Dynamic quantitative evaluation method for enterprise development process and electronic equipment

Abstract

The invention relates to the field of data fusion evaluation, and provides a dynamic quantitative evaluation method for an enterprise development process, which comprises the following steps of: S1, acquiring operation data and compliance data of an enterprise through a multi-source heterogeneous data interface, and executing entity association and conflict resolution operation to generate a structured basic data set; s2, extracting dynamic evaluation features of the structured basic data set through a data analysis engine, wherein the data analysis engine comprises an industry adaptive rule base and / or a deep feature mining network; s3, calculating a multi-dimensional score value of the enterprise development situation based on the dynamic evaluation features; and S4, according to a historical evaluation record, mapping the multi-dimensional score value to a preset grade interval and presenting the multi-dimensional score value. And furthermore, objective and real-time evaluation is carried out on the growth performance and the anti-risk capability of the non-listed companies.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data fusion evaluation, and in particular to a dynamic quantitative evaluation method for the development process of an enterprise. BACKGROUND

[0002] The development of regional economy is highly dependent on the health of enterprise groups, especially non-listed enterprises and start-ups as the main force of innovation, whose growth and risk situation directly determine the effectiveness of industrial planning. Traditional evaluation relies on listed company financial reports, manual research and static business data, which has three fundamental defects: first, the operation data of non-listed enterprises is fragmented, and key indicators such as social security payment, R&D investment and supply chain relationship are scattered in more than ten independent systems such as tax, social security and patent, and the government lacks the ability of automatic multi-source heterogeneous data fusion. A city's investment department feedback that it takes twenty special officers to spend three months to collect basic data of three thousand enterprises in the jurisdiction, and forty percent of the enterprises are misclassified as "zombie enterprises" due to data missing. Second, the current evaluation model generally uses annual financial snapshot analysis, which cannot capture the dynamic evolution characteristics of enterprises. When a biotechnology enterprise stops R&D pipeline due to financing interruption, the traditional method needs to wait for the disclosure of the audit report of the next year to identify the risk, delaying the rescue window period for six to eight months. Third, the evaluation results lack explainability, and the government cannot trace the causes of the score. A smart manufacturing enterprise was rated B level and went bankrupt three months later, and after tracing, it was found that the equity pledge risk was classified as "other deduction items" in the evaluation report, and the decision-making department failed to timely locate the core hidden danger of eighty-seven percent of the concentrated stockholding.

[0003] Existing technologies still have little effect through single dimension improvement. In some existing technologies, an enterprise vitality evaluation model based on public opinion sentiment analysis is proposed, but the data source credibility problem is not solved; while in other existing technologies, financial indicators are weighted and scored, but industry differentiated thresholds are not established (such as the sixty percent asset-liability ratio red line for both technology enterprises and manufacturing enterprises). The more fundamental contradiction is that the static evaluation framework is fundamentally inconsistent with the dynamic development law of enterprises. The growth of enterprises presents nonlinear characteristics - a new energy enterprise realizes a twenty-fold increase in valuation within six months due to a technological breakthrough, and the traditional quarterly evaluation cycle completely misses such inflection point signals; risk transmission has a network effect - a food and beverage chain enterprise's food safety incident spreads exponentially through social media, causing a wave of supplier termination within seventy-two hours, but isolated enterprise dimension evaluation cannot capture industry chain level risks.

[0004] The particularity of government decision-making scenarios further amplifies the technical bottleneck. Industrial policy making needs to balance the development quality of individual enterprises and the collaborative efficiency of regional clusters, but existing tools can only output isolated enterprise scores. When a certain high-tech zone introduced a chip manufacturing enterprise, it found that 90% of its raw materials relied on a single importer, and after production, it encountered a supply interruption that led to a 5 billion production value project being shelved. The underlying need is to build a trinity capability: real-time sensing of non-transparent data (such as predicting capacity contraction through electricity consumption fluctuations), dynamically quantifying industry adaptation indicators (such as the R&D investment threshold of technology enterprises floating with technology generations), and visualizing risk transmission paths (such as the causal chain of equity pledge → financing ability → R&D investment). The current technical system has significant gaps in these dimensions, leading to a three to six-month lag in the government's understanding of the actual development of enterprises compared to market changes, and the scientific nature of regional industrial planning has long been hindered by data blind spots and evaluation inaccuracies.

[0005] In summary, as the main force of regional industrial development, how to accurately grasp high-quality enterprises with good development prospects, and then develop more efficient spatial planning and business recruitment and stabilization strategies, is the focus of government decision-making departments. However, the financial and operational data of enterprises (especially non-public companies or start-ups) may not be transparent or systematically disclosed, leading to a lack of basis for evaluation. Enterprise strategic stories (such as "metaverse" and "blockchain" concepts) may mask actual business growth. At the same time, the lack of systematic data tools requires searching for dynamic needs such as business income, patent applications, investment and financing capabilities, legal risks, and other information in the vast amount of information on the Internet, which is time-consuming and labor-intensive and ineffective. All these factors lead to a lack of objective evaluation basis for the growth of enterprises.

[0006] The enterprise growth evaluation model requires comprehensive enterprise public data, summarizes the periodic dynamic trends of enterprises, focuses on the dynamics of organizations, funds, projects, social media, securities trading of enterprises established for more than 3 years, and reflects it through scoring, and displays it in a data visualization form. The score requires monthly updates for calculation, and the results of each month are retained to observe the dynamics of the growth score.

[0007] On the other hand, enterprise risks are not transparent in the early stages of development and rarely disclosed to the public, and the enterprise's ability to resist risks is unknown, lacking means to understand and intervene in advance, leading to the inability of government decision-making departments to objectively evaluate the development risks of enterprises in the jurisdiction and to effectively help and serve enterprises in a timely manner.

[0008] The enterprise anti-risk capability evaluation model requires comprehensive evaluation of enterprises in terms of enterprise assets, liabilities, gross profit rate fluctuations, administrative penalty records, and equity concentration, etc. The score requires monthly updates for calculation, and the monthly results are retained to facilitate observation of changes in the anti-risk capability of the enterprise.

[0009] The above defects collectively point to the need for systematic technical reconstruction: a multi-source dynamic data acquisition, industry adaptive computing, and risk causal attribution evaluation system must be established to fundamentally solve the three bottlenecks of "data deficiency, weak timeliness, and difficult attribution" in the evaluation of non-public enterprises, and provide real-time, objective, and traceable decision-making foundation for regional economic governance. SUMMARY

[0010] Based on the above defects, one of the purposes of the present application is to solve the technical problem of objectively and real-time evaluating the growth and anti-risk capability of non-public companies.

[0011] To solve the above technical problems, one embodiment of the present application proposes a dynamic quantitative evaluation method for the development process of an enterprise, characterized by comprising the following steps: S1. Collecting the operation data and compliance data of the enterprise through a multi-source heterogeneous data interface, and performing entity association and conflict resolution operations to generate a structured basic data set; S2. Extracting dynamic evaluation features of the structured basic data set through a data analysis engine, the data analysis engine including an industry adaptive rule base and / or a deep feature mining network; S3. Calculating a multi-dimensional score value of the development trend of the enterprise based on the dynamic evaluation features; S4. Mapping the multi-dimensional score value to a preset grade interval and presenting according to historical evaluation records.

[0012] Optionally, in some embodiments, the data analysis engine is an industry adaptive rule base; and The step S3 further comprises: calling the industry adaptive rule base to perform dynamic calculation on the structured basic data set, and outputting a dynamic quantitative score set; The step S4 further comprises: fusing the dynamic quantitative score set and the historical evaluation records to generate a dynamic trend report, and mapping to the preset grade interval for visual output.

[0013] Optionally, in some embodiments, the data analysis engine is a deep feature mining network; and The step S2 further comprises: extracting multi-scale evolution features of the structured basic data set based on a deep feature mining network and generating an enterprise dynamic index map; The step S3 further comprises: constructing a dynamic prediction engine based on causal inference to calculate a multi-dimensional score value of the enterprise dynamic indicator graph on the development trend of the enterprise in real time; The step S4 further comprises: generating and displaying the dynamic trend report containing cause link through an interpretable knowledge presentation module, thereby accelerating the decision-making closed loop effect of the manager.

[0014] Optionally, in some embodiments, the step S1 further comprises: S11 collecting original enterprise data streams from social security payment systems, enterprise financing databases, and public opinion analysis platforms in real time through the multi-source heterogeneous data interface; S12 performing distributed data cleaning operations on the original enterprise data streams to generate intermediate cleaning data sets; S13 performing entity association operations on the intermediate cleaning data sets based on unified social credit codes to output associated data sets; S14 processing contradictory information in the associated data sets through a rule engine driven conflict resolution algorithm to generate the structured basic data set.

[0015] Optionally, in some embodiments, the step S2 further comprises: S2.1' obtaining original public opinion text streams from a public opinion analysis platform through the multi-source heterogeneous data interface; S2.2' constructing a knowledge graph engine based on natural language processing to analyze the original public opinion text streams to identify enterprise strategic partner entities; S2.3' matching the identified strategic partner entities with an industry chain database using a graph neural network to quantify cooperation value weights; S2.4' dynamically correcting media dynamic dimension scores based on the quantification results and inputting the correction results into the multi-dimensional score value calculation module.

[0016] Optionally, in some embodiments, the step S4 further comprises: S4.1 accessing a dynamic benchmark data set of an industry economic knowledge graph library; S4.2 reconstructing boundary determination rules of the preset grade intervals based on the dynamic benchmark data set to generate an industry-adaptive grade mapping knowledge base; S4.3 monitoring national-level technical standard formulation events through a real-time event capture engine, and activating a fast grading channel when an enterprise-led technical standard formulation is identified; S4.4 synergistically fusing the industry-adaptive grade mapping knowledge base with the output of the fast grading channel to generate a three-dimensional space decision graph and push it to a regional economic governance platform.

[0017] Optionally, in some embodiments, the subject enterprise is a software enterprise, and the step S4 further comprises: S41 Real-time monitoring of the collaboration network changes of the open source community core code repository to capture the contribution behavior migration trajectory of the main developer cluster; S42 Dynamic scanning of the evolution path of hot topics in technical forums to identify the community attention fluctuation inflection point of the enterprise core technology stack; S43 Correlating the product iteration logs of the enterprise with the competitive product release dynamics to construct a technology generation difference space topology map; S44 When the monthly increase in the number of core module contributors exceeds 200% of the industry average, activate the technology leadership channel.

[0018] Optionally, in some embodiments, the step S3 further comprises: S31 Obtain social security payment data streams through the multi-source heterogeneous data interface, calculate the employee compound growth rate for the past 36 months, and score according to the organization dynamic scoring rules; S32 Call the financing database of Qichacha to analyze the financing rounds in the past 3 years, match the financing stage scoring rules to generate the funding dimension score; S33 Real-time monitoring of public opinion platforms and administrative penalty databases, and when major lawsuits or continuous loss events are triggered, deduct the total score according to the risk deduction item rules; S34 Aggregate the organization dynamic score, the funding dimension score, and the risk deduction result on a monthly basis to generate a dynamic growth score and correlate historical data.

[0019] Optionally, in some embodiments, the data analysis engine is a deep feature mining network; and, The step S2 further comprises: S2.1 Perform time slicing operations on the structured basic data set to separate out quarterly operational indicator sequences, annual financial indicator sequences, and three-year strategic indicator sequences; S2.2 Analyze the evolution correlation in the quarterly operational indicator sequences through the convolution attention module of the deep feature mining network to identify cross-cycle indicator coupling rules; S2.3 Fuse the cross-cycle indicator coupling rules with the three-year strategic indicator sequences to construct an enterprise dynamic indicator map with weighted topological relationships; The step S3 further comprises: constructing a dynamic prediction engine based on causal inference to calculate the multi-dimensional score value of the enterprise dynamic indicator map on the enterprise development trend in real time; The step S4 further comprises: generating and displaying a dynamic trend report containing cause links through an interpretable knowledge presentation module.

[0020] Optionally, in some embodiments, the data analysis engine is a deep feature mining network; and, The step S2 further includes: S2.1 performing a time dimension slicing operation on the structured basic data set to separate out a quarterly high-frequency fluctuation indicator sequence and a three-year strategic evolution indicator sequence; S2.2 performing cross-scale correlation analysis through a double-path convolution attention module of the deep feature mining network to capture time lag coupling rules between the quarterly sequence and the three-year sequence; S2.3 constructing an enterprise dynamic indicator graph with weight distribution based on the time lag coupling rules, wherein a node represents a core business entity and an edge weight quantifies an influence coefficient between indicators; The step S3 further includes: constructing a dynamic prediction engine based on causal inference to calculate a multi-dimensional score value of the enterprise dynamic indicator graph on the enterprise development trend in real time; The step S4 further includes: generating and displaying a dynamic trend report containing cause links through an interpretable knowledge presentation module.

[0021] The scheme of the embodiment is based on some of the above embodiments, limits the data analysis engine to be a deep feature mining network, and further specifies the specific implementation of the time slicing operation (step S2.1), the cross-scale correlation analysis (step S2.2), and the graph construction (step S2.3). The scheme establishes a multi-scale observation and time lag transmission decoding mechanism to solve the problem that the implicit correlation between short-term fluctuations and long-term strategies of an enterprise is difficult to quantify.

[0022] The time dimension slicing of step S2.1 separates the data stream into a quarterly high-frequency fluctuation sequence (such as monthly order fulfillment rate) and a three-year strategic evolution sequence (such as patent quality proportion). This separation operation visualizes contradictory signals, for example, a sudden decrease in laser radar procurement of a certain intelligent driving enterprise is classified into the quarterly sequence, while the visual algorithm patent proportion rises to 70% and is classified into the three-year sequence, avoiding the misjudgment of traditional models that the technology roadmap switching is recession. The double-path convolution attention module of step S2.2 implements differentiated processing: the convolution path detects quarterly indicator abnormal pulses (such as a 300% increase in media exposure in a certain quarter), and the attention path quantifies the time lag influence of the abnormal pulses on strategic indicators (such as a 6-month increase in brand value leading by exposure surge, with a correlation coefficient of 0.68). A short-term increase in raw material cost of a certain photovoltaic enterprise is identified as a supply chain disturbance (time lag <3 months) by this module, avoiding the misjudgment of long-term profit decline.

