Enterprise project service management system based on intelligent agent

Abstract

The invention relates to the technical field of big data analysis, in particular to an enterprise project service management system based on an intelligent agent. Business voucher data streams, project contract documents and fund flow records of enterprises are accessed to serve as private domain operation data, and logistics transportation space-time trajectories and industrial energy consumption data in the regional association data are synchronously accessed; the verification agent group calls logistics and energy consumption data in the regional association data based on the enterprise private domain operation data, performs multi-dimensional physical time-space and logic consistency matching, and generates credible interaction features including confidence coefficient weights; the analysis agent group receives credible interaction characteristics, maps a microscopic interaction relationship to a regional macroscopic dimension, and constructs a regional industry chain topology network comprising enterprise nodes and a weighted service flow edge; based on a regional industrial chain topology network, a service supply and demand coupling index between enterprise nodes is dynamically calculated, an industrial chain breakpoint hindering project circulation is identified, and a supplementary chain path plan for optimizing resource configuration is generated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data analysis, in particular to an enterprise project service management system based on an intelligent agent. BACKGROUND

[0002] As an important carrier of regional economic development, industrial parks carry the core functions of industrial agglomeration, resource optimization allocation and enterprise cultivation. With the advancement of smart park construction, various park management systems are widely used in daily operations. Traditional park management platforms usually focus on property asset management, basic government service and the aggregation of macro statistical data. These systems realize the online of park basic administrative affairs through digital means, and to a certain extent, improve the service efficiency of the park.

[0003] However, the existing technology still has significant technical bottlenecks in realizing the deep operation analysis of enterprises in the park and the fine management of the industrial chain.

[0004] The existing park management system generally faces the problem of superficial data level and information island. It mainly relies on the financial statements or tax data filled in by the registered enterprises. The park management party lacks effective technical means to reach the actual business operation bottom of the enterprise, and cannot establish a real-time mapping channel between the reported account data of the enterprise and the external associated data. This makes the park's understanding of the enterprise stay at the level of static report, and it is difficult to grasp the real production activity of the enterprise. Secondly, the existing technology lacks multi-dimensional physical verification capability for the authenticity of the registered enterprise project, which makes it difficult to effectively identify the real performance state of the enterprise project and dynamically perceive the supply and demand coupling situation inside the industrial chain.

[0005] In summary, how to solve the problem that the existing park management system only relies on the self-reported data of the enterprise for superficial statistics, lacks physical penetration verification of the authenticity of the enterprise operation combined with external logistics and energy consumption data, and thus cannot accurately identify false operation and construct regional industrial chain topology, is a technical problem that needs to be solved in the field at present.

[0006] Therefore, an enterprise project service management system based on an intelligent agent is proposed. SUMMARY

[0007] The purpose of the present application is to provide an enterprise project service management system based on an intelligent agent, which dynamically calculates the service supply and demand coupling index between enterprise nodes through a regional industrial chain topology network, identifies the industrial chain breakpoints that hinder the flow of projects, and generates a chain path planning for optimizing resource allocation.

[0008] To achieve the above purpose, the present application provides an enterprise project service management system based on an intelligent agent, comprising: The multi-source data collection module establishes a mapping channel of enterprise private domain operation data and regional associated data, accesses business voucher data flow, project contract documents and fund flow records of the enterprise as private domain operation data, and synchronously accesses space-time trajectory of logistics transportation and industrial energy consumption data in the regional associated data; the business voucher data flow includes invoice and logistics documents; The distributed intelligent agent processing module constructs a verification intelligent agent group for data verification and an analysis intelligent agent group for network modeling; the verification intelligent agent group calls logistics and energy consumption data in the regional associated data based on enterprise private domain operation data, performs multi-dimensional physical space-time and logical consistency matching, and generates trusted interaction features containing confidence weights; the analysis intelligent agent group receives the trusted interaction features, maps micro-interaction relationships to regional macro dimensions, and constructs a regional industrial chain topology network containing enterprise nodes and weighted service flow transfer edges; The chain supplement path planning module dynamically calculates service supply and demand coupling indexes between enterprise nodes based on the regional industrial chain topology network, identifies industrial chain breakpoints that hinder project flow, and generates chain supplement path planning for optimized resource allocation.

