Digital platform for driving industrial supply chain by department and wound supply chain

Abstract

The invention provides a digital platform for driving an industrial supply chain by an industrial supply chain, which comprises the steps of policy recommendation strategy: policy tagging processing: performing word segmentation, part-of-speech tagging and syntactic analysis on a policy file by using an NLP technology in policy text analysis, and extracting key elements; enterprise portrait modeling: constructing enterprise feature vectors through the collected data; and an intelligent recommendation algorithm. Through the above model system, full-process digital management from policy intelligent matching, industrial chain relation mining to supply and demand dynamic scheduling is realized, a closed-loop ecological system of data driving-model decision-scene landing is formed, and the fusion efficiency of a department and wound supply chain and an industrial supply chain is effectively improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of a digital platform, in particular to a digital platform for driving an industrial supply chain by a scientific and technological innovation supply chain. BACKGROUND

[0002] Science and technology projects are mainly disclosed on government websites, but the depth of policy guidance is not enough now, and the reporting process is not clear. When reporting projects, it is impossible to provide help to enterprises themselves, so as to achieve the purpose of reverse driving of supply chain matching by reporting data.

[0003] AI technology is now increasingly mature, and the integration of AI technology into the process of landing science and technology projects can greatly improve work efficiency and achieve unexpected results. SUMMARY

[0004] The technical problem to be solved by the application is to provide a digital platform for driving an industrial supply chain by a scientific and technological innovation supply chain, which facilitates the matching of projects by enterprises and can improve the supply chain in the opposite direction.

[0005] To solve the above technical problems, the application provides the following technical scheme:

[0006] A digital platform for driving an industrial supply chain by a scientific and technological innovation supply chain comprises:

[0007] Policy recommendation strategy:

[0008] Policy tagging processing: the policy text analysis uses NLP technology to perform word segmentation, part-of-speech tagging and syntax analysis on policy files, and extracts key elements;

[0009] Enterprise portrait modeling: an enterprise feature vector is constructed through collected data;

[0010] Intelligent recommendation algorithm:

[0011] Step 1: the similarity calculation adopts cosine similarity to calculate the matching degree of the enterprise and the policy:

[0012]

[0013] Wherein E is the enterprise feature vector, and P is the policy label vector;

[0014] Step 2: collaborative filtering prediction based on the historical reporting of similar enterprises, the recommendation score is predicted by using matrix decomposition:

[0015]

[0016] Wherein Q p is the policy latent vector, P u is the enterprise latent vector, and mu is the global average score, b u、b p bias for businesses and policies;

[0017] Step 3: Generate a recommendation list by sorting by matching degree to generate a Top-N policy recommendation list.

[0018] Furthermore, the policy labeling process includes the following steps:

[0019] Step 1: Policy text parsing uses NLP technology to perform word segmentation, part-of-speech tagging, and syntactic analysis on policy documents to extract key elements;

[0020] Step 2: Multi-dimensional label generation to build a three-layer label system:

[0021] Policy labeling system

[0022] Basic tags: release time, policy type;

[0023] Condition tags: industry restrictions, enterprise size, number of patents;

[0024] Benefit label: subsidy amount, tax preferential ratio;

[0025] Step 3: Tag weight calculation uses the TF-IDF algorithm to calculate tag importance:

[0026]

[0027] Among them: tf t,d is the word frequency of label t in policy d, df t is the number of policies containing label t, and N is the total number of policies.

[0028] Furthermore, enterprise portrait modeling includes

[0029] Step 1: Data collection and integration: Enterprise reported data: revenue, R&D investment, number of patents, industrial and commercial data, bidding records;

[0030] Step 2: Feature engineering to construct enterprise feature vector: E = [e1, e2, ..., e n ], where e i is the i-th feature: the proportion of R&D investment;

[0031] Step 3: Dynamic weight adjustment automatically adjusts feature weights according to policy type, w i '=w i ×α p

[0032] where α p is the characteristic adjustment coefficient corresponding to policy p.

