit; the first model set screening unit is used for matching a general model, an industry model, a general-industry model, a general-industry-enterprise model and a sub-model set with a target enterprise demand keyword, screening out a model meeting the requirements and integrating; the keyword missing rate calculation unit is used for quantifying the enterprise demand gap not covered by the model; the matching value calculation unit is used for calculating the model matching value in combination with the historical performance and demand coverage; the comprehensive coefficient calculation unit is used for extracting the features of the model and the ideal model, calculating the similarity with the feature library, calculating the defect comprehensive influence coefficient and the advantage comprehensive gain coefficient; and the model recommendation unit calculates a recommendation score based on the evaluation result, and pushes the optimal model to the enterprise after sorting.

[0098] In embodiment 1: Select the mainstream open source large model in the industry as the basic carrier, after completing the full parameter initialization, carry out basic training with cross-domain general corpus - these corpora cover basic semantic analysis, data logic analysis, general trend prediction and other knowledge, finally form a general model with cross-domain data processing ability, then extract the core keywords related to data trend analysis and logical reasoning in the model, build a general model keyword set, then, based on the general model, combined with the professional corpus of the power energy industry, carry out industry data training, these professional corpora include power grid operation historical data, peak-valley power consumption records, wind power and photovoltaic power generation output data, energy storage equipment operation and maintenance information and other energy exclusive data, through the knowledge distillation technology, the general data analysis ability and trend prediction ability of the general model are migrated to the model architecture suitable for the power scene, forming an industry model for power energy data processing, at the same time, extracting the keywords related to power load forecasting and new energy grid connection regulation in the industry model, building an industry model keyword set;

[0099] Then carry out general-industry model integration work, judge the correlation degree of the above general model and the power industry model, compare the overlapping content and overall range of the keyword sets of the two types of models, evaluate the close correlation between the two, if the correlation reaches the preset standard, it is determined that the two are highly correlated, and they are integrated into a general-industry model set, the keywords of the two types of models are fused and the set content is updated to ensure that the set has general data processing ability and power industry characteristics; if the correlation does not meet the standard, investigate the reasons from four aspects: the adaptation degree of the general model to the power industry, the integrity and accuracy of the power professional corpus, the fit degree of the model training target and the actual demand of the power industry, for example, if it is found that the lack of wind power output data leads to insufficient correlation, the part of the corpus is supplemented and perfected, and then the knowledge distillation and correlation evaluation are carried out again until the standard is met; if it still does not meet the standard after multiple adjustments, record the pre-training corpus characteristics, parameter size and knowledge distillation key operation characteristics of the general model for reference for subsequent model optimization.

[0100] Next, the enterprise model and the sub-set are constructed, based on the power industry model in the general-industry model set, combined with the specific business needs of a regional energy enterprise, including regional power grid load accurate prediction, wind and photovoltaic grid stability regulation, energy storage device charging and discharging optimization, and other personalized needs, the industry model is fine-tuned, the enterprise model is formed, the keywords related to the characteristics of the regional power grid and the installed capacity of new energy in the enterprise model are extracted, the enterprise model keyword set is constructed, and the relevance of the enterprise model to the general model and the power industry model in the general-industry model set is judged: if it is highly relevant to both types of models, it is added to the general-industry model set to form a general-industry-enterprise model set, and the set of keywords is updated to cover enterprise-specific needs; if it is only highly relevant to the general model, the two are integrated into a general-enterprise sub-model set; if it is only highly relevant to the power industry model, it is integrated into an industry-enterprise sub-model set, and the keywords of the corresponding sub-set are updated at the same time;