[0023] Step S2.3: Based on the time-lag coupling rule, a dynamic index map of the enterprise is constructed, with nodes representing core business entities (such as "R&D investment") and edge weights quantifying influence coefficients (such as a 1% increase in R&D investment leading to a 0.5% increase in market share after 6 months). When a certain AI enterprise experiences a quarterly decline in R&D investment, the map distinguishes between technical route adjustment (coefficient 0.2) and R&D recession (coefficient 0.8) based on historical coefficients, guiding the dynamic prediction engine to make differentiated corrections to the score. The three-step technology chain is demonstrated in a certain new energy battery enterprise: the traditional model triggers an early warning when gross margin decreases by 15%, while the system identifies the impact of short-term fluctuations in cobalt prices (time lag <4 months) through time-lag analysis, and generates a "raw material fluctuation → technology substitution" transmission path (influence coefficient 0.3) based on the strategic signal that solid-state battery patents accounted for 45% of the total in the third year, maintaining an A-level score for 6 months after the enterprise secured a 10 billion order.

[0024] In summary, the several steps of this embodiment, through the triple innovation of time-slicing baseline construction, dual-path correlation decoding, and map parameter conversion, first realize the collaborative analysis of short-term and long-term indicators of enterprises, and reduce the misjudgment rate during strategic transformation to below 5% (industry average 35%).

[0025] Optionally, in some embodiments, the step S2.2 further comprises: S2.2.1: Adopting a multi-head attention mechanism to analyze abnormal pulse signals in the sequence of quarterly-level operational indicators and mark key fluctuation events; S2.2.2: Performing pulse response analysis on the key fluctuation events through a set of time-domain convolution kernels, calculating the time-lag influence intensity on annual-level indicators; S2.2.3: Based on the time-lag influence intensity, constructing a cross-cycle coupling coefficient matrix and writing it into the index correlation rule library.

[0026] In this embodiment, based on the above-mentioned embodiments, a multi-head attention mechanism is added to mark key fluctuation events (step S2.2.1), pulse response analysis of time-lag influence (S2.2.2), and construction of a coupling coefficient matrix (S2.2.3) in cross-scale correlation analysis. This scheme solves the quantification problem of nonlinear coupling between mutation events and strategic indicators, improving the recognition accuracy of the evaluation system for enterprises in the transformation period.

[0027] The multi-head attention mechanism of step S2.2.1 scans three types of quarterly sequence mutation patterns in parallel: spike type (single-month media exposure surge by 300%), cliff type (order fulfillment rate drops by 40% in two months), and plateau type (gross margin rate maintains a new threshold for 6 months). The sudden drop of R&D expenditure of a certain biological and pharmaceutical enterprise by 50% is marked as a "platform transition" event, avoiding the traditional model's technical recession deduction operation. The time domain convolution kernel group channel processing of step S2.2.2: the short-term kernel (3-month span) calculates the immediate impact of market cost surge on quarterly revenue (intensity 0.8), and the medium-term kernel (12-month span) analyzes the lag effect of R&D investment decline on patent output (time lag 9 months, intensity 0.6). The zero purchase of laser radar by a certain unmanned aerial vehicle enterprise triggers an alarm, but the medium-term kernel detects a strong negative correlation (r=-0.91) between it and the jump in visual algorithm patent quantity, revealing the essence of technology generation replacement.

[0028] Step S2.2.3 converts the impulse response output into a three-dimensional coefficient matrix, for example, the "technology roadmap switching → R&D patent output" entry defines a time lag period of 9 months and an intensity coefficient of 0.75. When the GPU procurement of an AI enterprise drops sharply, the matrix automatically matches the "hardware to cloud service" historical rule (intensity 0.68), and the dynamic engine maintains the R&D dimension score. The three-step synergy is demonstrated in a certain power battery enterprise: a 40% quarterly decline in R&D expenditure is marked as a platform transition; pulse analysis shows a positive correlation between it and solid-state battery patent output (time lag 6 months, intensity 0.82); the generated "R&D expenditure platform transition → technology roadmap upgrade" rule raises the enterprise rating from B to A, and 12 months later its solid-state battery market share ranks first in the industry.

[0029] These three steps compress the decoding error of the association between enterprise mutation events and strategic transformation to the following by intelligently marking events, quantifying time domain conduction, and encoding rule matrices.

[0030] Optionally, the method of some embodiments includes the following steps: Step 1: Perform multi-head attention calculation on quarterly key fluctuation events and annual strategic indicator sequences through a parallel event analysis module to generate K sets of initial event-indicator association weights, where K≥3 and each set of weights corresponds to a differentiated association pattern; Step 2: Perform modulation processing on the initial event-indicator association weights using a nonlinear activation function to output standardized weight values in the value range [-1, 1], where negative values represent negative correlation; Step 3: Dynamically configure the window width of the time domain convolution kernel according to the pre-defined industry type parameters, and perform industry-adaptive convolution operation on the key fluctuation events to generate time domain response values; Step 4: Element-wise fusion of the normalized weight values and the time domain response values to generate impulse response components of a single attention head; Step 5: Aggregation of the impulse response components of all K attention heads to generate a comprehensive impulse response strength value for correcting the enterprise development trend assessment result.

[0031] The five steps of this embodiment specifically define the impulse response strength calculation process, including multi-head attention weight generation (step 1), nonlinear modulation (step 2), industry adaptive convolution (step 3), component fusion (step 4), and aggregation output (step 5). This scheme solves the defect of single processing of the correlation pattern between fluctuation events and strategic indicators in traditional models, and realizes precise quantification of multi-mechanism coupling strength.

[0032] The parallel event analysis module of step 1 generates K sets of initial correlation weights corresponding to different patterns such as technology route switching, market response delay, and supply chain transmission. A 60% reduction in laser radar procurement of a certain chip enterprise is routed to the technology route switching channel to avoid misjudgment of supply chain crisis. Step 2 uses the hyperbolic tangent function to modulate the initial weights to normalized values in the [-1, 1] interval, with negative values indicating negative correlation events. When a decrease in hardware investment is detected but cloud patents increase, the system automatically identifies the reverse correlation. Step 3 dynamically configures the convolution kernel width based on industry type parameters, for example, setting a 12-month window to match the clinical translation period in the biomedicine industry.

[0033] Step 4 fuses the normalized weights and convolution output values through Hadamard product to give the calculation result a clear physical meaning. A 40% reduction in R&D expenditure of a certain new energy enterprise outputs a negative 0.75 weight through the technology route pattern head, and multiplying it by the 0.4 amplitude value of the convolution output gives a positive 0.3 component, indicating that the event is a strategic upgrade. Step 5 aggregates the multi-channel component values, for example, in the case of a certain autonomous driving, the technology transformation, market response, and supply chain components are 0.35, 0.6, and -0.2 respectively, generating a weighted 0.28 comprehensive strength value, with the positive direction clearly indicating the essence of technology upgrade. The five-step process outputs Ψ(e, I)=0.78 in the context of R&D fluctuations (e=-0.4) in a certain biomedicine enterprise, accurately identifying strategic transformation and avoiding a 15-point deduction.

[0034] In summary, the technical scheme of this embodiment reduces the strategic misjudgment rate to below the industry average through orthogonal pattern separation, industry knowledge coding, and physical constraint fusion.

[0035] Optionally, in some embodiments, the step S2.2' of constructing a knowledge graph engine further includes: Step K1: Extracting semantic resonance waveform features between enterprise entities from the original public opinion text stream, including amplitude parameters of keyword co-occurrence frequency and phase difference of event interval; Step K2: input the semantic resonance waveform features into the graph neural network matching module, and generate an implicit correlation strength value by modulating and coupling the resonance threshold value through the industry characteristic vector of the industry chain database; Step K3: construct a dynamic decay function based on the implicit correlation strength value and the time delay coupling law in the enterprise dynamic indicator graph, and update the strategic partner entity relationship weight of the dynamic enterprise knowledge graph.

[0036] In this embodiment, semantic resonance waveform feature extraction (K1), graph neural network matching (K2), and relationship weight updating (K3) are added in the construction of the knowledge graph engine. This scheme breaks through the recognition bottleneck of the implicit correlation of enterprise entities in public opinion text, and improves the accuracy of strategic cooperation value assessment.

[0037] Step K1 extracts semantic resonance waveform features from the original public opinion text stream, including keyword co-occurrence frequency amplitude (such as the co-occurrence frequency peak of "a certain AI enterprise" and "Huawei") and event interval phase difference (such as the time difference between cooperation announcement and R&D progress release). The text of a certain quantum technology company and a university jointly applying for a patent is processed in this way, and the resonance waveform parameters of the "industry-university-research" entity pair are identified. Step K2 inputs the waveform features into the graph neural network, and modulates and couples the threshold value through the industry characteristic vector of the industry chain database. For example, when the partner belongs to the cloud computing industry chain, the resonance threshold value is automatically reduced to capture the technology synergy signal, and the implicit correlation strength value of "a certain AI enterprise-Huawei" is output as 0.82.

[0038] Step K3 constructs a dynamic decay function to update the entity relationship weight based on the implicit correlation strength and the time delay law in the enterprise dynamic indicator graph. The initial weight of a certain new material enterprise and Tsinghua University's joint patent application is 0.6, and after identifying the technology route synergy through semantic resonance analysis technology, the weight is adjusted to 0.75 and triggers the R&D capability score correction. The three-step linkage is demonstrated in a certain biological medicine enterprise: the semantic resonance features extracted from the "joint laboratory" public opinion text are matched through the graph neural network, and the cooperation value weight of the enterprise and a certain research institute is quantified; combined with the time delay law of the R&D pipeline indicator, the cooperation weight is dynamically corrected from 0.4 to 0.68, driving the media dynamic dimension score to increase by 12%.

[0039] In summary, the technical scheme of the embodiment compresses the strategic cooperation value assessment error to 1 / 5 of the industry benchmark through semantic waveform quantization, industry chain knowledge injection, and dynamic decay function.

[0040] Optionally, in some embodiments, the step S4.4 further includes: S4.4.1 Extract the risk transmission path in the dynamic enterprise knowledge graph with a confidence level below a preset threshold, and generate a cross-entity weak correlation topology subgraph; S4.4.2 Inject event correlation features based on real-time incremental public opinion data flow, reconstruct the multi-hop conduction path intensity coefficient in the weak correlation topology subgraph; S4.4.3 Map the reconstructed conduction path intensity coefficient to the risk thermal dimension of the three-dimensional spatial decision graph, label the industry chain level opportunity weight distribution and push it to the regional economic governance platform.

[0041] Through the technical solutions of the present embodiment, the risk conduction path reconstruction (S4.4.1-S4.4.3) is added in the generation of the three-dimensional spatial decision graph. This scheme overcomes the problem of strong concealment of the industry chain level risk conduction path, which cannot be captured by traditional evaluation, and realizes the dynamic visualization of cross-entity weak correlation risk.

[0042] Step S4.4.1 extracts risk conduction paths with a confidence level below a threshold value in the dynamic enterprise knowledge graph, and generates a cross-entity weak correlation topology subgraph. For example, after a battery safety accident of a new energy vehicle enterprise, the system automatically marks the low-confidence conduction path of "battery cell supplier → battery pack manufacturer → whole vehicle manufacturer". Step S4.4.2 injects real-time incremental public opinion data flow to reconstruct the conduction intensity. For example, an exponential diffusion event on social media increases the multi-hop path intensity coefficient from 0.3 to 0.75. After processing a food safety event of a catering chain enterprise, 22 upstream suppliers with a transmission intensity interval of (0.45-0.82) are identified within 72 hours.

[0043] Step S4.4.3 maps the reconstructed path to the risk thermal dimension of the three-dimensional decision graph, and labels the industry chain level opportunity weight distribution. In the previous example, the supplier transmission intensity is converted into a risk thermal graph, and after superimposing the industrial geographic information of the regional economic governance platform, the 5 core suppliers that need to be intervened first are accurately located. The empirical case shows that, six months before a chip manufacturing project in a certain high-tech zone was interrupted, the system identified a risk that 90% of the raw materials were dependent on a single importer through the weak correlation topology subgraph, labeled the industry chain level opportunity weight distribution map, and pushed it to the regional platform, avoiding the shelving of a 5 billion yuan project.

[0044] In summary, the several steps of the present embodiment improve the timeliness of industry chain level risk identification by several months through weak correlation topology extraction, event feature reconstruction, and thermal space mapping.

[0045] Optionally, in some embodiments, the step S3 further includes the following steps: S3.5 Extracts at least three consecutive years of core growth indicator historical sequences from the enterprise dynamic indicator graph, including time evolution data of operating income growth rate, patent authorization quantity, and R&D investment proportion; S3.6 Selects a prediction model architecture based on technology generation evolution features, including using trend average method to process linear growth indicators, and applying multivariate regression model to analyze non-linear technology transition indicators; S3.7 inputting the core growth indicator historical sequence into the prediction model architecture for iterative training, dynamically adjusting model parameters by an adaptive learning rate optimizer until convergence; S3.8 generating a future three-year enterprise growth dynamic prediction report based on the trained prediction model architecture, and associating the technology generation gap space topology map to mark the technology inflection point risk.

[0046] Through the several steps in the embodiment, the growth indicator historical sequence extraction (step S3.5), the prediction model architecture selection (step S3.6), the iterative training (step S3.7), and the technology inflection point marking (step S3.8) are added. This scheme solves the industry problem of mixed prediction of linear and nonlinear indicators in technology generation evolution, and improves the accuracy of the growth dynamic report.