[0033] Further, the digital platform further comprises an industry chain knowledge graph construction model, and the industry chain knowledge graph construction model comprises:

[0034] Step 1: enterprise entity recognition and enterprise information extraction;

[0035] Step 2: relationship mining and construction;

[0036] Wherein:

[0037] In the first step, multiple information of enterprises is collected for data integration;

[0038] In the second step, a four-layer relationship system is constructed.

[0039] Further, the enterprise entity recognition and enterprise information extraction comprise

[0040] Step 1: Multi-source data integration

[0041] Collect enterprise declaration materials, business information, bidding data, and patent literature;

[0042] Step 2: Named Entity Recognition

[0043] Three types of entities are identified using a BiLSTM+CRF model:

[0044] Entity type

[0045] Enterprise entity

[0046] Product entity

[0047] Technical entity;

[0048] Step 3: Entity Disambiguation

[0049] A clustering algorithm is used to solve the problem of entities with the same name:

[0050]

[0051] Wherein sim(e i ,e j ) is the similarity of entity vectors, and entities with a distance less than a threshold are merged.

[0052] Further, step 1:

[0053] Four-layer relationship system: supply and demand relationship, technology association, industry chain relationship, and policy adaptation relationship;

[0054] Step 2: Relationship Extraction

[0055] The algorithm uses a neural network model with remote supervision and attention mechanism, inputs a sentence fragment S, and outputs a relationship probability:

[0056] P(r|S) = softmax(W*tanh(U*eS + V*er)

[0057] where e S is the sentence embedding vector, e r is the relationship type embedding vector.

[0058] Step 3: Relationship weight calculation calculates the relationship strength according to transaction frequency, cooperation duration, etc.:

[0059] w r = alpha * f + beta * t + gamma * s

[0060] Where f is the transaction frequency, t is the cooperation duration, s is the transaction size, and alpha + beta + gamma = 1.

[0061] Graph dynamic updating mechanism

[0062] Step 1: Data monitoring real-time crawling of enterprise business changes, bidding announcements, policy updates, etc.:

[0063] Step 2: Change identification uses a difference algorithm to identify data changes:

[0064] Delta E = E new - Eold

[0065] Where E new and E old are the new and old enterprise entity attribute vectors.

[0066] Step 3: Graph updating strategy uses incremental updating:

[0067] New entity: add entity nodes and initialize relationships;

[0068] Entity attribute change: update node attributes and recalculate relationship weights;

[0069] Relationship change: update relationship strength or delete invalid relationships.

[0070] Compared with the prior art, the present application has at least the following beneficial effects:

[0071] Through the above model system, the whole process of digital management from policy intelligent matching, industry chain relationship mining to supply and demand dynamic scheduling is realized, forming a closed-loop ecological system of "data-driven-model decision-scene landing", effectively improving the fusion efficiency of scientific and technological supply chain and industry supply chain. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 It is a principle diagram of the embodiment one of the present application.

[0073] Figure 2A schematic diagram of the optimization process for the map construction of the present application. DETAILED DESCRIPTION

[0074] A digital platform and method for driving industrial supply chains with science and technology supply chains will be described in detail below in conjunction with specific embodiments. It should be noted here that in order to make the embodiments more detailed, the following embodiments are the best, preferred embodiments, and other alternative ways can also be used by those skilled in the art to implement them; and the drawings are only used to more specifically describe the embodiments and are not intended to specifically limit the present application.

[0075] Embodiment one:

[0076] Now the difficulties in policy propaganda:

[0077] 1. The last mile problem of policy landing: solve the problem of enterprises "can't find, can't understand, can't report" policy, and realize inclusive service through AI tools. 2. Industrial chain resource mismatch: break down the information barriers between enterprises, and use reporting data to drive supply chain matching in reverse. 3. Insufficient competitiveness of small and medium-sized enterprises: through policy empowerment and supply chain synergy, help small and medium-sized enterprises enter the supply chain system of large enterprises. 4. Government industrial governance efficiency improvement: provide local governments with dynamic industrial map dashboard, and research and judge the weak links of industrial chain based on enterprise reporting data.