[0101] Finally, model screening and recommendation are completed, based on the various types of models and sub-model sets that have been constructed, combined with the real-time business demand keywords of the regional energy enterprise, such as "load prediction response speed" and "new energy grid failure rate", the models that meet the demand are screened and integrated into the first model set, the demand gap that is not covered by each model is reflected by the keyword missing rate, combined with the past success of each model in serving similar energy enterprise projects and the highest historical success of all candidate models, the basic matching value of the model and the enterprise demand is calculated, taking the ideal model that "matches all the needs of the enterprise and has the highest historical success rate" as the benchmark, the defect feature set (features not covered by the ideal model) and the advantage feature set (unique features better than the ideal model) of each model are extracted, the similarity of the two types of features to the corresponding feature library is calculated and the highest value is taken, and then combined with the feature importance weight, the defect comprehensive influence coefficient and the advantage comprehensive gain coefficient are obtained, based on the matching value, the two coefficients, the recommendation score of each model is calculated, and the optimal model is pushed to the regional energy enterprise for regional power grid load real-time prediction and new energy grid dynamic regulation, realizing the lightweight deployment and real-time inference of large models in the power energy scene.

[0102] It is apparent to those skilled in the art that the application is not limited to the details of the foregoing exemplary embodiments, and that the application can be implemented in other concrete forms without departing from the spirit or essential characteristics of the application. Therefore, the embodiments should be considered exemplary and non-limiting, and the scope of the application is defined by the appended claims rather than the foregoing description, and all changes falling within the meaning and range of equivalents of the essential elements of the claims are intended to be encompassed by the application. Any reference signs in the claims should not be considered limiting of the claims to which they relate.

Claims

1. A large model compression and real-time inference method based on knowledge distillation, characterized in that: The method comprises the following steps: Extract the keywords of the general model and the industry model to form corresponding keyword sets respectively; Calculate the correlation coefficient of the general model and the industry model. When the correlation coefficient is not lower than a preset threshold, it is determined that the two are highly correlated, and they are formed into a general-industry model set and the keywords are updated; Based on the industry model in the general-industry model set, an enterprise model set is formed in combination with the business requirements of the target enterprise, and the keywords of each enterprise model are extracted to form corresponding sets; when a certain enterprise model is highly correlated with the general model and the industry model in the general-industry model set, it is added to form a general-industry-enterprise model set and the keywords are updated; when it is only highly correlated with one type of model, a sub-model set is formed with the model and the keywords of the sub-model set are updated; wherein the sub-model set represents the general-enterprise model set or the industry-enterprise model set; Based on the general, industry, general-industry, general-industry-enterprise and sub-model sets that have been formed, in combination with the demand keywords of the target enterprise, the models are screened and integrated into a first model set, the keyword missing rate is defined, the matching value of the model and the enterprise demand is calculated in combination with the past similar project success rate of the model, the highest success rate of all candidate models and the demand coverage; Based on the comparison between the models in the first model set and the ideal model, the defect and advantage feature sets are extracted, the similarity of each feature and the corresponding feature library is calculated and the highest value is taken, and then the defect comprehensive influence coefficient and the advantage comprehensive gain coefficient are calculated, wherein the ideal model represents a virtual benchmark model that matches all demand keywords of the target enterprise and has the highest historical success rate; The recommendation score is calculated based on the matching value, the defect comprehensive influence coefficient and the advantage comprehensive gain coefficient, the models are sorted from high to low according to the score, and the optimal model is pushed to the target enterprise.

2. The large model compression and real-time inference method based on knowledge distillation according to claim 1, characterized in that: The keywords of the general model and the industry model are extracted to form corresponding keyword sets respectively, and the specific steps include: Select a mainstream open source large model in the industry as a basic carrier, initialize all parameters, complete basic training through general corpus, form a general model set with cross-domain general service capability; extract the keywords of the i-th general model L(0,i), and construct the corresponding general model keyword set, denoted as {ki(0,1), ki(0,2),..., ki(0,K0)}; wherein ki(0,1), ki(0,2),..., ki(0,K0) represent the 1st, 2nd,..., K0th keywords of the i-th general model, K0 represents the number of keywords of the i-th general model, L(0,i) represents the i-th general model, i=1,2,...,I, I represents the number of formed general models; Based on the general model L(0, i), the professional corpus of the target industry is combined to carry out industry data training, the general knowledge of the general model is migrated to the model architecture suitable for the industry scene through the knowledge distillation technology, and the industry model set for the industry knowledge field application is formed; the keywords of the jth industry model L(1, j) are extracted, and the corresponding industry model keyword set is constructed, denoted as {kj(1, 1), kj(1, 2), …, kj(1, K1)}; wherein kj(1, 1), kj(1, 2), …, kj(1, K1) represent the 1st, 2nd, …, K1th keywords of the jth industry model respectively, K1 represents the number of keywords of the jth industry model, L(1, j) represents the jth industry model, j=1, 2, …, J, and J represents the number of formed industry models.