[0009] The multi-source data collection module is configured with an entity alignment and semantic standardization unit; the entity alignment and semantic standardization unit establishes a unified metadata standard for heterogeneous data from different regions; a knowledge graph reasoning algorithm is used to bind unique identifiers for the same subject scattered in the financial and tax management system, park asset management system and judicial public system, realizing cross-domain data association; a small sample learning model is used to clean and normalize non-standard address text and enterprise abbreviations, construct standardized context data index, and transmit the cleaned data stream to the distributed intelligent agent processing module through an encrypted channel.

[0010] The verification intelligent agent group contains three-level verification logic for verifying the real performance status of project services, specifically including: The first level is information consistency verification, which uses multi-modal semantic alignment technology to verify the logical closed loop of the surface of the single document, filters format abnormal records, and obtains consistency features; The second level is physical reality verification, which introduces space-time data verification of physical facts of logistics transportation for records verified in the first level, and obtains reality features; The third level is business logic verification, which introduces energy consumption and material data to verify the rationality of production capacity support for records verified in the second level, and obtains logical features; The verification intelligent agent group performs comprehensive identification according to the consistency features, reality features and logical features, and obtains trusted interaction features containing confidence weights.

[0011] The information consistency verification includes: verifying that the agent group extracts semantic key information including project object description, service amount and performance clause from unstructured project contract documents and logistics documents using optical character recognition and natural language processing technology; constructing a semantic vector space and calculating the cosine similarity between the semantic key information and structured value-added tax invoice details and bank transaction records; Based on the similarity and semantic key information, the consistency of the fund flow, the bill flow and the contract flow is checked; for the records with semantic conflict and missing key elements, a risk warning mark is automatically generated and output as the consistency feature.

[0012] The physical authenticity verification includes: verifying that the agent group analyzes the business certificate data verified by the first level, extracts structured features including timestamp, vehicle identification information, delivery location coordinates, receiving location coordinates and cargo type; Retrieving traffic logistics data corresponding to the time window and the geographic path in the regional association data, the traffic logistics data including vehicle passing records and satellite positioning trajectories; Calculating the coincidence index of the logistics path shown by the business certificate and the actual vehicle trajectory in the time dimension and the space dimension as the authenticity feature to verify the physical existence of the project material circulation.

[0013] The business logic verification includes: verifying that the agent group obtains the theoretical input-output coefficient and energy consumption benchmark interval of the target enterprise based on the industry standard process model; based on the private domain operation data, aggregating the input data and output data of the target enterprise in the preset period, calculating the conversion data between raw material procurement quantity, energy consumption quantity and product sales quantity; Performing deviation analysis on the conversion data and the theoretical input-output coefficient, including the logical mutual exclusion of raw material type and product type, and the difference between energy consumption and corresponding theoretical energy consumption lower limit of product value; determining the logical anomaly of the enterprise and outputting the logical feature.

[0014] The regional industry chain topology network is a weighted directed graph containing time sequence attributes, and the construction method includes: mapping the enterprise entities in the region to network nodes, and mapping the verified real project interaction relationship to directed edges; According to the confidence weight, each directed edge is assigned a connection strength value, and the higher the confidence of the interaction relationship, the greater the connection strength; combined with geographic information system data, the network nodes are projected to industrial park spatial coordinates; Integrating time sequence flow attributes, dynamically tracing the fund and material flow heat of the industry chain in different time windows, and presenting the regional industry agglomeration degree and evolution trend.