[0078] Targeting the current large model to solve the above problems (1) Intelligent policy matching and reporting system (core entry product)

[0079] 1. Product function policy radar: real-time monitoring of science and technology policies at all levels, automatic pushing of policies with a matching degree of more than 90% through AI algorithm. Reporting navigation: provide "policy examination → material preparation → online reporting → progress tracking" whole process service. Success rate prediction: based on historical reporting case big data, intelligently predict the success rate of various policy reporting.

[0080] 2. Special services Enterprise digital portrait: automatically generate a digital report containing various indicators, and export reporting materials with one key. Policy sand table deduction: simulate the reporting effect of different policy combinations, and optimize the reporting strategy.

[0081] (2) Industrial chain synergy cloud platform (core value-added product)

[0082] Three function modules Demand market: intelligently analyze the procurement / technology demand in enterprise reporting materials; visualize the industrial chain heat map (supply-demand gap, technology bottleneck, etc.). Intelligent connection: accurate recommendation based on knowledge graph; blockchain electronic contract template library.

[0083] (3) Science and technology supply chain digital middle platform (bottom support product)

[0084] 1 Four core components Policy knowledge graph engine: build policy industry enterprise correlation network. Supply chain intelligent brain: demand forecasting + capacity scheduling algorithm model. Credible transaction storage system: interface with provincial and municipal government blockchain platform. API open platform: provide 12 types of standardized interfaces for ecological partners to call.

[0085] (Four) Industry governance SaaS tool (government end product) 1. Core function industry map: dynamically monitor the completeness and vulnerability of key industries. Policy simulation: predict the impact of policy adjustment on the industry chain. Investment navigation: intelligently recommend chain-strengthening target enterprises.

[0086] Key technologies (one) Intelligent policy matching engine 1. Technology stack: NLP (natural language processing) + knowledge graph + recommendation algorithm 2. Technical content Policy labeling: Analyze technology policy documents through NLP to build a multi-dimensional label system (such as industry, scale, qualification requirements, etc.). Enterprise portrait modeling: Generate dynamic portraits based on enterprise fill-in data (revenue, patents, R&D investment, etc.) to match applicable policies. Intelligent recommendation: Use collaborative filtering algorithm, refer to similar enterprise application success cases, and push high-pass rate policy combinations.

[0087] (Two) Industry chain knowledge graph 1. Technology stack: graph database + entity recognition + relationship mining 2. Technical content Supply chain relationship mining: Analyze technical / purchasing needs in enterprise application materials and automatically associate platform vendors; Build a "product technology enterprise" three-layer graph to support intelligent recommendation of upstream and downstream. Dynamic update mechanism: Update graph relationships in real time through enterprise business changes, bidding, and other public data.

[0088] (Three) Blockchain credible storage 1. Technology stack: Hyperledger Fabric + smart contract 2. Technical content Qualification storage: Store enterprise intellectual property, test reports, and other key qualifications on the chain for policy review parties to directly access and verify. Contract automation: Code supply chain cooperation agreement terms to automatically execute phased payments when conditions are met; Policy incentive redemption process is on the chain to ensure that subsidies can be traced.

[0089] (Four) Supply and demand forecasting and scheduling model

[0090] 1. Technology stack: time series forecasting + operations optimization

[0091] 2. Technical content Demand forecasting: Analyze policy application hotspots to predict industry chain demand explosion points; Combine historical procurement data to predict raw material / technology demand fluctuations. Resource scheduling: When multiple enterprises apply for similar technological transformation projects at the same time, intelligently recommend staggered procurement solutions or alternative suppliers.

[0092] (Five) Multi-modal interaction middleware

[0093] 1. Technology Stack: Microservices + RPA + Low-Code 2. Technical Content: Government-Enterprise Interface: Automatically captures policy updates from government websites and uses RPA robots to complete enterprise qualification pre-qualification and reporting. Enterprise: Low-code tools support rapid deployment of industry-specific scenarios. Service Closed Loop: Digitalizes the entire process from policy matching to supply chain integration to application progress tracking and impact evaluation.

[0094] Core technology model construction and implementation

[0095] Step 1: Intelligent Policy Matching Engine Model

[0096] (1) Model architecture and steps

[0097] Policy labeling

[0098] Step 1: Policy text parsing uses NLP technology to perform word segmentation, part-of-speech tagging, and syntactic analysis on policy documents to extract key elements (such as subsidy amount, industry requirements, qualification thresholds, etc.).