3. The large model compression and real-time inference method based on knowledge distillation according to claim 2, characterized in that: The correlation coefficient of the general model and the industry model is calculated, when the correlation coefficient is not less than the preset threshold, it is determined that the two are highly correlated, and they are constituted into a general-industry model set and the keywords are updated, the specific steps include: The correlation coefficient between the ith general model and the jth industry model is calculated, which is defined as follows: R(0, 1)=s0 / s1; wherein s0 represents the number of intersection elements of the ith general model keyword set and the jth industry model keyword set, and s1 represents the number of union elements of the ith general model keyword set and the jth industry model keyword set; If the correlation coefficient between the ith general model and the jth industry model is greater than or equal to the preset correlation threshold, at this time, it is judged that the ith general model and the jth industry model have high correlation, the ith general model and the jth industry model are constituted into a general-industry model set, and the keywords of the general-industry model set are updated; If the correlation coefficient between the ith general model and the jth industry model is less than the preset correlation threshold, it is judged that the ith general model and the jth industry model do not have high correlation, at this time, the reason why the correlation coefficient between the ith general model and the jth industry model is less than the preset correlation threshold is analyzed, and secondary adjustment is carried out, if the correlation coefficient between the ith general model and the jth industry model is greater than or equal to the preset correlation threshold after the secondary adjustment, it is judged that the ith general model and the jth industry model have high correlation; if the correlation coefficient between the ith general model and the jth industry model is less than the preset correlation threshold after the secondary adjustment, the characteristics, parameter size and knowledge distillation characteristics of the pre-training corpus of the ith general model are marked.

4. The large model compression and real-time inference method based on knowledge distillation according to claim 3, characterized in that: Based on the industry model in the general-industry model set, the enterprise model set is formed combined with the business demand of the target enterprise, and the keywords of each enterprise model are extracted to construct the corresponding set; When the enterprise model is highly correlated with both the general model and the industry model in the general-industry model set, it is added to form a general-industry-enterprise model set and the keywords are updated; when it is only highly correlated with one of the two types of models, a sub-model set is constructed with the model and the keywords of the sub-model set are updated; wherein the sub-model set represents the general-enterprise model set or the industry-enterprise model set, and the specific steps include: Based on the industry model in the general-industry model set, the enterprise model set is formed by directional fine-tuning combined with the business needs of the target enterprise; the keywords of the kth enterprise model L(2, k) are extracted to construct the corresponding enterprise model keyword set, denoted as {kk(2, 1), kk(2, 2),..., kk(2, K2)}; wherein kk(2, 1), kk(2, 2),..., kk(2, K2) represent the 1st, 2nd,..., K2th keywords of the kth enterprise model respectively, K2 represents the number of keywords of the kth enterprise model, L(2, k) represents the kth enterprise model, and k = 1, 2,..., K, K represents the number of enterprise models formed; If the kth enterprise model has high correlation with the ith general model in the general-industry model set and high correlation with the jth industry model in the general-industry model set, the kth enterprise model is added to the general-industry model set to form a general-industry-enterprise model set, and the keywords of the general-industry-enterprise model set are updated; If the kth enterprise model has high correlation with one of the ith general model in the general-industry model set or the jth industry model in the general-industry model set, a sub-model set is constructed with the kth enterprise model and the related model, and the keywords of the sub-model set are updated, wherein the related model represents the model that has high correlation with the kth enterprise model among the ith general model in the general-industry model set or the jth industry model in the general-industry model set, and the sub-model set represents the general-enterprise model set or the industry-enterprise model set.