[0015] The complement chain path planning module includes supply and demand coupling index calculation and breakpoint detection, comprising: monitoring the flow flux of each node in the regional industry chain topological network in real time, calculating the regional supply proportion and the supply and demand growth rate matching degree based on the confidence weighted flow flux, and generating the service supply and demand coupling index between the upstream and downstream nodes; according to the service supply and demand coupling index, the key nodes of supply and demand imbalance are identified and marked as industry chain breakpoints, and the targeted complement chain path planning is generated.

[0016] Compared with the prior art, the beneficial effects of the present application are: 1, the present application uses the law of conservation of energy to mine the deep logical vulnerability of enterprise operation data, financial statements are easy to polish, but the industrial energy consumption of enterprises as an objective physical quantity is difficult to fake; the present scheme forcibly couples and analyzes the energy consumption data and the output value data, can accurately identify abnormal enterprises with high output value but low energy consumption or with tickets but no output, help park managers see through the appearance of financial statements, master the real production activity of enterprises, and thus realize accurate policy support and risk early warning.

[0017] 2, the present application converts abstract economic data into intuitive and dynamic industry chain panoramic graphs; traditional statistical methods can only provide static lists or summary charts, and cannot show the complex correlation between enterprises. The weighted directed graph constructed by the present scheme not only shows who is doing business with whom, but also shows the real degree and closeness of the business through the weight; the integrated space-time attribute enables managers to dynamically trace back the industrial evolution process and intuitively see the formation of the industrial agglomeration area, the radiation range of the core enterprise and the context of the capital flow, providing accurate digital tools.

[0018] 3, the present application realizes the leap from passive management to active service; existing systems mostly stay at the data display level and lack problem-solving ability. The present scheme can accurately find the weak links and supply and demand imbalance points in the industry chain by calculating the supply and demand coupling index, and the complement chain path planning generated based on this is not a general suggestion, but an executable scheme calculated based on accurate data; not only optimizes the resource allocation of the park, but also directly assists the business flow of enterprises, truly realizing the fine management and strong chain complement of the industry chain. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 It is a structural schematic diagram of the enterprise project service management system based on the intelligent agent of the present application; Figure 2 It is a data logic diagram of the enterprise project service management system based on the intelligent agent of the present application; Figure 3 It is an architecture schematic diagram of the intelligent agent group of the present application. DETAILED DESCRIPTION

[0020] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of the present application.

[0021] Embodiment one: The present application provides an enterprise project service management system based on an agent, the structure of the system is as shown in Figure 1 The data logic of the system is as shown in Figure 2 The system comprises: A multi-source data acquisition module, which establishes a mapping channel of enterprise private domain operation data and regional associated data, accesses business voucher data flow, project contract documents and fund flow record of the enterprise as private domain operation data, and synchronously accesses logistics transportation space-time trajectory and industrial energy consumption data in the regional associated data; the business voucher data flow comprises invoice and logistics documents; A distributed agent processing module, which constructs a verification agent group for data verification and an analysis agent group for network modeling; the verification agent group calls logistics and energy consumption data in the regional associated data based on enterprise private domain operation data, performs multi-dimensional physical space-time and logical consistency matching, and generates trusted interaction features containing confidence weight; the analysis agent group receives the trusted interaction features, maps micro interaction relationship to regional macro dimension, and constructs a regional industry chain topology network containing enterprise nodes and weighted service flow edges; A chain supplement path planning module, which dynamically calculates service supply and demand coupling indexes between enterprise nodes based on the regional industry chain topology network, identifies industry chain breakpoints that hinder project flow, and generates chain supplement path planning for optimized resource allocation.

[0022] In specific physical implementation, the verification agent group and the analysis agent group are independent service containers (such as Docker containers) deployed based on micro-service architecture; the agents communicate asynchronously through a message queue (such as Kafka or RabbitMQ); the verification agent is deployed on a data acquisition edge server, and the analysis agent is deployed on a cloud computing cluster, and the two exchange serialized feature data through an encrypted TCP / IP protocol.