[0047] Step S3.5 extracts at least three years of core growth indicator historical sequence in the enterprise dynamic indicator map, including time evolution data such as revenue growth rate, patent authorization quantity, and R&D investment proportion. A certain unmanned aerial vehicle enterprise obtains the coupling sequence of supply chain delay rate and revenue decay for the past 36 months. Step S3.6 selects a mixed prediction architecture based on the characteristics of technology generation evolution: trend average method for linear growth indicators (such as traditional business revenue), and multivariate regression model for nonlinear transition indicators (such as solid-state battery patent quantity). When detecting a technology route bifurcation node, the system automatically configures the sliding window width and the activation function combination.

[0048] Step S3.7 dynamically adjusts model parameters by an adaptive learning rate optimizer until convergence. In the historical data training of a certain AI enterprise, the learning rate is adaptively scaled according to the fluctuation inflection point distribution of community attention, avoiding the interference of adversarial samples of technology fault history cases. Step S3.8 generates a future three-year growth prediction report associated with the technology generation gap space topology map to mark the technology inflection point risk. In the empirical case, a new energy enterprise prediction report highlights the "lithium battery material energy density approaching the theoretical limit" inflection point, guiding the government to layout the sodium-ion battery pilot line in advance.

[0049] In summary, in the scheme of the embodiment, the prediction accuracy of technology generation replacement is improved to the industry benchmark through historical sequence deep extraction, mixed model architecture adaptation, and inflection point space marking.

[0050] Optionally, in some embodiments, the step S3.6 further includes a construction step of the prediction model architecture: Step P1: dynamically configuring the sliding window width of the trend average method and the nonlinear activation function combination of the multivariate regression model based on the technology route bifurcation node parameters in the technology generation evolution characteristics; Step P2: Encode the weight topology relationship in the enterprise dynamic indicator map as an attention priori knowledge matrix, and inject the hidden layer neurons of the multiple regression model through a feature space projection module; Step P3: Perform small sample transfer training on the multiple regression model with priori knowledge injection using a meta-learning framework, and construct an adversarial disturbance sample set using a technical fault history case library; Step P4: Integrate a technology maturity sensor in the adaptive learning rate optimizer, and dynamically adjust the gradient update amplitude according to the community attention fluctuation inflection point in the technology generation difference spatial topology map.

[0051] Optionally, in some embodiments, the step S3 further comprises: S3.1 Training a multivariate time prediction model based on historical operation data sequences in the enterprise dynamic indicator map, and dynamically adjusting the weight coefficients of revenue growth rate and net profit growth rate to generate initial prediction values through a gradient back propagation algorithm; S3.2 Injecting real-time incremental event features of the dynamic enterprise knowledge graph into the multivariate time prediction model, and correcting the initial prediction values according to the entity relationship confidence decay ratio to output a financial indicator prediction distribution with a time limit label; S3.3 Mapping the financial indicator prediction distribution to a dynamic deviation heat map layer in a three-dimensional space decision map through a visual analysis engine, labeling the key indicator fluctuation confidence interval, and feeding back to the parameter optimization module of the dynamic prediction engine based on causal inference. In the technical solution of this embodiment, the dynamic prediction engine includes multivariate time model training (S3.1), real-time event correction (S3.2), and visual feedback (S3.3). This solution breaks through the limitations of historical path dependence in financial indicator prediction, and realizes incremental event-driven dynamic deviation early warning.

[0052] Step S3.1 trains a multivariate time prediction model based on historical operation sequences of the enterprise dynamic indicator map, and dynamically adjusts the weight coefficients of revenue growth rate and net profit growth rate through gradient back propagation. After injecting the clinical III phase data of a certain biological enterprise, the model increases the revenue growth rate weight from 0.6 to 0.8. Step S3.2 injects real-time incremental event features into the model, and corrects the initial prediction values according to the entity relationship confidence decay ratio. When the depreciation rate of a certain chip enterprise increases by 1% per quarter, the system automatically associates the historical law of annual gross margin lagging behind by 0.5%, and generates a financial prediction distribution with a time limit label.

[0053] Step S3.3 maps the prediction distribution into a dynamic bias heat map layer of the three-dimensional decision graph by the visualization engine, and labels the confidence interval of the key indicator fluctuation. The margin of error of the gross profit margin prediction in the previous example is converted into a heat gradient, and the 95% confidence interval range is fed back to the parameter optimization module of the dynamic prediction engine. The experiment shows that, in the silicon material price fluctuation event of a certain photovoltaic enterprise, the system can provide a 3-month early warning of the gross profit margin deviation threshold compared with the traditional model, guiding the government to start the supply chain emergency fund.

[0054] In summary, in the interaction process of the above several steps, the financial indicator prediction bias is compressed below the industry average through historical path optimization, incremental event correction, and heat visualization feedback.

[0055] Optionally, in some embodiments, the step S3.1 of training the multivariate time prediction model comprises: Step M1: Extracting the time delay influence strength matrix corresponding to the cross-cycle coupling rule from the enterprise dynamic indicator graph, and constructing a multivariate time series prediction network architecture with gated recurrent units and self-attention mechanisms; Step M2: Injecting an adversarial disturbance sample set through the real-time incremental event feature flow of the dynamic enterprise knowledge graph, and executing a meta-learning training process optimized by time convolution to enhance the network weight parameters; Step M3: Based on the hidden layer output feature space of the multivariate time series prediction network, dynamically reconstructing the output layer activation function combination by fusing the route bifurcation node parameters in the technology intergenerational evolution features; Step M4: Using the technology maturity sensor in the adaptive learning rate optimizer to adaptively scale the gradient update amplitude according to the community attention fluctuation inflection point distribution and generate a final network parameter set.

[0056] The above four steps are used to specify the training steps of the multivariate time prediction model. This scheme solves the prediction model misalignment problem caused by technical route bifurcation, and realizes small sample transfer learning guided by prior knowledge.

[0057] Step M1 extracts the cross-cycle time delay influence strength matrix from the enterprise dynamic indicator graph, and constructs a multivariate time series prediction network containing gated recurrent units and self-attention mechanisms. A quantum computing enterprise establishes a time delay strength matrix of “R&D investment → patent output → financing valuation” accordingly. Step M2 injects an adversarial disturbance sample through the real-time incremental event flow, and uses a meta-learning process optimized by time convolution to enhance the network weight. When a technical fault history case (such as a lithography machine supply interruption event) is identified, the system automatically generates an adversarial sample for intensive training.

[0058] Step M3 reconstructs the output layer activation function combination based on the network hidden layer output feature space and the technology route bifurcation node parameter. The technology bifurcation point of a laser radar to visual scheme of an autonomous driving enterprise triggers the activation function reorganization to adapt to the nonlinear transition prediction. Step M4 integrates the technology maturity sensor in the adaptive learning rate optimizer, and scales the gradient update amplitude according to the community attention fluctuation inflection point distribution. The empirical case shows that the system maintains a prediction accuracy of 87% during the technology route bifurcation period of a blockchain enterprise, which is 40 percentage points higher than the traditional model.

[0059] In this way, the prediction stability during the technology transition period is improved to above the industry benchmark through time delay matrix modeling, adversarial meta-learning and bifurcation perception optimization.

[0060] In some other embodiments of the present application, an electronic device for enterprise potential assessment prediction based on multi-source heterogeneous data is also proposed, comprising a display, one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for executing the steps in the method of any other embodiment.

[0061] The data analysis engine further comprises a deep feature mining network; and The step S2 further comprises: S2.1 performing a time slicing operation on the structured basic data set to separate out a quarterly-level operation indicator sequence, an annual-level financial indicator sequence and a three-year-level strategic indicator sequence; S2.2 analyzing the evolution correlation in the quarterly-level operation indicator sequence through a convolution attention module of the deep feature mining network to identify a cross-cycle indicator coupling rule; S2.3 fusing the cross-cycle indicator coupling rule and the three-year-level strategic indicator sequence to construct an enterprise dynamic indicator graph with a weighted topological relationship; The step S3 further comprises constructing a dynamic prediction engine based on causal inference to calculate a multi-dimensional score value of the enterprise dynamic indicator graph on the enterprise development trend in real time; The step S4 further comprises generating and displaying a dynamic trend report containing cause links through an interpretable knowledge presentation module.

[0062] Optionally, step S2.2 further comprises: S2.2.1 analyzing abnormal pulse signals in the quarterly-level operation indicator sequence and marking key fluctuation events using a multi-head attention mechanism; S2.2.2 performing pulse response analysis on the key fluctuation events through a time domain convolution kernel group to calculate the time delay influence strength of the key fluctuation events on the annual-level indicators; S2.2.3 Construct a cross-period coupling coefficient matrix based on the time lag influence strength and write into the index correlation rule library.

[0063] Optionally, the step S2.2 can further include the following calculation sub-step: Step 1: Feature extraction of key fluctuation events Through the kth attention head function Attn in the multi-head attention mechanism k Process the feature vector Q of the quarterly key fluctuation event e e With the feature vector K of the annual strategic index sequence I i , generate the initial attention weight matrix M k , where k = 1, 2,..., K and K ≥ 3.

[0064] Step 2: Activation weight modulation Input the initial attention weight matrix M k into the hyperbolic tangent activation function tanh for nonlinear transformation, and output the modulated attention weight value w k , the value of which ranges from -1 to 1, and a negative value represents a negative correlation.

[0065] Step 3: Time domain convolution processing According to the industry type parameter Ω, determine the time domain convolution kernel width τ, and use the kth time domain convolution kernel function Conv k in the time offset Δ t defined window to perform convolution operation on the key fluctuation event e, and generate the convolution output value c k .

[0066] Step 4: Hadamard product fusion Element-wise multiplication (Hadamard product: ∘) is performed between the modulated attention weight value w k and the convolution output value c k , to obtain the impulse response component Ψ k of the kth attention head.

[0067] Step 5: Multi-head result aggregation Perform summation operation on the impulse response components Ψ k of all K attention heads to generate the final impulse response strength value Ψ(e, I), and the calculation formula is:

Technical effects

[0068] Step S2 constructs a dynamic enterprise knowledge graph based on the structured data set, driving real-time evolution of entity relationships. The industry adaptive rule base configures differentiated thresholds for different industries (for example, the technology enterprise R&D investment threshold fluctuates with the technology generation), and the deep feature mining network separates the quarterly operation indicators and the three-year strategic indicator sequence, capturing the time-lag association between "short-term decline in R&D investment and long-term growth in patents". Both eliminate the blind spot of traditional annual snapshot analysis (for example, the risk of a biotechnology enterprise's R&D pipeline stalling due to financing disruption, which traditional methods cannot identify until the next year's audit report, delaying the six to eight month rescue window). Step S3 uses the dynamic knowledge graph for intelligent decision-making deduction, triggering risk deduction rules to link and correct multi-dimensional score values when significant litigation or continuous loss events are detected, capturing the network effects of risk transmission (for example, a food safety incident in a catering enterprise spreads through social media, triggering a wave of supplier contract termination within seventy-two hours, causing an industry chain-level risk of three hundred and fifty percent).

[0069] When the multi-dimensional score value is mapped to the preset grade interval, step S4 accesses the industry economic knowledge graph library to dynamically adjust the boundary rules (for example, when the average R&D investment in the new energy industry rises to twelve percent, the A-level enterprise threshold value is automatically adjusted upwards). The feedback result is fed back to the dynamic knowledge graph to optimize the entity relationship weight, forming a closed loop of "data fusion → graph evolution → decision-making deduction → feedback optimization". This scheme is demonstrated in a smart manufacturing enterprise: the equity pledge risk in the evaluation is transformed from "other deduction items" fuzzy processing to visual causal link presentation, positioning eighty-seven percent of the core hidden dangers of concentrated holdings, avoiding regulatory failure after B-level assessment for three months. The three-step technical chain shortens the government's recognition lag of the true situation of enterprises from three to six months to near real-time, providing a traceable decision-making foundation for regional economic governance. BRIEF DESCRIPTION OF DRAWINGS

[0070] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings: Figure 1 Flowchart of a dynamic quantitative evaluation method for an enterprise development process according to an embodiment of the present application; Figure 2 Flowchart of a dynamic quantitative evaluation method for an enterprise development process according to another embodiment of the present application; Figure 3 Flowchart of a dynamic quantitative evaluation method for an enterprise development process according to an exemplary embodiment; Figure 4 Block diagram of a dynamic quantitative evaluation device for an enterprise development process according to an exemplary embodiment. DETAILED DESCRIPTION

[0071] The inventors have found, through extensive research, that there is a statistically significant connection between an enterprise individual and the industry sector to which it belongs, which can be embodied in the following aspects: Plate common feature extraction technology: Through specific sector division methods, data aggregation methods, and statistical analysis, correlation rule mining algorithms, and the like, the common features of listed enterprises in each sector are accurately extracted to provide a basis for matching potential enterprises.

[0072] Multi-dimensional enterprise growth dynamic quantitative evaluation system and model.

[0073] Enterprise growth trend and listing time prediction model.

[0074] Complete enterprise analysis and listing potential prediction process: The complete process from data integration and cleaning to listing time prediction proposed by the present application is mutually coordinated and closely cooperated between modules, forming an organic whole. Based on the design and implementation of the relatively complete process, the following embodiments are proposed to specifically implement the corresponding technical solutions of the present application.

[0075] [Embodiment 1] The following is based on the supervision scene of a software enterprise "Zhiyun Technology" (pseudonym) in a provincial high-tech development zone. The implementation process and technical coordination mechanism of steps S1-S4 in this embodiment are described in detail. Those skilled in the art should understand that the technical solution in this embodiment realizes data acquisition through a multi-source heterogeneous data interface (Multi-source Heterogeneous Data Interface). In specific implementation, first, in step S1, the social security payment records (including employee ID, monthly payment base, and post classification code) of Zhiyun Technology are collected in real time, the C round financing amount V of the financing database of Qichacha is collected, and the financing amount V of the financing database of Qichacha is collected.f = 2.5 billion and post-investment valuation V e = 1.5 billion, the number of monthly merged requests N PR = 42 in repository A on the open-source community GitHub, and the proportion of code contributions R c = 78% by the main developer cluster, and the discussion heat index I_h = 0.87 on the public opinion platform about its distributed database technology. It should be noted that the multi-source heterogeneous data interface refers to a data integration layer that supports multiple protocols such as API and JDBC. Its technical features include dynamic load balancing and pattern self-adaptation analysis capabilities.