[0099] Step 2: Multi-dimensional label generation to build a three-layer label system:

[0100] Policy labeling system

[0101] ─ Basic tags (release time, policy type)

[0102] ├─Conditional tags (industry restrictions, enterprise size, number of patents)

[0103] └─Benefit label (subsidy amount, tax incentive ratio)

[0104] Step 3: Tag weight calculation uses the TF-IDF algorithm to calculate tag importance:

[0105]

[0106] Among them: tf t,d is the word frequency of label t in policy d, df t is the number of policies containing label t, and N is the total number of policies.

[0107] Enterprise portrait modeling

[0108] Step 1: Data collection and integration of enterprise reported data (revenue, R&D investment, number of patents), industrial and commercial data, and bidding records.

[0109] Step 2: Feature engineering to construct enterprise feature vector: E = [e1, e2, ..., e n ], where e i is the i-th feature (such as the proportion of R&D investment).

[0110] Step 3: Dynamic weight adjustment automatically adjusts feature weights according to policy types, for example: w i ' = w i x a p where a p is the feature adjustment coefficient corresponding to policy p (such as the patent quantity weight coefficient a p = 1.5 in the high-tech enterprise identification policy).

[0111] Intelligent recommendation algorithm

[0112] Step 1: Similarity calculation uses cosine similarity to calculate the matching degree between enterprises and policies:

[0113]

[0114] where E is the enterprise feature vector and P is the policy label vector.

[0115] Step 2: Collaborative filtering prediction based on the history of similar enterprises applying for the policy, using matrix decomposition to predict the recommendation score:

[0116]

[0117] where Q p is the policy latent vector, P u is the enterprise latent vector, μ is the global average rating, b u and b p are the enterprise and policy biases.

[0118] Step 3: Generate Top-N policy recommendation list by sorting by matching degree.

[0119] Implementation case

[0120] A certain biological and pharmaceutical enterprise A has a research and development investment ratio of 15% (feature weight 0.3) and owns 5 invention patents (feature weight 0.4), and needs to match the research and development subsidy policy: 1. The label vector of policy P is [biological and pharmaceutical, research and development investment ≥10%, invention patent ≥3, subsidy 500,000]

[0121] Calculate the cosine similarity between the enterprise feature vector and the policy label vector:

[0122]

[0123] Collaborative filtering prediction shows that the success rate of similar enterprises applying for the policy is 85%, and the policy is finally recommended.

[0124] II. Construction model of industry chain knowledge graph

[0125] Graph construction steps

[0126] 1. Business entity recognition and business information extraction

[0127] Step 1: Multi-source data integration

[0128] Collect business declaration materials, business information, bidding data, and patent literature

[0129] Step 2: Named entity recognition

[0130] Use BiLSTM+CRF model to identify three types of entities:

[0131] Entity type

[0132] (Huawei, Tencent)

[0133] (chip, sensor)

[0134] (AI, blockchain)

[0135] Step 3: Entity disambiguation

[0136] Use clustering algorithm to solve the problem of entities with the same name:

[0137]

[0138] Where sim(e i ,e j ) is the similarity of entity vectors, and entities with a distance less than the threshold are merged.

[0139] 2. Relationship mining and construction

[0140] Step 1: Relationship type definition

[0141] Construct a four-layer relationship system:

[0142] Relationship system

[0143] (purchasing, supply)

[0144] (R&D cooperation, patent citation)

[0145] (upstream, downstream, same layer)

[0146] (suitable policies, declaration records)

[0147] Step 2: Relationship extraction

[0148] The algorithm uses a neural network model with remote supervision and attention mechanism, inputting sentence fragments S and outputting relationship probability:

[0149] P(r|S) = softmax(W · tanh(U · eS + V · er)

[0150] where e S is the sentence embedding vector, e r is the relation type embedding vector.

[0151] Step 3: Relationship weight calculation calculates the relationship strength according to transaction frequency, cooperation duration, etc.

[0152] w r = a · f + b · t + g · s

[0153] where f is the transaction frequency, t is the cooperation duration, s is the transaction size, and a + b + g = 1.