[0023] The multi-source data collection module is configured with an entity alignment and semantic standardization unit; the entity alignment and semantic standardization unit establishes a unified metadata standard for heterogeneous data from different regions; a knowledge graph reasoning algorithm is used to bind a unique identifier to the same subject scattered in the financial and tax management system, the park asset management system and the judicial public system, to realize cross-domain data association; a small sample learning model is used to clean and normalize non-standard address text and enterprise abbreviations, to construct standardized context data index, and to transmit the cleaned data stream to the distributed intelligent agent processing module through an encrypted channel.

[0024] The same enterprise is named in full in the tax system, may be abbreviated in the park access control system, and may be an address plus a contact person on a logistics document. In order to accurately identify, it needs to be unified.

[0025] To build an enterprise knowledge graph, first take the unified social credit code as the core entity. For data missing the credit code, use a knowledge graph reasoning algorithm; specifically, use an entity alignment method based on graph embedding to map entities in different data sources to a low-dimensional vector space.

[0026] Calculate the similarity between entities. If the similarity of the name text of two entities exceeds a threshold, or if there is a strong correlation between attributes such as address and legal representative, i.e., there is a multi-hop path connection in the graph, it is determined that they are the same subject, and a unique internal identifier is assigned to realize cross-domain association.

[0027] Non-standard text cleaning based on small sample learning includes: using a small sample learning model based on a prototype network, the model consists of an embedding layer (such as a BERT model) and a metric layer; The training process includes: selecting a small number of labeled standard addresses and non-standard addresses from historical data; dividing the data set into a support set and a query set. The model learns the class prototype (i.e., the mean of the features) of the samples in the support set to predict the normalization result of the samples in the query set; use the prototype loss function, i.e., minimize the Euclidean distance between the query sample and its class prototype, and maximize the distance between other class prototypes; for new input non-standard logistics addresses or enterprise abbreviations, the model maps them to a vector space, finds the closest standard address prototype, and outputs the standardized context data index; the cleaned data is transmitted to the subsequent module through a national encryption algorithm encryption channel.

[0028] The application significantly improves the availability and correlation accuracy of multi-source heterogeneous data. Traditional technologies often rely on rule matching, and the recognition rate is very low when facing arbitrary written logistics documents or abbreviations, which leads to data disconnection and information island. The scheme introduces knowledge graph reasoning to solve the cross-system subject recognition problem by using the logical relationship between attributes; introduces small sample learning to effectively solve the pain point of lack of a large number of labeled samples in the industrial scene, and a high-precision text normalization model can be trained with only a small number of samples. This enables the system to piece together fragmented information scattered in the finance, tax, and judicial systems into a complete enterprise portrait, laying a solid data foundation for subsequent in-depth verification.

[0029] The verification agent group includes three-level verification logic, and the specific architecture is as shown in Figure 3 for verifying the real performance status of the project service, including: The first level is information consistency verification, which uses multi-modal semantic alignment technology to verify the logical closed loop of the surface of the single document, filters format abnormal records, and obtains consistency features; The second level is physical reality verification. For the records verified in the first level, the spatiotemporal data verification is introduced to verify the physical fact of logistics transportation, and the reality feature is obtained; The third level is the operating logic verification. For the records verified in the second level, the energy consumption and material data are introduced to verify the rationality of the production capacity support, and the logic feature is obtained; The verification agent group performs comprehensive identification according to the consistency feature, the reality feature and the logic feature, and obtains a trusted interaction feature containing a confidence weight.

[0030] The information consistency verification includes: the verification agent group uses optical character recognition and natural language processing technology to extract semantic key information from unstructured project contract documents and logistics documents, including project target description, service amount and performance clause; construct a semantic vector space, and calculate the cosine similarity between the semantic key information and the structured value-added tax invoice details and bank flow records; Based on the similarity and the semantic key information, the consistency of the fund flow, the bill flow and the contract flow is checked; for the records with semantic conflict and missing key elements, a risk warning mark is automatically generated and output as the consistency feature.