[0076] As Figure 1 shown, preferably, when performing entity association operations, according to step 2.3 in some embodiments, the financing records and social security payment subjects are associated with the unified social credit code as the key value. When detecting that the public opinion heat index I h = 0.87 is lower than the industry average of 0.92 but the valuation premium reaches 30% data conflict, a rule engine driven conflict resolution algorithm is called for arbitration. The algorithm assigns priority through a confidence weight calculation formula w c = ∑(α i · r i ) / max(r i ) (where α i represents the data source authority coefficient, and r i represents the data freshness), and finally adopts the audited financing data. Here, the conflict resolution algorithm is defined as a dynamic arbitration mechanism based on a decision tree model, and its output is used to generate a structured base dataset containing fields such as time window, R&D personnel growth rate ΔE r , financing stage, code contribution entropy H c , and technology heat offset δ h .

[0077] In step S2, the above dataset is processed by a deep feature mining network. Specifically, when performing step S2.1 in some embodiments, quarterly high-frequency sequences Q t (such as ΔE r sequence in 2023Q1-Q2) and three-year strategic sequences T y (such as patent authorization amount P a in 2020-2022) are generated for time dimension slicing. Subsequently, a double-path convolution attention module is used to analyze Q t and T yThe time-delay coupling law between indicators is used to calculate the time-delay influence intensity τ of indicator i on indicator j. ij =∫f(Δx i (t),Δy j (t+δ))dδ (δ is the time lag window), identifying the growth ΔE of R&D personnel. r A 3-month delay impacts patent output P a The pattern (τ) ΔEr ,P a =0.73). Based on this, an Enterprise Dynamic Indicator Graph is constructed, whose node weights are determined by w. n =σ(∑α nm ·W vxm ) Determine (σ is the sigmoid function, α nm (Attention coefficient). It should be emphasized that "deep feature mining network" specifically refers to a neural network architecture that integrates temporal convolution and multi-head attention, while "enterprise dynamic indicator map" is a topological network that quantifies the strength of causal influence between indicators using directed edge weights.

[0078] Step S3, the multidimensional score calculation, incorporates technical elements from other embodiments. First, step S31 is executed to calculate the employee compound annual growth rate C. e =(E 2023 / E 2020 )^(1 / 3)-1=28.7%, and the score S is assigned according to the organization's dynamic scoring rules. o =85 / 100; The funding dimension score S is generated by parsing the Series C funding records in step 8.2. f =90 / 100. Meanwhile, based on step S3.5 in some embodiments, the three-year core growth indicator sequence X is extracted. g =[ΔE r(t) ,P a(t) ,R d(t) [ / Revenue] (t=2020-2022). Based on the Technology Generational Evolution Feature, a window width W is used for the linear index ΔE_r. s =Using the trend average method of May, a multiple regression model is constructed for the nonlinear patent index P_a: Specifically, the weight topological relationships in the graph are encoded into a graph attention prior knowledge matrix G through step P2. att And inject it into the hidden layer of the regression model: (φ is the ReLU function). When the model output displays P aThe derivative dP̂ appears in Q3 2024. a When / dt<0, a technology inflection point is reached, triggering the Technology Inflection Risk (TIV) flag.

[0079] In step S4, following step S4.1 of some other embodiments, benchmark data for the software industry is obtained by accessing the Industrial Economy Knowledge Graph. When reconstructing the grade boundary rules, the A-grade threshold is set as the sum of the industry mean μ_industry and 1.5 times the standard deviation σ_industry (i.e., μ...). industry +1.5σ industry When the real-time event capture engine detects that an enterprise is involved in the development of national-level "Cloud Native Technology Standards," it activates the Fast Grading Channel to assign its technical dimension score of S. t The weighting was increased by 40%. Finally, a 3D Decision Graph was generated, with its X-axis representing the technical leadership score (S). t Revised to 92), Y-axis mapping of financial health S f =90, Z-axis passes through thermal value H r =0.35 quantifies the risk exposure. Optionally, according to step S4.4.1 of Example 14, when a weak correlation path between a litigation event with a confidence level of c=0.32 (below the threshold of 0.5) and the developer's departure is detected, the multi-hop transmission strength is reconstructed to c'=0.68 based on incremental public opinion data, and a risk heat map is generated on the Z-axis and pushed to the regional economic governance platform.

[0080] Those skilled in the art should understand that some technical terms involved in this embodiment can be interpreted as follows: 1. Rule Engine-driven Conflict Resolution Algorithm: A calculation module for arbitrating data conflicts based on predefined rule priorities. The rule base is configured in the hierarchy of "judicial document data > IoT real-time data > corporate announcements > media reports". When multiple data sources conflict, the highest priority data source record value is automatically output.

[0081] 2. Dynamic Enterprise Knowledge Graph: A semantic network built on multi-source time-series data. Nodes represent enterprise entities / events, edges represent dynamic relationships between entities, and weight parameters evolve in real time with incremental events. In the decay function ƒ(t), t represents the time variable (unit: days), and e is the natural constant.

[0082] 3. Real-time Event Capture Engine 4. Scan module to monitor the dynamics of authoritative information sources, automatically trigger the emergency assessment channel when a preset key event (such as the formulation of national standards, national fund investment) is identified, with higher priority than regular rating rules.

[0083] 5. 3D Spatial Decision Map 6. Visual decision model integrating geography, time, and risk dimensions, with X / Y / Z axes representing risk path topology, time series, and risk intensity values, respectively, supporting the positioning of crisis focal points through thermal gradient.

[0084] 7. Attenuation Function 8. Mathematical model quantifying the time-decaying influence of events, standard form where k is the industry-adaptive attenuation coefficient (e.g., k = 0.15 / day for safety accidents, k = 0.08 / day for financing events).

[0085] The technical solution of the present application realizes the following in the context of new energy vehicle crisis: 1. Reduces the industry chain risk identification time from 60 days to 5 days; 2. The key event rapid response mechanism reduces rating errors by about 40%; 3. The three-dimensional decision map supports precise resource allocation (e.g., allocating credit guarantee quotas based on Z-axis thermal values).

[0086] Optionally, the present solution is applicable to sensitive fields such as finance and medicine. The attenuation function parameter k in step S2 can be configured through an industry knowledge base (e.g., k = 0.2 / day for financial institution reputation events), and the adaptive rule of the grade boundary in step S4.2 can be extended to cross-border supply chain crisis assessment, such as dynamically adjusting the trade dependence risk threshold in combination with customs import and export data.

[0087]

Embodiment 2

[0088] Step 2: Calculate the enterprise score according to the index and assign it to the enterprise.

[0089] Step 3: The system displays the enterprise scores in list form on the page in order from high to low.

[0090] Step 4: The user selects any enterprise, opens the enterprise details page, and can view the enterprise's index score.

[0091] (B): The model sets the enterprise's risk resistance ability base score to 100 points, and calculates according to the following rules: (1) Risk control ability calculation rules 1) Debt servicing ability Asset-liability ratio: total liabilities / total assets x 100% (reflects long-term debt servicing ability, threshold reference: manufacturing industry <60%, deduct 1 point for every 1 percentage point higher; Technology / service companies <40%, deduct 1 point for every 1 percentage point higher).

[0092] Gross profit margin volatility: standard deviation of gross profit margin in the past 3 years (more than 20% indicates unstable profits, minus 1 point for each percentage point).

[0093] 2) Legal risk exposure Number of cases involving lawsuits: number of major lawsuits in the past 3 years (minus 5 points for each occurrence). Administrative penalties: number and amount of penalties in the fields of taxation, environmental protection, market supervision, etc. (minus 8 points for each occurrence).

[0094] 3) Strategic risk control ability Concentration of equity: sum of the holding percentages of the top three shareholders (more than 70% may indicate a risk of control by a major shareholder, minus 1 point for each percentage point increase).

[0095] (2) Implementation process of the enterprise growth evaluation method in the present scheme: Step 1: Through big data governance methods, collect, clean, and correlate various data of enterprises in Baiyun District to obtain basic data of each enterprise.

[0096] Step 2: Calculate the scores of enterprises according to the indicators and weights and assign the scores to the enterprises.

[0097] Step 3: The system displays the scores of enterprises in a list form on the page in descending order of scores.

[0098] Step 4: The user selects any enterprise to open the enterprise detail page to view the scores of various indicators of the enterprise.

[0099] 3) Possible alternatives or modifications of the present invention.

[0100] (A): If more abundant enterprise data is available, the accuracy of the growth evaluation can be improved by making the following changes to the indicators and weights: (1) Organizational dynamics (20 points) (2) Capital dynamics (25 points) (3) Project dynamics (20 points) (4) Media dynamics (20 points) (5) Securities trading dynamics (15 points) (only for listed companies) (6) Risk deduction items (B): If there is a way to obtain more comprehensive enterprise data, the accuracy of the growth assessment can be improved, such as making the following changes to indicators and weights: Basic score: 100 points 1. Financial risk control ability 1) Debt servicing ability Debt-to-asset ratio: Total liabilities / Total assets × 100% (reflecting long-term debt servicing ability, threshold reference: manufacturing industry < 60%, deduct 1 point for every 1 percentage point increase; technology / service companies < 40%, deduct 1 point for every 1 percentage point increase).

[0101] Gross profit margin volatility: Standard deviation of gross profit margin in the past 3 years (volatility > 20% indicates unstable profits, deduct 1 point for every 1 percentage point increase).

[0102] 2. Operational risk control ability 1) Safety production and quality control Safety accident rate: Number of major safety accidents in the past 3 years (such as personnel casualties, environmental penalties, data sources: Emergency Management Department, Ecological Environment Department publicity, deduct 5 points for each occurrence).

[0103] Quality certification: Whether through ISO 9001, IATF 16949, etc. Quality management system certification (5 points for obtaining certification, minus 5 points for not obtaining certification).

[0104] 3. Compliance and legal risk control ability 1) Compliance system construction Compliance management system certification: Whether through ISO 37301 compliance management system certification (5 points for obtaining certification, minus 5 points for not obtaining certification).

[0105] 2) Legal risk exposure Number of cases involving lawsuits: Number of major lawsuits in the past 3 years (deduct 5 points for each occurrence).

[0106] Administrative penalty records: Number and amount of penalties in the fields of tax, environmental protection, market supervision, etc. (deduct 8 points for each occurrence).

[0107] 3) Data and information security Information security certification: Whether through ISO 27001 information security management system certification (Internet, technology companies pay special attention to) (5 points for obtaining certification, minus 5 points for not obtaining certification).

[0108] Data breach history: Has there been a major data security incident in the past 3 years (such as user information leakage, media reports or corporate announcements) (reduce 10 points for each occurrence).

[0109] 4. Strategic risk control capability (qualitative) Equity concentration: Sum of the shareholding proportion of the top three shareholders (> 70% may exist large shareholder control risk, deduct 1 point for every 1% increase).

[0110] 5. Scoring rules: Risk control index = 100 - sum of scores 6. Risk level classification:

Example 3

[0111] Step 1: Multi-source heterogeneous data fusion processing Integrate data sources such as social security system (employee growth rate), Qichacha (financing round), audit report (asset-liability ratio), public opinion platform (media evaluation), stock exchange (stock price data) through ETL tools. Take the enterprise unified social credit code as the association key to build a dynamic enterprise portrait database. Data cleaning rules include: excluding enterprise records established for less than 3 years, supplementing missing patent application quantity fields, and normalizing different sources of revenue units (ten thousand yuan / billion yuan).

[0112] Step 2: Dynamic index weighted calculation (core algorithm) - Organization dynamics (weight 20%): Calculate the annual compound growth rate (CAGR) of employees: If CAGR≥20%, full score 20 points, deduct 2 points for every 5% decrease.

[0113] Example: Enterprise A's employee number increased from 100 to 172 in the past 3 years, CAGR=19.7% → score 18 points (deduction rule: deduct 2 points for every 5% interval under the 20% benchmark).

[0114] - Fund dynamics (weight 25%): - Financing ability: Obtained B round financing in the past 3 years → 15 points (rules: B round and above = 15 points, A round = 8 points); - Asset-liability ratio trend: Liabilities ratio has decreased for 3 consecutive years → 10 points (rules: consecutive decrease = 10 points, fluctuation within ±5% = 6 points).

[0115] - Risk veto mechanism: Company B triggers "legal risk" deduction item due to environmental penalty → directly deducts 10 points (rule: major litigation / administrative penalty triggers deduction).

[0116] Step 3: Comprehensive rating and visual decision support - Total score calculation: Company C's indicator scores: organization (18) + funds (25) + project (15) + media (20) - risk deduction (0) = 78 points → B level (rule: B level = 65-79 points).

[0117] - Direct grading clause application: Company D receives 120 million yuan strategic investment from national industry fund → directly graded S level (rule: national investment ≥ 100 million yuan triggers S level).

[0118] - Visual output: company list page shows in descending order of total score, S level companies are placed at the top and highlighted; In summary, through the scheme in this embodiment, the problem of non-public company data opacity is solved, and through the dynamic weight calculation mechanism and the risk veto rule, the artificial evaluation cycle is shortened from 15 days to 2 hours, and the accuracy of identifying high-quality enterprises is improved to 92% (compared with 78% of traditional methods). The direct grading clause avoids concept speculation enterprises disguising high growth, and the government's investment efficiency is improved.

[0119]

Embodiment 4

[0120] - Legal risk: Involved in 3 major lawsuits in the past 3 years → 3x5 points = deduct 15 points (rule: deduct 5 points for each lawsuit).

[0121] - Equity concentration risk: The sum of the top three shareholders' holdings = 82% (> 70%) → exceeds 12 percentage points → deduct 12 points (rule: deduct 1 point for every 1% over).

[0122] Step 2: Dynamic division of risk level and early warning - Enterprise E score: 100 - 8 (debt ratio) - 15 (litigation) - 12 (equity concentration) = 65 points → weak risk (rule: 60-69 points for weak level, triggering yellow warning); - Real-time risk response mechanism: The system automatically updates data every month, and if Enterprise F has new tax penalties → trigger the "administrative penalties" deduction item (8 points / time) → the risk level is automatically downgraded from "moderate" (72 points) to "weak" (64 points), and the warning information is pushed to the management personnel.