[0154] 3. Dynamic updating mechanism of graph

[0155] Step 1: Data monitoring real-time crawling of enterprise business changes, bidding announcements, policy updates, etc.

[0156] Step 2: Change identification uses difference algorithm to identify data changes, such as:

[0157] AE = E new - E old

[0158] where E new and E old are the new and old enterprise entity attribute vectors.

[0159] Step 3: Graph updating strategy adopts incremental updating:

[0160] New entity: add entity node and initialize relationship

[0161] Entity attribute change: update node attributes and recalculate relationship weight

[0162] Relationship change: update relationship strength or delete invalid relationship.

[0163] Implementation case: building a medical device industry chain knowledge graph:

[0164] 1. Identify entities: a company, CT scanner (product), image recognition technology (technology)

[0165] 2. Extract relationships: a company → supply → CT scanner, CT scanner → rely on → image recognition technology

[0166] 3. Calculate relationship weight: - supply relationship: transaction frequency f = 3 times / year (a = 0.4 → weight 0.16)

[0167] Technology dependency: patent citation times t = 5 times (β = 0.3 → weight 0.15) - comprehensive weight w = 0.31. When a new image recognition technology patent is added to a certain enterprise, trigger graph update, recalculate the weight of technology dependency to 0.42

[0168] III. Supply and demand forecasting and scheduling optimization model

[0169] (I) Model core

[0170] 1. Demand forecasting module

[0171] Step 1: Data preprocessing Clean historical procurement data and handle missing values:

[0172]

[0173] Where is the i-th prediction value, and k is the size of the sliding window.

[0174] Step 2: Time series prediction model uses ARIMA(p, d, q) model:

[0175] (1-φ1B-...-φ p B p )(1-B) d X t =θ0+(θ1B+...+θ q B q )∈ t

[0176] Where B is the lag operator, and ∈ t is a white noise sequence.

[0177] Step 3: Policy influence correction Introduce policy correction factor:

[0178]

[0179] Where λ i is the i-th policy influence coefficient, and δ i,t is the policy change range at time t

[0180] 3. Resource scheduling optimization

[0181] Step 1: Problem modeling Build a multi-objective optimization model:

[0182]

[0183] Constraint conditions:

[0184]

[0185] Where c i,jis the transportation cost from supplier i to enterprise j, x i,j is the transport volume, s j is the inventory cost, y j is a binary variable indicating whether supplier j is enabled, d j For the needs of enterprise j, M i,j is the maximum supply.

[0186] Step 2: Algorithmic solution uses a mixed integer programming solver (such as CPLEX) or a heuristic algorithm (such as a genetic algorithm).

[0187] Step 3: The scheduling plan generates outputs including the optimal plan for procurement quantity, supplier selection, and logistics routing. Implementation Example: Scheduling silicon material demand for a photovoltaic enterprise cluster in a certain region.

[0188] 1. Historical data shows that Q3 silicon material demand follows the ARIMA (2,1,1) model:

[0189] X t =0.6X t-1 -0.2X t-2 +∈ t +0.4∈ t-1

[0190] The forecast base demand for Q3 is 1,200 tons.

[0191] 2. Policy revision: National subsidies are reduced by 15% (δ = -0.15), and the photovoltaic industry policy impact coefficient λ = 0.8D t '=1200×(1+0.8×(-0.15))=1056 tons.

[0192] 3. Scheduling optimization model: - Objective: Minimize procurement cost + inventory cost - Constraint: The maximum supply quantities of the three suppliers are 500 tons, 600 tons, and 400 tons respectively - Solution: 500 tons are purchased from supplier 1 and 556 tons are purchased from supplier 2, reducing the total cost by 8.7%.

[0193] 4. Output plan: Purchase 500 tons in July and 556 tons in August, and use staggered transportation strategies to reduce logistics costs.

[0194] 4. Model Integration and Interaction Process

[0195] (I) Cross-model collaboration steps (II) 1. The data interaction layer builds a unified data middle platform to realize data sharing between models.