5. The large model compression and real-time inference method based on knowledge distillation according to claim 4, characterized in that: Based on the constructed general, industry, general-industry, general-industry-enterprise and sub-model sets, combined with the demand keywords of the target enterprise, the models are screened and integrated into the first model set, the keyword missing rate is defined, the model matching value with the enterprise demand is calculated based on the model success rate of the past similar projects, the highest success rate of all candidate models and the demand coverage, the defect and advantage feature sets are extracted based on the comparison between the models in the first model set and the ideal model, the similarity of each feature with the corresponding feature library is calculated and the highest value is taken, and then the defect comprehensive influence coefficient and the advantage comprehensive gain coefficient are calculated, wherein the ideal model represents a virtual benchmark model that matches all demand keywords of the target enterprise and has the highest historical success rate; the recommendation score is calculated based on the matching value, the defect comprehensive influence coefficient and the advantage comprehensive gain coefficient, the models are sorted from high to low according to the score, and the optimal model is pushed to the target enterprise, and the specific steps include: Based on the constructed general model set, industry model set, general-industry model set, general-industry-enterprise model set and sub-model set, the demand keywords of the target enterprise are matched, the models meeting the demand keyword requirements of the target enterprise are screened out from the corresponding model set, and the first model set is obtained by integration, the zth model in the first model set is denoted as Lz(q), wherein Lz(q) represents that the zth model meets the demand keyword requirements of the target enterprise, and the number of keywords matched with the demand keywords of the target enterprise is q, q = 1, 2,..., Q, Q represents the number of demand keywords of the target enterprise, and z = 1, 2,..., Z, Z represents the number of models meeting the demand keyword requirements of the target enterprise; When the model Lz(q) is applied to the target enterprise, the matching effect will be affected by the unmatched keywords, and the missing rate of the keywords defining the model is defined as M z = (Q - q) / Q; According to the formula, the matching value of the model Lz(q) with the target enterprise demand is calculated, which is defined as follows: F z =S z / S max ×(1-M z ); wherein, S z represents the historical success rate of the model Lz(q) in the past similar enterprise projects, that is, the proportion of the number of projects successfully landed and achieved the expected effect to the total number of projects when the model is applied to the historical similar projects of the target enterprise; S max represents the highest historical success rate among all candidate models, that is, the maximum value of the proportion of projects successfully landed and achieved the expected effect from the historical performance of all candidate models; Feature information is extracted from the model Lz(q) and the ideal model L(Q) respectively: the feature set D(z) of the model Lz(q) is extracted, including demand keyword semantic features, industry scene features and technical capability features, and the feature set D(Q) of the ideal model L(Q) is extracted, including demand keyword semantic features, industry scene features and technical capability features, as a benchmark feature set, wherein the ideal model L(Q) is defined as a virtual benchmark model matching Q demand keywords of the target enterprise and having the highest historical success rate; The distinguished feature information set F0(z) = D(Q)-D(Q)∩D(z) is extracted from the model Lz(q); The advantage feature set F1(z) = D(z)-D(Q)∩D(z) of the model Lz(q) relative to the ideal model is extracted; The similarity of each feature information in the distinguished feature information set F0(z) with each feature in the model defect feature library is calculated, and the highest similarity value is taken: Let F0(z) contain B defective features: θ1, θ2,..., θB. B ; Each feature θ b Take the highest value after calculating the similarity with the defect feature library: max s (θ b ); The defect comprehensive influence coefficient of the model Lz(q) is calculated, which is defined as shown below: ； wherein a b is a weight coefficient of the feature θ b . The similarity of each feature information in the advantage feature set F1(z) with each feature in the model advantage feature library is calculated, and the highest similarity value is taken: Let F1(z) contain V dominant features: φ1, φ2,..., φ V ; Each feature φ v Take the highest value after calculating the similarity with the advantage feature library: max s (φ v ); The advantage comprehensive gain coefficient of the model Lz(q) is calculated, which is defined as shown below: ； where β v is a weight coefficient of the feature φ v . A recommendation score is determined, defined as shown below: G z = F z x x z / y z ; According to the recommendation score, the models meeting the demand keyword requirements of the target enterprise are sorted in descending order of the recommendation score, and the model corresponding to the highest recommendation score is preferentially selected and pushed to the target enterprise.