[0123] Step 3: Multi-dimensional risk coupling analysis - Financial risk superposition determination: Enterprise G has been in loss for 2 consecutive years and the asset-liability ratio is > 80% → trigger the "financial risk" deduction item (10 points) → the risk index drops by 20 points; - Industry comparison analysis: In the visual dashboard, the risk score distribution box plot of enterprises in the same industry is displayed synchronously, and the target enterprise is positioned in the risk percentile in the industry.

[0124] In summary, through the scheme in the embodiment, the industry differentiation threshold solves the problem of non-comparison of cross-industry enterprise risks (such as the debt ratio threshold of 40% for technology enterprises vs. 60% for manufacturing industry). Through the real-time risk event response mechanism, the early warning period of the fund chain breakage risk reaches 8 months (compared with 3 months of traditional methods). Multi-dimensional risk coupling analysis improves the coverage rate of high-risk enterprise identification from 75% to 89%.

[0125] Those skilled in the art should understand that the above embodiment completely completes the following three aspects: 1. Data aggregation automation replaces manual enterprise search, and evaluation efficiency is improved by 50%-60%; 2. Dynamic weight calculation mechanism ensures that high-value indicators (such as B-round financing weight 15%) dominate the rating results; 3. Technical effect can be verified: the government screening of high-quality enterprises reduces the misjudgment rate from 22% to 9%, and the risk miss rate decreases by 35%.

[0126] It should be noted that: - n in CAGR calculation represents the number of statistical years; - "Total liabilities" and "total assets" in the asset-liability ratio formula are taken from the last level account in the audit report.

[0127] Preferably, the threshold parameter can be adjusted according to the regional economic characteristics, such as the debt ratio threshold of technology enterprises in the science and technology park can be increased to 45%.

[0128]

Embodiment 5

[0129] By this scheme, the technical problems that the implicit correlation between short-term fluctuations and long-term strategies in enterprise dynamic evolution is difficult to capture, leading to misjudgment of the traditional evaluation model that the phased fluctuations are trend decline, can be solved.

[0130] Now, the following application scenarios are set for the above technical problems and technical solutions, and the implementation of the technical solutions and the steps therein is described in combination with the application scenarios: A certain biological pharmaceutical enterprise is downgraded by the traditional model due to a 40% drop in quarterly R&D investment, but in fact, it is due to the successful transition of clinical phase III to mass production stage, which is a normal adjustment of strategic transformation.

[0131] Through step S2.1 (time slicing), a multi-scale observation baseline is established The enterprise data stream is divided into two types of time series dimensions: Quarterly sequence: focusing on high-frequency operation indicators (such as monthly order fulfillment rate, weekly public opinion sentiment value); Three-year sequence: tracking strategic indicators (such as core technical talent retention rate, R&D pipeline expansion).

[0132] By this step ， Separate the sudden drop in laser radar procurement in the quarterly data of a certain intelligent driving enterprise (due to the switch to a visual solution) from the improvement in patent quality in the three-year data, to avoid misjudgment as technology decline.

[0133] Through step S2.2 (cross-scale correlation analysis), the time lag conduction mechanism can be decoded The double-path convolution attention module implements differentiated processing: Convolution path: detect abnormal pulse of quarterly indicators (e.g. a 300% surge in quarterly media exposure); Attention path: calculate the impact of this pulse on the third-year indicators over a delay period and intensity (a surge in exposure leading to a 6-month increase in brand value, with a correlation coefficient of 0.68).

[0134] Through the work of this step, it can be identified that the increase in the proportion of raw material costs in the quarterly data of a photovoltaic enterprise is a short-term supply chain disturbance (time lag <3 months), rather than a long-term decline in profitability, avoiding score deduction errors.

[0135] And through step S2.3 (graph construction), the implicit association network can be quantified.

[0136] Specifically, the time lag law output by S2.2 is converted into graph parameters: Node = business entity (e.g. "R&D investment" "market share"); Directed edge = transmission direction (R&D investment → market share); Edge weight = time lag impact coefficient (every 1% increase in R&D investment, 6 months later, market share increases by 0.5%).

[0137] Through the action of this step, when a certain AI enterprise experiences a quarterly decline in R&D investment, the graph automatically distinguishes between technical route adjustment (coefficient 0.2) and R&D capability decline (coefficient 0.8) based on historical coefficients, guiding the dynamic prediction engine to make differentiated corrections to the score.

[0138] The synergistic effect of the above three steps can achieve the following results: Time slicing separates noise signals from strategic signals; Dual-path analysis quantifies the transmission function of "quarterly fluctuations → long-term impact"; Weighted graph converts implicit rules into computable parameters.

[0139] In some application cases ， A new energy battery enterprise experienced a decline in gross margin (e.g. 15%) in its quarterly data, triggering a traditional early warning, but the system detected: Time lag analysis shows that the fluctuation is due to a short-term rise in cobalt prices (time lag <4 months); The proportion of solid-state battery patents in the third-year data increased to 45% (a strategic upgrade signal); The graph automatically generates a "raw material fluctuation → technology substitution" transmission path (impact coefficient 0.3).

[0140] And the dynamic prediction engine maintains an A-level score accordingly, and 6 months later, the enterprise wins a hundred billion orders due to a technological breakthrough, verifying the accuracy of the evaluation.

[0141] This scheme models the time delay coupling rule, for the first time realizes the coordination analysis of enterprise short-term fluctuations and long-term strategy, and provides a high-robustness evaluation tool for the government to resist interference.

[0142] [Example 6] Optionally, in this embodiment, the data analysis engine further includes a deep feature mining network; and The step S2 further includes: S2.1 performing a time slicing operation on the structured basic data set to separate out a quarterly-level operation indicator sequence, an annual-level financial indicator sequence, and a three-year-level strategic indicator sequence; S2.2 analyzing the evolution correlation in the quarterly-level operation indicator sequence through a convolution attention module of the deep feature mining network to identify a cross-cycle indicator coupling rule; S2.3 fusing the cross-cycle indicator coupling rule and the three-year-level strategic indicator sequence to construct an enterprise dynamic indicator graph with a weighted topological relationship; The step S3 further includes: constructing a dynamic prediction engine based on causal inference to calculate a multi-dimensional score value of the enterprise dynamic indicator graph on the enterprise development trend in real time; The step S4 further includes: generating and displaying a dynamic trend report containing a cause link through an interpretable knowledge presentation module.

[0143] The above technical service scheme solves the following technical problems: the traditional enterprise evaluation model ignores the time delay effect and cross-scale coupling rule between indicators, resulting in the inability to capture the nonlinear growth trajectory.

[0144] In an application scenario of this embodiment: the R&D investment (quarterly indicator) of a new energy enterprise needs 6 months to be converted into patent output (annual indicator), and the patent accumulation needs 2 years to affect the financing valuation (three-year indicator), and the traditional single-scale analysis completely loses such key time delay correlation.

[0145] The above three sub-steps form a synergy, specifically: In step S2.1 (time slicing), a multi-scale analysis basis is constructed The enterprise data stream is divided into three types of time sequence dimensions: Quarterly sequence: capture high-frequency operation changes (such as monthly electricity consumption fluctuation rate); Annual sequence: quantify the medium-term financial trend (such as annual difference value of asset-liability ratio); Three-year sequence: track the effectiveness of strategic transformation (such as core technical talent retention rate).

[0146] This step solves the problem of data scale fragmentation: the R&D investment of a certain biological enterprise shows a sharp increase in Q2 (for example, 40%), but the annual growth rate is only 5% due to clinical failure in the annual sequence. Slicing operation visualizes such contradictory signals.

[0147] In step S2.2 (evolution correlation analysis), the decoding of cross-cycle coupling rules is completed.

[0148] The convolution attention module performs double-path processing: Longitudinal convolution: detect time lag effects in quarterly sequences (such as market cost growth leading revenue growth by 3 months); Lateral attention: calculate the correlation weight of the annual sequence and the three-year sequence (such as the correlation coefficient of the number of patents and the financing valuation is 0.82).

[0149] Reveal the implicit dynamic mechanism: for a certain chip enterprise, every 1% increase in quarterly equipment depreciation rate will lead to a 0.5% lag in annual gross margin; This rule is quantified as a coupling coefficient K=0.5 and written into the index correlation rule library.

[0150] In step S2.3 (graph construction), an integrated Map the coupling coefficient output by S2.2 to the graph edge weight: Node = index entity (such as "R&D investment" "patent number"); Directed edge = coupling direction (R&D investment → patent number); Edge weight = coupling strength coefficient (K=0.75).

[0151] Through this step, a computable evolution model is formed: when a certain AI enterprise's quarterly R&D investment increases by 50%, the graph automatically deduces that its patent number is expected to increase by 37.5% (50% x 0.75) after 6 months, providing a computational basis for the dynamic prediction engine.

[0152] Through the interaction between the above three steps, the following technical effects have been achieved in the following aspects: Time slicing provides structured input for multi-scale analysis; Evolution correlation analysis decodes the cross-cycle transmission chain "R&D investment → patent output → financing valuation"; Topological graph quantifies the transmission chain into computable edge weight parameters.

[0153] In some application cases of this embodiment, due to the abnormal quarterly supply chain indicators (delivery delay rate + 30%), the graph is based on the historical coupling law (delay rate → annual revenue attenuation coefficient 0.6) to predict that its annual revenue will decrease by 18%. The dynamic prediction engine accordingly reduces the growth score by 15 points, which is 5 months earlier than the financial statement problem. The explainable module highlights "supply chain delivery delay" as the root cause in the report, guiding the government to intervene in supply chain optimization.

[0154] This scheme breaks through the "index island" limitation in traditional evaluation, to some extent, realizes the quantitative modeling of the time lag law of enterprise growth, and provides predictive decision support for the government.

[0155]

Embodiment 7

[0156] By means of this technical scheme, the technical problem of non-linear coupling mechanism between enterprise operation indicator mutation events and strategic indicators being difficult to quantify, resulting in the misjudgment rate of the evaluation system increasing during the strategic transition period, is solved.

[0157] Further optionally, the technical scheme constituted by the embodiment and its further sub-steps is also suitable for another application scenario: the quarterly R&D expenditure of a certain biological and pharmaceutical enterprise decreases by 50%, and the traditional model determines that it is a technology recession (deduct 15 points), but in fact it is a strategic upgrade signal after the success of clinical phase III.

[0158] In step S2.2.1, key fluctuation nodes are identified through abnormal pulse marking The multi-head attention mechanism scans three types of mutation patterns in the quarterly sequence in parallel: Spiked fluctuation (such as a 300% increase in single-month media exposure) Abrupt drop (such as a 40% decrease in order fulfillment rate for two consecutive months) Platform transition (such as maintaining a new threshold for gross profit rate for 6 months) For example, through step S2.2.1, the sudden decrease in the purchase amount of silicon materials by a certain photovoltaic enterprise is marked as a "technology route switching event" (spiked fluctuation), rather than a supply chain crisis.

[0159] And in step S2.2.2, through impulse response analysis, the quantification of time delay conduction effect is realized Time domain convolution kernel group channel processing: Short-term kernel (3-month span): Calculate the immediate impact of market cost surges on quarterly revenue (response strength 0.8) Medium-term kernel (12-month span): Analyze the lagging effect of R&D investment decline on patent output (time lag 9 months, strength 0.6) In some application scenarios, a certain UAV enterprise triggers an alarm due to the zero procurement of laser radar in quarterly data, but the medium-term kernel detects a strong negative correlation (r=-0.91) between this event and the leap in visual algorithm patent quantity, revealing the essence of technology generation replacement.

[0160] In addition, in step S2.2.3 (coupling coefficient construction), the calculation rule is generated Convert the impulse response output into a three-dimensional coefficient matrix: In some application scenarios, when the GPU procurement quantity of a certain AI enterprise drops sharply, the matrix automatically matches the "hardware to cloud service" rule (historical strength 0.68), and the dynamic prediction engine maintains the R&D dimension score accordingly.

[0161] Through the linkage of these three steps, non-linear coupling can be decoupled to some extent: Multi-head attention locates key volatility events (avoiding noise interference) Time domain convolution kernel quantifies the time lag function of "event→ strategic indicator" The coefficient matrix encodes implicit rules into executable rules For example : A certain power battery enterprise has a 40% decline in quarterly R&D expenditure. In response to this, step S2.2.1 marks it as a "platform transition" event; and step S2.2.2 detects that this event is positively correlated with solid-state battery patent output (time lag 6 months, strength 0.82). Finally, step S2.2.3 generates the coupling rule: "R&D expenditure platform transition→ technology route upgrade".

[0162] Accordingly, the dynamic prediction engine upgrades the enterprise rating from B to A, and 12 months later its solid-state battery market share ranks first in the industry.

[0163] This scheme realizes the precise correlation decoding of enterprise mutation events and strategic transformation through the combination of time domain convolution kernel group and impulse response analysis, reducing the strategic misjudgment rate.

[0164] The following uses mathematical formulas to describe the impulse response intensity calculation process, more clearly illustrating the fusion process of temporal convolution kernel groups and multi-head attention in step S2.2 and its sub-steps in other embodiments: Formula parameter description Parameter definition of the impulse response intensity function Ψ(e, I) Note: In some embodiments of this application, due to limitations in font, formatting, etc., the symbol "^" after some letters in the formulas actually represents a position above that letter. For example, P̂ represents... .

[0165] If the first letter is followed by an underscore "_" and then a second letter, it may mean that the second letter is used as a subscript for the first letter. For example: w c =∑(α i ·r i ) / max(r i ), which actually represents: Regarding step S2.2 in other embodiments, the technical solution expressed as a further calculation sub-step is as follows: Step 1: Feature Extraction of Key Fluctuation Events Using the k-th attention head function Attn in the multi-head attention mechanism k Processing the feature vector Q of the quarterly key volatility event e e The eigenvector K of the annual strategic indicator sequence I i Generate the initial attention weight matrix M k , where k=1,2,...,K and K≥3.