[0196] 2. The decision-making linkage layer establishes a trigger mechanism between models:

[0197] Policy matching results → Update policy adaptation relationships in the knowledge graph

[0198] Supply and demand relationships in the knowledge graph → Optimize the input of the supply and demand forecasting model

[0199] Supply and demand forecast results → recommended strategies for adjusting policy matching

[0200] 3. Feedback optimization layer builds a closed-loop optimization mechanism: implementation effect → model parameter adjustment → prediction accuracy improvement. Reinforcement learning algorithm is used to dynamically optimize model weights: Where θ is the model parameter, α is the learning rate, and r t is the reward value, J(θ) is the objective function

[0201] (3) Full-process application examples

[0202] 1. Enterprise declaration scenario - the enterprise submits policy declaration materials - the policy matching model calculates and recommends policies (matching degree 0.85) - the knowledge graph model mines supply chain demand (such as the need for silicon wafer suppliers) - the supply and demand forecasting model predicts material demand fluctuations (demand decreases by 12% after correction) - and generates an integrated "policy declaration + supply chain procurement" solution.

[0203] 2. Government decision-making scenarios - issuing new policies (such as raising environmental protection standards) - policy matching models updating labeling systems - knowledge graph models recalculating enterprise policy adaptability - supply and demand forecasting models predicting changes in industry chain demand (such as a 20% increase in demand for environmental protection equipment) - generating industry governance recommendations (such as focusing on supporting environmental protection equipment suppliers). Through the above model system, full-process digital management is achieved from intelligent policy matching, industry chain relationship mining to dynamic supply and demand scheduling, forming a closed-loop ecosystem of "data-driven-model decision-making-scenario implementation", effectively improving the integration efficiency of the scientific and technological innovation supply chain and the industrial supply chain.

[0204] Example 2:

[0205] The lag problem of policy dynamic fluctuations in supply chain demand forecasting. The "supply and demand forecasting and scheduling model" in the existing solution only makes predictions based on historical procurement data and policy declaration hotspots, but does not build a real-time mapping mechanism between policy dynamic fluctuations and supply chain demand. When sudden policy adjustments occur, such as subsidy reductions and increased qualification thresholds, the existing model cannot quickly quantify the impact of the policy on demand, resulting in a lag in resource scheduling. The semantic conflict problem of cross-industry industry chain knowledge graphs. Although the existing "industry chain knowledge graph" technology stack mentions graph databases and entity recognition, it does not solve the semantic ambiguity of technical terms between different industries. For example, "chips" refer to integrated circuits in the electronics industry, and may refer to packaging components in the medical device industry. This ambiguity will lead to "pseudo-correlations" when matching cross-industry supply chains, reducing the accuracy of recommendations. Solution: Demand forecasting model for dynamic policy response (I) Core technology formula to construct a policy sensitivity correction model:

[0206]

[0207] D t ': demand forecast value at time t after correction;

[0208] D t : original time series forecast demand value;

[0209] w i : The sensitivity weight of the i-th policy to a specific industry (obtained through historical data training, range [1,1]);

[0210] ΔP i,t : The change range of the i-th type of policy at time t (such as the change in subsidy rate, the change rate of qualification threshold)

[0211] (2) Specific operation methods

[0212] 1. Construction of policy impact factor database: Establish policy type industry impact matrix,

[0213] For example:

[0214] Policy Type Photovoltaic industry weight Biopharmaceutical Weight R&D subsidy adjustments 0.7 0.3 Improved environmental standards 0.4 0.6

[0215] 2. Real-time data capture and analysis: RPA robots capture policy documents from government websites in real time and use NLP to extract policy change keywords (such as "subsidies reduced by 20%" and "qualification requirements increased to ISO certification") to calculate ΔP. i,t : For example, a 20% reduction in subsidies corresponds to ΔP = -0.2.

[0216] 3. Dynamically correct prediction: When policy changes are detected, the weight matrix calculation is automatically called

[0217] ∑wi×ΔPi,t

[0218] The original predicted value D t Make corrections and generate D t 'And trigger an early warning in the supply chain scheduling system.