6. The large model compression and real-time inference system based on knowledge distillation according to any one of claims 1-5, characterized in that: The system comprises a model construction module, a model integration module, a construction module and an evaluation module; the model construction module is used to extract general and industry model keywords; the model integration module is used to calculate the correlation coefficient of general and industry models, and if the standard is met, a general-industry model set is constructed and the keywords are updated; the construction module is used to form an enterprise model set based on the industry model of the set, combined with enterprise demand, and to construct a general-industry-enterprise model or a sub-model set and update the keywords according to the correlation of the enterprise model with general and industry models; the evaluation module is used to filter out a first model set from the constructed various models and sub-sets, combined with enterprise demand keywords, define the missing rate, calculate the matching value combined with the past success rate of the model, the highest success rate of the candidate model and the demand coverage, extract defects and advantage features compared with ideal models, calculate the recommendation score, and sort and push the optimal model.

7. The large model compression and real-time inference system based on knowledge distillation according to claim 6, characterized in that: The model construction module comprises a general model construction unit and an industry model construction unit, which selects mainstream open source large models in the industry as a basic carrier and performs full parameter initialization; completes basic training through general corpus to form a general model set with cross-domain general service capability; extracts keywords of each general model to construct a corresponding general model keyword set; the general model construction unit migrates general knowledge through knowledge distillation based on general model adaptation to industry scenarios to extract industry model keywords.

8. The large model compression and real-time inference system based on knowledge distillation according to claim 7, characterized in that: The model integration module comprises a correlation coefficient calculation unit and a keyword updating unit; the correlation coefficient calculation unit is used to quantify the correlation degree of general models and industry models; and the keyword updating unit is used to update the keywords of the general-industry model set after integrating high correlation models.

9. The knowledge distillation based large model compression and real-time inference system according to claim 8, characterized in that: The construction module comprises an enterprise model fine-tuning unit, an enterprise model correlation determination unit, a sub-model set construction unit and a model set keyword updating unit; the enterprise model fine-tuning unit is used to perform directional fine-tuning based on the industry model in the general-industry model set combined with the business demand of the target enterprise to form an enterprise model set; extract the keywords of each enterprise model to construct an enterprise model keyword set; the enterprise model correlation determination unit is used to determine the correlation of the enterprise model with the general model and the industry model in the general-industry model set; the sub-model set construction unit is used to construct a general-enterprise model set when the enterprise model is only highly correlated with the general model, and to construct an industry-enterprise model set when the enterprise model is only highly correlated with the industry model; and the model set keyword updating unit updates the keywords of the general-industry-enterprise model set and the sub-set.

10. The large model compression and real-time inference system based on knowledge distillation according to claim 9, characterized in that: The evaluation module comprises a first model set screening unit, a keyword missing rate calculation unit, a matching value calculation unit, a comprehensive coefficient calculation unit and a model recommendation unit; the first model set screening unit is used for matching general, industry, general-industry, general-industry-enterprise and sub-model sets with target enterprise demand keywords, screening out required models and integrating; the keyword missing rate calculation unit is used for quantifying enterprise demand gaps not covered by the model; the matching value calculation unit is used for calculating the model matching value in combination with historical performance and demand coverage; the comprehensive coefficient calculation unit is used for extracting the characteristics of the model and the ideal model, calculating the similarity with the feature library, calculating the defect comprehensive influence coefficient and the advantage comprehensive gain coefficient; and the model recommendation unit calculates a recommendation score based on the evaluation result, and pushes the optimal model to the enterprise after sorting.

Images