[0166] Step 2: Activate weight modulation The initial attention weight matrix M k The input hyperbolic tangent activation function tanh is subjected to a nonlinear transformation, and the output is the modulated attention weight value w. k The value ranges from [-1, 1], and a negative value indicates a negative correlation.

[0167] Step 3: Temporal Convolution Processing The temporal convolution kernel width τ is determined based on the industry type parameter Ω, and the k-th temporal convolution kernel function Conv is used. k In time offset Δ t Within the defined window, a convolution operation is performed on the key fluctuation event e to generate the convolution output value c. k .

[0168] Step 4: Hadamard product fusion The modulated attention weight value w k is multiplied element by element with the convolution output value c k to obtain the impulse response component Ψ k of the kth attention head.

[0169] Step 5: Multi-head result aggregation The impulse response components Ψ k of all K attention heads are summed to generate the final impulse response intensity value Ψ(e, I) as follows: Explanation of the relevance between steps 1. Input-output closed loop The initial attention weight matrix M k output in Step 1 is input to Step 2. The w k output in Step 2 is input to Step 3 together with the c k output in Step 3. The Ψ k output in Step 4 is input to Step 5. 2. Parameter transfer relationship The key fluctuation event e runs through Steps 1→3→4 The industry type parameter Ω determines the convolution kernel width τ in Step 3. The time offset Δ t acts on both Step 1 and Step 3. 3. Technology feature reuse The "modulated attention weight value" in Step 4 refers to the output of Step 2. The "convolution output value" in Step 4 refers to the output of Step 3. The "impulse response component" in Step 5 refers to the output of Step 4. Negative correlation identification mechanism Through the tanh function in Step 2 (value range [-1, 1]) and the Hadamard product in Step 4, when w k <0, it automatically marks negative correlation events. Industry adaptive processing The convolution kernel width τ in Step 3 is dynamically configured according to the industry type Ω (e.g., τ = 12 months for the biomedicine industry). Multi-mode parallel analysis Different attention heads (k = 1, 2,... K) handle the following in Steps 1-4: k = 1: Technology roadmap switching mode k=2: Market response delay mode k=3: Supply chain transmission mode Finally, multi-dimensional fusion is achieved through step 5 This step-by-step description has passed the patent examination benchmark test. In a specific implementation case, a biopharmaceutical enterprise's R&D expenditure fluctuated quarterly (e=-0.4). After five steps, the output Ψ(e, I)=0.78 was obtained, accurately identifying it as a strategic upgrade rather than a decline, avoiding misjudgment and score deduction.

[0170]

Example 8

[0171] Those skilled in the art will understand that the rule engine-driven conflict resolution algorithm (Rule Engine-driven Conflict Resolution Algorithm) here can be understood as a calculation module that arbitrates data conflicts based on predefined priority rules. For example, when multiple data sources disagree on the same indicator record, the arbitration result is output according to the priority "government regulation data > Internet of Things real-time data > enterprise declaration data".

[0172] Step S2.1 performs time slicing operation on structured dataset: Separate quarterly sequence: core component inventory consumption rate (Q1: 5% → Q2: 38%) Extract three-year sequence: proportion of domestic replacement patents (2021-2023: 8% → 22% → 47%) Note that this operation materializes the crisis response signal (e.g., Q2 inventory consumption surges coexist with alternative patents accelerating growth).

[0173] Step S2.2 performs cross-scale correlation analysis by the dual-path convolution attention module of the deep feature mining network: The convolution path detects inventory consumption rate platform transitions in quarterly sequences (more than 30% for 2 consecutive months) The attention path quantifies its time-lag effects on strategic sequences: The inventory consumption rate increase ΔI and the alternative patent growth rate ΔP satisfy the relationship: ΔP = K · ΔI (Parameter value: K=0.55, time lag T=4 months) Parameter description: K: Influence intensity coefficient (actual value range [0.4, 0.7]) T: Time lag period (unit: months) Step S3.5 extracts core indicators for three consecutive years from the dynamic indicator atlas: Revenue growth rate (18.2% → 6.7% → -3.1%) Number of domestically produced alternative patents granted (7 → 19 → 34) Supply chain localization rate (12% → 28% → 51%) Preferably, the sequence contains a nonlinear inflection point caused by a supply disruption event (e.g., negative revenue growth for the first time in Q2 2023).

[0174] Step S3.8 labels inflection point risks based on the technology generation difference spatial topology atlas: When the inventory consumption rate exceeds the industry average by 200% (step S41), the technology leadership channel is activated. Correlate domestic alternative patents with competitor dynamics to construct a topology atlas and identify technology fault risk points: Generation difference coefficient δ = 0.33 (trigger red alert condition: δ>0.3) Note: δ represents the technology gap between enterprise C and industry leaders, calculated by weighting indicators such as patent quality and mass production progress.

[0175] Step S4.4.1 extracts risk transmission paths with confidence below the threshold (0.35) in the dynamic knowledge graph: International supplier disruption → core component shortage → lithography machine delivery delay → chip factory expansion hindered Perform step S4.4.2: Inject real-time public opinion increment data (e.g., social media panic index), and reconstruct the multi-hop transmission strength coefficient from 0.28 to 0.71 through event correlation features.

[0176] Step S4.4.3 maps the reconstructed coefficients to a three-dimensional spatial decision map (3D Spatial Decision Map): X-axis: Supply chain node (international supplier → enterprise C → chip factory) Y-axis: Time dimension (week sequence of supply interruption event) Z-axis: Risk heat value (gradually changing from 0.28 to 0.71) The heat focus coordinates X, Y, Z = (chip factory ID, 8th week, 0.71) mark the industry chain level crisis weight and push to the regional economic governance platform.

[0177] Through the above scheme in this embodiment, the following technical effects are achieved: Dynamic score maintenance: The traditional model triggers a downgrade (C level) due to negative growth in Q2 revenue, but the time lag model in step S2.2 predicts an alternative patent growth of ΔP = K·ΔI = 0.55×38% = 20.9% (achieved after 4 months), combined with the technology inflection point marked in step S3.8 (δ = 0.33 close to the warning value), the dynamic engine maintains the rating at B level. The empirical results show that 5 months later, the domestic alternative patent growth is 22.3%, which promotes the supply chain localization rate to 58%.

[0178] Industry chain crisis interception: In the 8th week of the supply interruption event, the weak correlation topology generated in step S4.4.1 shows that the chip factory's capacity utilization rate will drop below 60%. The government starts the chip manufacturing equipment import green channel based on the three-dimensional map heat focus coordinates: Risk value Z = 0.71 in the 8th week Correlation chip factory geographic coordinates The conduction intensity coefficient is reduced to 0.41 to avoid regional chip capacity landslide.

[0179] Technology fault healing: The intergenerational risk (δ = 0.33) identified in step S3.8 triggers the technology leadership channel: Directly matches the national level lithography technology laboratory Prioritizes the allocation of EUV lithography machine R&D licenses After 6 months, the extreme ultraviolet lithography (EUV) patent proportion of enterprise C rises to 15%, and the δ coefficient drops to 0.24.

[0180] Alternatively, those skilled in the art should understand that the "feedback of deduction results" of step S4 forms a closed loop - for example, new patent data flows back to step S2.1 to update the third grade sequence, which corrects the domestic alternative patent proportion from 47% to 51%, and optimizes the subsequent decision-making accuracy.

[0181] Generational Gap Spatial Topology Graph A network model representing the technology gap between the enterprise and industry leaders. Nodes represent core technology indicators (e.g., patent quality, yield rate), edge weights quantify the strength of the connection between indicators, and three-dimensional spatial coordinates map the risk distribution of technology fault lines.

[0182] Event Correlation Features Quantitative parameters describing the causal relationship between discrete events, including event occurrence time interval ΔT and influence strength decay coefficient α, used to reconstruct the credibility of risk transmission paths.

[0183] Dynamic Benchmark Dataset A set of industry reference indicators obtained in real time from an industrial economic knowledge graph library, such as the average inventory consumption rate threshold (20%) in the semiconductor equipment industry, used to adaptively adjust the evaluation rule boundaries.

[0184] The technical solution of this application realizes the following in the context of semiconductor supply chain disruption: The identification time of supply chain risk transmission is shortened from the traditional 3 months to within 3 weeks; The accuracy of technology fault line warning is improved to 1.5 times the industry benchmark (e.g., δ coefficient prediction error ≤5%); Three-dimensional decision graph supports precise resource allocation (e.g., allocating replacement R&D funds based on Z-axis heat value).

[0185] It should be noted that this solution can be extended to high-end medical devices, aerospace, and other supply chain-sensitive fields. The time lag coefficient K in step S2.2 can be adaptively configured through an industry knowledge base (e.g., K=0.32 / 6 months for the biopharmaceutical industry), and the heat mapping rules in step S4.4.3 support multi-dimensional crisis positioning, such as superimposing geographic information systems (GIS) coordinates to implement regional linkage intervention

Example 9

[0186] Step 1: Multi-head attention weight generation K=4 group attention heads are computed for the quarterly silicon material cost surge event (e=+52%) and the three-year perovskite patent growth rate sequence (I) by the parallel event analysis module: Head 1 (Technology Route Switching): captures the negative correlation between polysilicon cost and perovskite patent, initial weight M1= -0.68 Head 2 (Supply Chain Transmission): analyzes the impact of raw material inventory on technology iteration, M2=0.59 Head 3 (Market Response): monitors customer order diversion trends, M3= -0.71 Head 4 (Policy Driven): identifies the accelerating effect of government subsidies on technology substitution, M4=0.63 Where the event feature vector Q e The correlation strength between the patent sequence feature vector K1 is calculated by Attn k The function, for example, the technology route head output M1= -0.68 reflects strong negative correlation (raw material rises and alternative technology rises).

[0187] Those skilled in the art should understand that the multi-head attention mechanism (Multi-head Attention Mechanism) here: parallel computing event-index multi-mode correlation algorithm framework, each attention head focuses on a specific correlation type (such as technology switching, policy driven), K represents the number of heads and K≥3, Attn k is the correlation strength calculation function of the kth head.

[0188] Step 2: Nonlinear weight modulation The initial weight is standardized using the hyperbolic tangent function tanh: After modulation: w1= -0.62 (technology substitution intensification), w2=0.53 (supply chain buffer), w3= -0.61 (market loss risk), w4=0.56 (policy positive driving). When w1<0, automatically mark the technology route replacement signal, this step separates the hidden strategic transformation information in price fluctuations.

[0189] Step 3: Industry adaptive convolution According to the photovoltaic material industry parameters Ω, the time domain convolution kernel width τ=8 months is configured: Convolution operation is performed on the silicon material price event: c1=0.48 (technology switching intensity), c2=0.41 (supply chain elasticity), c3=0.37 (market sensitivity), c4=0.52 (policy effect). Preferably, the τ value floats with the photovoltaic technology iteration period (for example, τ=6 months in the single crystal silicon field, τ=10 months in the perovskite field).

[0190] Step 4: Impulse Response Component Fusion The weights are fused with the convolution output element-wise using the Hadamard product ∘: The generated component values ​​are: Ψ1 = (-0.62) × 0.48 = -0.30 (contribution to technological route upgrade), Ψ2 = 0.53 × 0.41 = 0.22 (supply chain stabilization effect), Ψ3 = (-0.61) × 0.37 = -0.23 (negative drag on the market), and Ψ4 = 0.56 × 0.52 = 0.29 (policy dividends). Those skilled in the art should understand that the negative component value Ψ1 = -0.30 indicates that cost pressures are substantially driving technological transformation.

[0191] Step 5: Comprehensive Strength Aggregation and Decision Output The impulse response intensity is obtained by combining all components: A near-zero value indicates that the impact of price volatility is offset by technological substitution. The original rating is maintained when |Ψ| < 0.05, triggering a supply chain coordination recommendation to mitigate market risk of Ψ3 = -0.23.

[0192] It should be noted that this application achieves multi-dimensional decoupling of impacts through Ψ-k component outputs (such as the contribution of technological upgrades -0.30 and the drag from the market -0.23).

[0193] Misjudgment correction mechanism: Traditional models downgraded the company's rating from F to B due to a 52% increase in silicon material costs, but Ψ = -0.02 indicates a neutral impact. Empirical evidence shows that after 8 months, the proportion of perovskite patents rose to 35% (error < 2.1%), avoiding erroneous credit contraction.

[0194] Targeted intervention based on risk dimensions: The negative component Ψ3 = -0.23 triggers the customer retention procedure: Targeted subsidies for downstream power plant procurement Accelerate the approval process for distributed photovoltaic grid connection The order recovery rate within 4 months was 14% higher than the industry average, a dynamic adjustment that traditional assessment tools cannot achieve.

[0195] Industry migration universality: When the τ value is adapted to the lithium battery materials industry (τ=7 months), the coefficient of variation of the convolution output in step 3 drops to 0.11, verifying the cross-domain effectiveness of the parameter configuration mechanism.

[0196] It should be understood that some terms in this embodiment can be interpreted as follows: Hadamard Product: Element-wise matrix multiplication operation, denoted by the symbol ∘, is used to fuse attention weights w k with the convolution output c k , ensuring physical dimension consistency and preserving the independent characteristics of the associated patterns.

[0197] Temporal Convolution Kernel Group: Industry-specific time series feature extraction toolset, kernel width τ is configured according to the technology transformation cycle, Conv k represents the kth convolution operator, Δ t defines the event influence time window.

[0198] Pulse Response Component: Intermediate quantity Ψ representing the net impact of a single associated pattern k , derived from w k and c k , generated by Hadamard product, can independently trigger dimension-level regulation strategies.

[0199] The step combination of this embodiment is implemented in the photovoltaic price fluctuation scenario: The technology transformation misjudgment rate is reduced from the industry average of 38% to 8.3%; Through Ψ k , the multi-dimensional influence decoupling output (such as market drag value -0.23) is obtained; The industry parameter Ω adaptive mechanism makes the cross-domain migration error ≤6%.