[0219] (III) Example: PV Industry Policy Response Scenario 1. Scenario Background: On June 15, 2025, the National Energy Administration issued a policy to reduce PV installation subsidies from 0.3 yuan / kWh to 0.2 yuan / kWh (ΔP = 0.33). Model Calculation: The PV industry's sensitivity to the subsidy policy has a weight of w = 0.7.

[0220] Correction factor = 1 + 0.7 × (0.33) = 0.769 The original forecast for silicon material demand in July was 1,000 tons. The revised demand is D t' = 1000 x 0.769 = 769 tons 3. Scheduling optimization: the system automatically adjusts the July silicon material procurement plan from 1000 tons to 769 tons, triggering alternative material recommendations (such as partially replacing silicon materials with perovskite materials), peak shifting suggestions: delay the remaining 231 tons of demand to the Q4 policy stable period Cross-industry semantic mapping knowledge graph optimization (I) Core technology formula construction Cross-industry semantic similarity model:

[0221]

[0222] Sim(T1, T2): Semantic similarity of industry 1 term T1 and industry 2 term T2

[0223] f k (T): Feature vector of term T in the kth dimensional ontology feature space

[0224] δ k : Industry relevance coefficient of the kth feature (obtained by training across industry ontology library)

[0225] Specific operation method

[0226] 1. Cross-industry ontology library construction: establish a three-level ontology structure:

[0227] Top ontology (general concept)

[0228] (e.g. electronic industry)

[0229] (defined as integrated circuit)

[0230] (defined as shell process)

[0231] (e.g. medical devices)

[0232] (defined as sensor components)

[0233] (defined as sterile packaging)

[0234] 2. Semantic mapping training: use BERT model to train cross-industry term mapping relationship, for example: input: the context corpus of "chip" in electronic industry output: the most semantically similar "sensor components" in medical device industry

[0235] 3. The graph construction optimization process is shown in Figure 2 .

[0236] Implementation case: electronic and medical device industry term disambiguation 1. Scene background: medical device enterprise A mentions "chip packaging" in the declaration materials, and electronic enterprise B provides "chip packaging" services.

[0237] 3. Semantic analysis process: The system identifies that enterprise A belongs to the medical device industry and enterprise B belongs to the electronics industry Extracts the term "chip packaging": In the medical device industry, "chip" points to "sensor chip", and "packaging" points to "sterile packaging process" In the electronics industry, "chip" points to "integrated circuit", and "packaging" points to "shell packaging process" Calculate semantic similarity: Sim = 0.21 (belonging to low similarity)

[0238] 4. Atlas correction result: reject direct association between enterprise A and enterprise B Recommend medical device industry exclusive supplier C (providing sensor chip sterile packaging service) Generate cross-industry term mapping report:

[0239] Medical device "chip packaging" → Electronics industry "sensor chip sterile packaging"

[0240] Similarity matching: 0.21 → Recommend industry-specific term conversion

[0241] Five, benefit evaluation of scheme implementation

[0242] Standard Current solution status Expected after optimization Delayed policy response Average 72 hours <1 hour Cross-industry matching accuracy 58％ ＞92％ Supply chain scheduling efficiency Off-peak purchasing rate 35% ＞65％ Enterprise application success rate 68％ ＞85％

[0243] Through the above two innovative schemes, the deficiencies of the existing technology in policy dynamic response and cross-industry knowledge fusion can be effectively made up, forming an intelligent supply chain system of "policy semantic demand" trinity, further strengthening the technical barriers and market competitiveness of the scheme.

[0244] The present application covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present application. In order to make the public have a thorough understanding of the present application, specific details are described in the following preferred embodiments of the present application, and the present application can also be completely understood without the description of these details to those skilled in the art. In addition, in order to avoid unnecessary confusion to the essence of the present application, well-known methods, processes, procedures, elements and circuits, etc. are not described in detail.

[0245] The above is only the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be regarded as the protection scope of the present application.