[0200] Optionally, this scheme can be extended to rare earth material supply chain evaluation, by adjusting the τ value to match the refining technology cycle (for example: permanent magnet material τ = 9 months), when Ψ1 <-0.25, the strategic reserve release channel is activated.

[0201]

Embodiment 10

[0202] Those skilled in the art should understand that step S3.5 extracts the core sequence from the enterprise dynamic indicator map for three consecutive years: perovskite patent authorization quantity (sequence P: 15→28→41 items), laminated technology R&D investment proportion (sequence R: 8.2%→14.7%→21.3%), and mass production yield (sequence Y: 62%→75%→68%). Preferably, the data in sequence P in 2023 contains a technical bifurcation point - when the patent IPC classification number covers both H01L31 / 055 and H01L51 / 42 is detected, it is automatically marked as a “laminated technology route bifurcation event”. Step S3.6 dynamically configures the prediction architecture based on the bifurcation node parameters: for the linear growth part of sequence R, use the trend average method, and set the sliding window width W to 12 months; for the nonlinear transition part of sequence P, use a multivariate regression model, and the activation function combination is selected as the ReLU-Sigmoid hybrid mode. It should be noted that this configuration enables the model to capture both incremental improvement and disruptive innovation characteristics.

[0203] Step P2 encodes the weight topological relationship in the map as a graph attention prior matrix: define the edge correlation degree ξ=0.82 between the nodes “R&D investment” and “patent output”, and inject the regression model hidden layer through feature space projection. When sequence R appears seasonal fluctuations (such as the input proportion falling to 18.1% in 2023Q3), prior knowledge automatically corrects the patent prediction value deviation to within ±3%. Step S3.7 uses a meta-learning framework for small sample transfer training: load the technical fault history case library (such as a quantum dot enterprise light conversion efficiency stagnation event) to generate an adversarial sample set, and dynamically adjust the parameters through an adaptive learning rate optimizer. The optimizer integrates a technology maturity sensor that scales the gradient update amplitude η based on community attention turning points (such as a 40% monthly increase in the number of special reports in the top journal “Nature Energy”), so that the model converges to a loss value L<0.05 at the 15th iteration.

[0204] Step S3.8 generates a growth prediction report for the next three years: outputs the patent growth curve Ŷ P = (47, 59, 74 items) for the years 2024-2026, and the associated technology intergenerational difference spatial topology map marks the risk inflection point. The map shows that the intergenerational difference coefficient δ=|0.55-0.33|=0.22 (threshold δ alarm =0.25) between the laminated technology node and the crystalline silicon technology node triggers a yellow warning to rebalance R&D resources. This application compresses the commercialization progress prediction error from 12 months in traditional models to 2.3 months through dynamic modeling driven by bifurcation parameters.

[0205] The unexpected technical effects are manifested in three aspects: first, the technology bifurcation point identification improves the prediction accuracy of sequence P to 96.7%, while the traditional model is only 82%; second, the adversarial training mechanism maintains the prediction stability (variance σ²<0.01) during the fluctuation period in 2023Q3, avoiding the vicious cycle of "funds withdrawal→R&D interruption" commonly seen in the industry; third, the generational difference coefficient δ quantitatively shows that the crystalline silicon technology path still has a 55% weight, guiding enterprise G to reserve 30% of R&D resources to upgrade PERC technology, and unexpectedly obtaining cross-authorization income from dual technology routes patents.

[0206] It should be understood that some technical terms in the embodiments can be understood as follows: Generational Gap Spatial Topology Graph: a network model representing the technology route gap, with node weights calculated from patent quality, mass production progress, etc., and edge correlation degree ξ quantifying the knowledge transfer strength between technology paths, with three-dimensional coordinates mapping the generational risk distribution.

[0207] Adaptive Learning Rate Optimizer: a training algorithm that dynamically adjusts the parameter update step size according to the technology maturity, with the maturity sensor calculating the η scaling factor based on the academic attention fluctuation inflection point distribution, formula η t =η0׃(ΔA cite ) where ΔA cite is the authoritative journal citation growth rate.

[0208] Generational Gap Coefficient δ: a scalar measuring the technology route lead, δ=|W new -W old |, δ>0.25 triggers a red alert.

[0209] Alternatively, the scheme can be extended to solid-state battery electrolyte technology evaluation by adjusting the bifurcation node detection rules (such as the existence of oxide / sulfide patent families) and δ threshold (for example, δ alarm =0.30), when the prediction of sulfide patent share exceeds 60% in 2025, the production capacity conversion fund is activated. Those skilled in the art should understand that the adversarial sample construction mechanism (step S3.7) in the meta-learning framework effectively solves the overfitting problem in the small sample scenario.

[0210]

Embodiment 11

[0211] As shown in Figure 3 Step S3.6 selects a hybrid prediction architecture based on the characteristics of technological generational evolution: trend averaging method is used for the fluctuation part of sequence S, and the sliding window width W is dynamically configured as 18 months to cover the R&D period; for the transition part of sequence O, a multiple regression model is used, and the combination of activation functions is selected as LeakyReLU-Sigmoid hybrid form. Those skilled in the art should understand that this configuration enables the model to simultaneously capture the characteristics of gradual improvement and disruptive innovation. Step P1 dynamically adjusts the model parameters according to the bifurcation node parameters: when the proportion of sulfide R&D investment increases to 45% in 2023Q4, the sliding window width W is automatically reduced to 12 months, and the negative slope parameter a of the LeakyReLU function is adjusted from 0.01 to 0.05 to enhance the small sample learning ability. Step P2 encodes the weight topology relationship (edge weight ξ = 0.79) between "R&D investment → patent output" in the graph into a graph attention prior matrix, and injects it into the hidden layer neurons of the multiple regression model through feature space projection. When the pilot yield abnormally fluctuates in 2023Q3 (sequence Y decreases from 68% to 62%), the prior knowledge automatically corrects the patent prediction value deviation to within ±2.8%.

[0212] Step S3.7 uses a meta-learning framework to load a technology fault line historical case library (such as an interface passivation failure event of a certain sodium ion company), generates an adversarial perturbation sample set for small sample transfer training. The technology maturity sensor integrated with the adaptive learning rate optimizer dynamically scales the gradient update amplitude η according to the community attention fluctuation inflection point: when a sharp increase in academic attention is detected, η t = η0×1.25 to accelerate convergence; when a technology controversy event (such as a top scholar questioning the stability of sulfide) is detected, η t= η0x0.8 to improve stability. The training process converges to a loss value L = 0.038 at the 20th epoch, which is 15 fewer iterations than the conventional optimizer. Step S3.8 generates a 2024-2026 growth prediction report: output sulfide patent growth curve = (48, 61, 77 items), associated technology generational gap spatial topology graph annotation inflection point risk. The graph shows that the generational gap coefficient of sulfides and oxides is δ = |0.57-0.39| = 0.18 (threshold δ alarm = 0.20), triggering a blue warning to fine-tune R&D resources.

[0213] The unexpected technical effects are reflected in three aspects: first, the adversarial training mechanism keeps the sequence O stable during the R&D resource tilt period in 2023Q4 (fluctuation variance σ²<0.008), avoiding the common "prediction error→financing fluctuation" chain reaction in the industry; second, the generational gap coefficient δ reveals that the sulfide path still has a 57% weight, guiding enterprise H to retain 40% of the capacity compatible with the oxide route, unexpectedly obtaining double-route equipment versatility cost reduction benefits; third, the η adjustment driven by community attention makes the model convergence speed increase by 25%, reducing the computing power consumption by 38%. The Generational Gap Spatial Topology Graph as the core innovation tool, its nodes represent core technology indicators (such as patent quality weight W patent , mass production progress value V volume ), edge weight quantifies the coupling strength between indicators ξ, and three-dimensional coordinates map technology fault risk distribution. The δ coefficient calculation formula is δ = |W new -W old |, where the technology weight value W is dynamically corrected by an expert system (for example, the interface stability breakthrough makes W sulfide increase by 0.07).

[0214] Optionally, this scheme is adapted to fuel cell catalyst technology evaluation, by adjusting the bifurcation node rules (such as Pt-based / Pd-based patent family coexistence) and δ threshold (δ alarm = 0.22), when predicting that Pd-based patents will account for more than 55% in 2025, activate catalyst recovery technology research channel. Those skilled in the art will understand that the gradient adjustment mechanism of community attention perception (step S3.7) maintains the robustness of the model during the academic controversy period, which is an unexpected advantage unique to the combination of steps.

[0215]

Example 12

[0216] Step M3 reconstructs the output layer based on the hidden layer output feature space: when a technical route bifurcation node is detected (e.g., enterprise J simultaneously submits an epoxy resin / bio-based resin patent), the R&D substitution weight ξ sub = 0.67 in the topological relationship is projected to the hidden layer neuron, and the reorganized activation function is a hybrid form of Sigmoid-LeakyReLU. The final output is a financial prediction distribution with a time-limited label: a gross profit margin 95% confidence interval [-7.3%, -3.8%], which is 42% narrower than the interval width of the traditional model. It should be noted that the entity relationship confidence decay mechanism plays a key role here - the confidence of the shutdown event ε supply decreases by Δc = -0.05 per day, and when c < 0.6, it automatically reduces its weight contribution to the prediction.

[0217] The unexpected technical effect embodies three advantages: first, the time delay matrix τ and the event injection mechanism compress the prediction error from 14.2% to 3.8%, and the event response time is shortened to 24 hours; second, the adversarial training maintains the prediction stability (fluctuation standard deviation σ < 0.8) in the simulated customs delay scenario, avoiding abnormal stock price fluctuations caused by supply chain disturbances; third, the function reorganization mechanism of technical bifurcation perception makes the research and development investment-output ratio prediction accuracy of the bio-based resin route reach 94.7%, unexpectedly guiding enterprise J to obtain green material subsidy qualification. The Generational Gap Spatial Topology Graph, as the core analysis tool, has nodes containing raw material cost elasticity coefficient η cost , technical substitution rate γ sub , etc., and edge weights represent the dynamic coupling strength between indicators. The Dynamic Prediction Engine outputs a probability distribution with a time-limited label by integrating real-time event streams and historical coupling laws, and the key parameter decay function where k is the industry adaptive decay rate (k = 0.12 / day for the photoresist industry).

[0218] Alternatively, this scheme is adapted to pharmaceutical bulk drug supply chain evaluation by adjusting the time delay matrix dimension (e.g., adding an environmental approval delay factor) and the decay rate (k = 0.08 / day), and activating the raw material strategic reserve when the predicted supply disruption impact exceeds the threshold. Those skilled in the art will understand that the confidence decay-driven weight adjustment mechanism (step M2) maintains the prediction reliability in long-period events.

[0219]

Example 13

[0220] Step K2 inputs the waveform features into the graph neural network: loads the industry characteristic vector of the photoresist industry chain database (such as photoetching accuracy threshold, resin purity standard), and modulates the coupling resonance threshold to θ=0.65 (industry benchmark threshold is 0.50). When the main classification number of the joint patent covers C08F222 / 14 (photosensitive resin) and G03F7 / 004 (photolithography application) at the same time, the output of the hidden correlation strength value λ=0.78. Preferably, this value quantifies the technology premium coefficient of industry-university-research collaboration, which improves the accuracy by 32% compared to the traditional patent quantity statistical method.

[0221] Step K3 constructs a dynamic decay function based on the correlation strength λ and the time delay coupling law in the dynamic indicator map (such as research and development input→patent output time delay T=6 months): ƒ(λ,t)=λ·e -0.12t (t is the time variable, unit: month). When the research and development log of Enterprise K in the third quarter shows that "the refractive index of resin breaks through 1.82", the function increases the institute's cooperation weight from 0.70 to 0.83, triggering the media dynamic dimension score correction module to increase by 15 points. Those skilled in the art should understand that this three-step collaboration compresses the cooperation value evaluation error from the industry average of, for example, 40% to 23%).

[0222] The technical effects are reflected in two aspects: first, the phase difference Δφ=90 days identifies the efficient collaboration from basic research to enterprise process transformation of the institute, guiding the government to grant "industry-university-research fast track" qualification; second, the industry parameter 0.12 in the decay function makes the cooperation weight run at a high level for 180 days after the technology breakthrough, which unexpectedly promotes the second round of joint fund investment.

[0223] Those skilled in the art should understand that for some technical terms in the above embodiments, the following explanations can be made: Semantic Resonance Waveform Features, refer to time-domain quantification parameters of technical association between entities, including keyword co-occurrence frequency amplitude A (Amplitude) and event interval phase difference Δφ (Phase Difference), used to capture cooperation depth and conversion efficiency.

[0224] Implicit Association Strength λ, is a scalar (0≤λ≤1) calculated by graph neural network fusion of industry chain knowledge, representing potential value gain generated by technical collaboration, which is modulated by industry characteristic vector.

[0225] Optionally, the present scheme is adapted to biopharmaceutical CDMO cooperation evaluation, by adjusting phase difference determination rules (such as interval between preclinical research and IND filing) and attenuation coefficient (for example, k=0.08 / month), when λ>0.75, the priority review channel is activated. Those skilled in the art should understand that the combined effect of time delay coupling and dynamic attenuation (step K3) maintains the evaluation accuracy in long-cycle R&D.

[0226] Figure 4 is a block diagram of an enterprise potential evaluation prediction device 800 based on multi-source heterogeneous data according to an exemplary embodiment. For example, the device 800 can be a mobile phone, a computer, a digital broadcast terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, etc.

[0227] Referring to Figure 4 , the device 800 can include one or more of the following components: a processing component 802, a memory 804, a power supply component 806, a multimedia component 808, an audio component 810, an input / output (I / O) interface 812, a sensor component 814, and a communication component 816.

[0228] The processing component 802 controls the overall operations of the device 800, including one or more processors 820 to execute instructions. The memory 804 stores data used by the components of the device 800, including instructions for the processor(s) 820. The storage component 806 provides non-volatile storage for the device 800. The power supply component 806 provides power for the components of the device 800. The multimedia component 808 includes a screen with output interface that can include a liquid crystal display (LCD) and a touch panel that, when present, is integrated with the screen to provide a touch screen. The multimedia component 808 also includes a front-facing camera and / or a rear-facing camera to capture images of the external environment. The audio component 810 is configured to output and / or input audio signals. For example, the audio component 810 includes a microphone (MIC) that is configured to receive an external audio signal when the device 800 is in a call mode, a record mode, and / or a voice recognition mode. The received audio signal can be further processed and / or stored. In an embodiment, the audio component 810 also includes a speaker for outputting audio signals.

[0229] The I / O interface 812 provides an interface between the processing component 802 and peripheral interface module, which can include a keyboard, a click wheel, buttons, and / or the like. These buttons can include, but are not limited to, a home button, a volume button, a start button, and a lock button.

[0230] The sensor component 814 includes one or more sensors to provide information about the status of the device 800. For example, the sensor component 814 can detect the opening / closing of the device 800, relative positioning of components, such as a display and a keypad of the device 800, a change in position of the device 800 or a component of the device 800, the presence or absence of user contact with the device 800, an orientation or acceleration / deceleration of the device 800, and a temperature change of the device 800, among other examples. The sensor component 814 can include a proximity sensor configured to detect the presence of a nearby object without any physical touch. The sensor component 814 can also include a light sensor (e.g., a CMOS or CCD image sensor) configured to work in conjunction with the display to adjust the brightness and / or contrast of the display based on ambient light conditions. In some embodiments, the sensor component 814 can further include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

[0231] The communication component 816 is configured to facilitate wired or wireless communication between the device 800 and other devices. The device 800 can access a wireless network based on a communication standard, such as WiFi, 2G, or 3G, or a combination thereof. In an exemplary embodiment, the communication component 816 receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 816 also includes a Near Field Communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on Radio Frequency Identification (RFID) technology, Infrared Data Association (IrDA) technology, Ultra-WideBand (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0232] In an exemplary embodiment, the device 800 can be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, micro-controllers, microprocessors, or other electronic elements, for performing the above-described methods.

[0233] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions, such as the memory 804 including instructions, is also provided, which can be executed by the processor 820 of the device 800 to complete the above-described methods. For example, the non-transitory computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, and an optical data storage device, etc.

[0234] Multi-source heterogeneous data interface (1) A set of special communication protocols for realizing cross-system data automatic collection is used to connect data sources of different architectures such as social security payment systems, enterprise financing databases, public opinion analysis platforms, etc. The interface processes structured and unstructured data streams through protocol conversion middleware, such as converting CSV format insured records of social security systems, API-JSON data streams of financing databases, and text logs of public opinion platforms into standardized data packets. In Baiyun District enterprise evaluation, the interface processes 5000 pieces of heterogeneous data per second, and captures the operation changes of 200 newly added insured employees of an intelligent manufacturing enterprise in real time.

[0235] Structured basic data set (1) The standardized enterprise data set after cleaning and association, containing dynamic operation indicators and risk events. The data set is organized in the third normal form of a relational database, for example, storing a certain biological enterprise's R&D investment ratio sequence for 36 consecutive months (e.g., 2023 Q1: 8.2%, Q2: 7.9%), environmental protection penalty events (2023-06-05, fine of 500,000 yuan) and other fields. In the right 4 implementation, this data set is used as the input source for blockchain storage verification to verify the authenticity of administrative penalties.

[0236] Knowledge graph engine (3) Based on natural language processing entity relationship analysis system, used to identify strategic cooperation relationship from public opinion text. When the engine analyzes the text "an AI company co-builds a cloud computing laboratory with Huawei", it extracts the cooperation entity pair "an AI company-Huawei" and labels the relationship type as "technical cooperation". In the patent analysis of right 5, this engine further identifies the industry-university-research entity network in joint patent applications of "university-enterprise", such as analyzing the quantum chip invention patent jointly applied by "University of Science and Technology of China-some quantum technology company".

[0237] Smart contract-driven risk analysis module (4) A programmable rule execution unit deployed on the blockchain, used to automatically verify and assess legal risks. When accessing the judicial blockchain data stream, the module automatically executes the preset verification rules: first, verify the validity of the "2023 intellectual property dispute case" court electronic seal, then calculate the proportion of the amount involved in the lawsuit to the net assets of the enterprise (e.g., 15%), and output the legal risk exposure coefficient 0.85. This coefficient triggers the adjustment of the risk resistance ability score of right 7 in real time, reducing the score of a certain enterprise from 80 to 72.

[0238] Graph convolutional neural network-based industry-university-research collaboration analysis module (5) A cooperation tightness calculation model based on graph structure data, used to quantify the technical innovation collaboration effect. The module constructs a patent joint application network graph, with nodes representing enterprises / universities (e.g., node A: a new material company, node B: Tsinghua University), and edge weights representing the number of joint application patents (e.g., 3). Through graph convolution operation, the cooperation tightness score of a certain enterprise with the Chinese Academy of Sciences is calculated as 0.68 (full score 1.0), which is input into the right 5 R&D capability correction engine to increase the project dynamic score by 12%.

[0239] Industry-adaptive level mapping knowledge base (6) A decision support system that dynamically adjusts rating rules, with an industry differentiation threshold strategy built-in. The knowledge base obtains benchmark data from the industrial economic map every month. When it detects that the average R&D investment in the new energy industry has risen to 12%, it automatically raises the R&D threshold for Class A enterprises from 8% to 10%. In the case of Right 6, a battery enterprise was initially rated Class B due to an R&D investment of 9.5%, but after the knowledge base identified it as belonging to the new energy industry, it was re-rated Class A according to the new threshold.

[0240] Blockchain evidence storage and verification (7) Refers to a method for verifying data authenticity based on a distributed ledger, used to confirm regulatory penalty records. After the system retrieves the record of "a chemical plant's wastewater exceeding the standard penalty" publicly announced by the Ministry of Ecology and Environment, it sends a verification request to the blockchain node of the Environmental Protection Bureau to compare the hash value of the penalty document number "Yuehuanfazi

[2023] No. 101". After successful verification, it is marked as an immutable event, triggering the legal risk deduction mechanism of Right 7, and deducting 8 points from the enterprise's risk resistance ability score.

[0241] Organizational dynamic scoring rules (8) Refers to the calculation specification for quantifying the health of an enterprise's human resources, assigning scores based on employee growth data. Rule definition: If the compound growth rate of employees is ≥ 20% for 36 consecutive months, a full score of 20 points is obtained, and 2 points are deducted for every 5% decrease. The growth rates of a drone enterprise in the past three years were 18% / 22% / 19% respectively. After the system calculated the compound growth rate of 19.6%, it assigned a score of 18 points. This score is aggregated with the financing ability score in Right 8 to generate a monthly growth curve.

[0242] Digital twin risk simulation engine (9) Refers to a virtual simulation system of an enterprise's operating state, used for stress testing crisis scenarios. The engine constructs a digital twin of a retail enterprise and loads its historical sales data and supply chain network. When simulating the scenario of "logistics interruption for 30 days", the engine calculates the inventory consumption curve and the activation cost of an alternative supply chain, and outputs a survival probability value of 38%. This value is converted into a dynamic risk entropy value of 0.72 in Right 9, replacing traditional financial indicators to evaluate the risk resistance ability.

[0243] Causal reasoning engine (10) Refers to an attribution model for analyzing the risk conduction path, used to predict chain reactions. The engine constructs a supply chain knowledge graph of a new energy vehicle enterprise. When a battery safety accident occurs in this enterprise, the model traces the path of "cell supplier → battery pack manufacturer → vehicle manufacturing" in the graph and quantifies the contagion intensity values (in the range of 0.45 - 0.82) of 22 upstream suppliers. In Right 10, this value is used to dynamically correct the risk resistance scores of associated enterprises.

[0244] Technology evolution vitality perception module (11) The special analysis unit for monitoring the innovation momentum of software companies, through multi-dimensional signal capture technology inflection point. Module continues to track the open source community activities of a database company, when detecting that the monthly increase of its core module contributors reaches 230% of the industry average (benchmark value 10% vs. measured value 33%), automatically activate the technical leadership channel. In the visual sand table of right 11, the three-dimensional display of the expansion trend of the developer collaboration network supports the government to identify it as an S-level enterprise.

[0245] It should be understood that in the above embodiments, the terminal and / or network device can perform some or all of the steps in the embodiments. These steps or operations are only examples, and the embodiments of the present application can also perform other operations or variations of various operations. In addition, the various steps can be executed in different orders as presented in the various embodiments, and it is possible that not all operations in the embodiments of the present application are to be executed. Moreover, the size of the serial number of the steps does not mean the order of execution, the execution order of the processes should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

Claims

1. A dynamic quantitative evaluation method for enterprise development process, characterized in that, The method comprises the following steps: S1. Collecting operation data and compliance data of an enterprise through a multi-source heterogeneous data interface, and performing entity association and conflict resolution operations to generate a structured basic data set; S2. Extracting dynamic evaluation features of the structured basic data set through a data analysis engine; S3. Calculating a multi-dimensional score value of the enterprise development trend based on the dynamic evaluation features; S4. Mapping the multi-dimensional score value to a preset grade interval and presenting it according to historical evaluation records.

2. The method for dynamic quantified assessment of the enterprise development process according to claim 1, wherein, The data analysis engine is an industry adaptive rule base; and, The step S3 further comprises: calling the industry adaptive rule base to dynamically calculate the structured basic data set, and outputting a dynamic quantitative score set; The step S4 further comprises: fusing the dynamic quantitative score set and the historical evaluation records to generate a dynamic trend report, and mapping it to the preset grade interval for visual output.

3. The method for dynamic quantified assessment of the enterprise development process according to claim 1, wherein, The data analysis engine is a deep feature mining network; and, The step S2 further comprises: extracting multi-scale evolution features of the structured basic data set based on the deep feature mining network and generating an enterprise dynamic indicator graph; The step S3 further comprises: constructing a dynamic prediction engine based on causal inference to calculate the multi-dimensional score value of the enterprise development trend in real time based on the enterprise dynamic indicator graph; The step S4 further comprises: generating and displaying the dynamic trend report containing cause links through an interpretable knowledge presentation module.

4. The method of claim 1, wherein, The step S1 further comprises: Real-time collection of original enterprise data streams from social security payment systems, enterprise financing databases, and public opinion analysis platforms through the multi-source heterogeneous data interface; Performing distributed data cleaning operations on the original enterprise data streams to generate an intermediate cleaning data set; Performing entity association operations on the intermediate cleaning data set based on the unified social credit code to output an associated data set; Processing the contradictory information in the associated data set through a rule engine driven conflict resolution algorithm to generate the structured basic data set.

5. The method of claim 1, wherein, The step S2 further comprises: S2.1' Acquiring original public opinion text streams from the public opinion analysis platform through the multi-source heterogeneous data interface; S2.2' Constructing a knowledge graph engine based on natural language processing to analyze the original public opinion text streams and identify enterprise strategic partner entities; S2.3' Matching the identified strategic partner entities with the industry chain database through graph neural networks to quantify cooperation value weights; S2.4' Dynamically correcting media dynamic dimension scores based on the quantification results and inputting the correction results into the multi-dimensional score value calculation module.

6. The method of claim 1, wherein, The step S4 further comprises: S4.1 Accessing a dynamic benchmark data set of an industry economic knowledge graph library; S4.2 Reconstructing the boundary determination rules of the preset grade interval based on the dynamic benchmark data set to generate an industry adaptive grade mapping knowledge base; S4.3 Monitoring national-level technical standard formulation events through a real-time event capture engine, and activating a rapid grading channel when identifying that an enterprise is leading the technical standard formulation. S4.4 The industry adaptive level mapping knowledge base is cooperatively fused with the output of the rapid grading channel to generate a three-dimensional space decision map and push to the regional economic governance platform.

7. The method of claim 1, wherein, The data analysis engine is a deep feature mining network, and the step S2 further includes: S2.1 performing a time dimension slicing operation on the structured basic data set to separate out a quarterly level high-frequency fluctuation index sequence and a three-year level strategic evolution index sequence; S2.2 performing cross-scale correlation analysis through a double-path convolution attention module of the deep feature mining network to capture time lag coupling rules between the quarterly level sequence and the three-year level sequence; S2.3 constructing an enterprise dynamic index map with weight distribution based on the time lag coupling rules, wherein a node represents a core business entity and an edge weight quantifies an influence coefficient between indexes.

8. The method of claim 1, wherein, The step S3 further includes: S31 acquiring social security payment data flow through the multi-source heterogeneous data interface, calculating a 36-month employee compound growth rate, and assigning scores according to organization dynamic scoring rules; S32 calling a financing database of Qichacha to analyze financing rounds in the past three years, matching financing stage scoring rules to generate a fund dimension score; S33 real-time monitoring of public opinion platforms and administrative punishment databases, and when major lawsuits or continuous loss events are triggered, deducting the total score according to risk deduction item rules; S34 aggregating organization dynamic scores, fund dimension scores, and risk deduction results on a monthly basis to generate a dynamic growth score and correlate historical data.

9. The method of claim 1, wherein, The data analysis engine is a deep feature mining network; And, The step S2 further includes: S2.1 performing a time slicing operation on the structured basic data set to separate out a quarterly level operation index sequence, an annual level financial index sequence, and a three-year level strategic index sequence; S2.2 analyzing evolution correlation in the quarterly level operation index sequence through a convolution attention module of the deep feature mining network to identify cross-cycle index coupling rules; S2.3 fusing the cross-cycle index coupling rules and the three-year level strategic index sequence to construct an enterprise dynamic index map with weight topology relationship; The step S3 further includes: constructing a dynamic prediction engine based on causal inference to calculate a multi-dimensional score value of the enterprise dynamic index map on the enterprise development trend in real time; The step S4 further includes: generating and displaying a dynamic trend report containing cause links through an interpretable knowledge presentation module.

10. An electronic device for enterprise potential assessment prediction based on multi-source heterogeneous data, comprising: a display; one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for executing any one of the methods according to claims 1-9.

Images