Claims

1. A digital platform that uses the scientific and technological supply chain to drive the industrial supply chain, characterized by: include: Policy recommendation strategy: Policy labeling: Policy text parsing uses NLP technology to perform word segmentation, part-of-speech tagging, and syntactic analysis on policy documents to extract key elements; Enterprise portrait modeling: construct enterprise feature vectors through collected data; Intelligent recommendation algorithm: Step 1: Similarity calculation uses cosine similarity to calculate the matching degree between enterprises and policies: Where E is the enterprise feature vector and P is the policy label vector; Step 2: Collaborative filtering prediction is based on the application history of similar companies and uses matrix decomposition to predict the recommendation score: where Q p is the policy latent vector, P u is the enterprise latent vector, μ is the global average score, b u 、b p bias for businesses and policies; Step 3: Generate a recommendation list by sorting by matching degree to generate a Top-N policy recommendation list.

2. The digital platform for driving industrial supply chains with scientific and technological supply chains according to claim 1 is characterized in that: The policy labeling process includes the following steps: Step 1: Policy text parsing uses NLP technology to perform word segmentation, part-of-speech tagging, and syntactic analysis on policy documents to extract key elements; Step 2: Multi-dimensional label generation to build a three-layer label system: Policy labeling system Basic tags: release time, policy type; Condition tags: industry restrictions, enterprise size, number of patents; Benefit label: subsidy amount, tax preferential ratio; Step 3: Tag weight calculation uses the TF-IDF algorithm to calculate tag importance: Among them: tf t,d is the word frequency of label t in policy d, df t is the number of policies containing label t, and N is the total number of policies.

3. The digital platform for driving industrial supply chains with scientific and technological supply chains according to claim 2 is characterized in that: Enterprise portrait modeling includes Step 1: Data collection and integration: Enterprise reported data: revenue, R&D investment, number of patents, industrial and commercial data, bidding records; Step 2: Feature engineering to construct enterprise feature vector: E = [e1, e2, ..., e n ], where e i is the i-th feature: the proportion of R&D investment; Step 3: Dynamic weight adjustment automatically adjusts feature weights according to policy type, w i '=w i ×α p ; where α p is the characteristic adjustment coefficient corresponding to policy p.

4. The digital platform for driving industrial supply chains with scientific and technological supply chains according to claim 3 is characterized in that: The digital platform also includes an industrial chain knowledge graph construction model, which includes: Step 1: Enterprise entity identification and enterprise information extraction; Step 2: Relationship mining and construction; in: In the first step, data integration is performed by collecting various information of the enterprise; In the second step, a four-tier relationship system is constructed.

5. The digital platform for driving industrial supply chains with scientific and technological supply chains according to claim 4 is characterized in that: Business entity recognition and business information extraction include Step 1: Multi-source data integration Collect enterprise application materials, industrial and commercial information, bidding data, and patent documents; Step 2: Named Entity Recognition Use BiLSTM+CRF model to identify three types of entities: Entity Type Business Entity Product Entity technical entities; Step 3: Entity Disambiguation Use clustering algorithm to solve the problem of entities with the same name: Where sim(e i ,e j ) is the entity vector similarity, and entities with a distance less than the threshold are merged.

6. The digital platform for driving industrial supply chains with scientific and technological supply chains according to claim 5 is characterized in that: Step 1: Four-layer relationship system: supply and demand relationship, technology connection, industrial chain relationship, and policy adaptation relationship; Step 2: Relation Extraction The algorithm uses a neural network model with remote supervision and attention mechanism, inputs a sentence fragment S, and outputs the relationship probability: P(r|S)=softmax(W·tanh(U·eS+V·er) where e S is the sentence embedding vector, e r Embedding vectors for relation types; Step 3: Calculate relationship weights. Calculate relationship strength based on transaction frequency, cooperation duration, and other dimensions: w r =α·f+β·t+γ·s Where f is the transaction frequency, t is the cooperation duration, s is the transaction size, and α+β+γ=1. Graph dynamic update mechanism Step 1: Data monitoring captures real-time data on business changes, bidding announcements, policy updates, etc. Step 2: Change Identification Use a differential algorithm to identify data changes: ΔE=E new -E old Among them E new and E old are the attribute vectors of the new and old enterprise entities respectively; Step 3: The graph update strategy uses incremental update: Add new entity: add entity node and initialize relationship; Entity attribute changes: update node attributes and recalculate relationship weights; Relationship changes: Update relationship strength or delete invalid relationships.

Images