[0031] Input unstructured PDF contract, picture format logistics document; Use a deep learning OCR model (such as a CRNN architecture) to recognize text, and then use a named entity recognition model (such as a BiLSTM plus CRF architecture) in natural language processing (NLP) technology to extract key entities: Party A, Party B, target name, amount, date, and key performance clauses.

[0032] Semantic vector space construction and comparison: Vectorization: using a pre-trained language model (such as BERT), the extracted contract subject matter description and invoice detail description are converted into high-dimensional semantic vectors; Similarity calculation: calculate the cosine similarity of two vectors, the specific algorithm is: the dot product of two vectors as the numerator, the product of the modulus of two vectors as the denominator, and the cosine similarity value is obtained by dividing the two.

[0033] Logical judgment: Semantic verification: if the cosine similarity is lower than the preset threshold, it means that the contract signs A product, but the invoice opens B product, and the system determines that there is a semantic conflict; Three-flow consistency verification: compare the contract amount, invoice amount, and bank flow amount; if the numerical difference of the three exceeds the allowed calculation error, or the time sequence logic is wrong, such as the invoice date is earlier than the contract signing date and there is no special clause, it is determined to be abnormal.

[0034] For records that pass the verification, generate high consistency features; for abnormal records, generate risk warning markers.

[0035] The present application realizes the deep automatic audit of commercial document content; the prior art is limited to the comparison of structured data, and it is difficult to find fraud behaviors such as yin and yang contracts or mismatched goods. This scheme uses OCR and NLP technology to give the system reading comprehension ability, which can identify semantic conflicts in unstructured documents. By calculating the cosine similarity of semantic vectors, the system can accurately determine whether the contract description and invoice content are substantially consistent, effectively filter out false transactions that match the amount but fake the business content, greatly reduce the manual review cost of the park management party, and improve the accuracy of risk identification.

[0036] The physical authenticity verification includes: verifying the agent group to analyze the business voucher data that has passed the first level verification, extracting structured features including timestamp, vehicle identification information, delivery location coordinates, receiving location coordinates and cargo type; Retrieving traffic and logistics data on the corresponding time window and geographic path in the regional association data, the traffic and logistics data including vehicle passing records and satellite positioning trajectories; Calculate the coincidence degree index of the logistics path shown by the business voucher and the actual vehicle trajectory in the time dimension and the spatial dimension as the authenticity feature to verify the physical existence of the project material circulation.

[0037] The verification agent first analyzes the business voucher that has passed the first level verification, and extracts the structured feature four-tuple: timestamp, vehicle identification (such as license plate number), delivery location coordinates, and receiving location coordinates.

[0038] Correlation data retrieval: the system sets a tolerance window before and after the timestamp (for example, two hours before and after), and retrieves the corresponding satellite positioning trajectory data and road portal vehicle passing records of the vehicle identification in the regional correlation database.

[0039] Overlap index calculation: Trajectory matching algorithm: calculate the overlap of the actual satellite positioning trajectory and the theoretical path; the specific algorithm steps are: set the theoretical path as a buffer zone with a certain width, calculate the proportion of the number of actual trajectory points falling into the buffer zone to the total number of trajectory points; the width threshold of the buffer zone is set to 50-200 meters, which is dynamically adjusted according to the road grade, with the upper limit for expressways and the lower limit for urban roads; if the calculated proportion exceeds 85%, it is determined that the trajectories overlap; the theoretical path is obtained based on the enterprise's historical transportation data.

[0040] Space-time logic verification: calculate the average moving speed of the actual trajectory; the specific calculation is: divide the actual moving distance by the time difference. If the average speed violates the physical common sense (for example, crossing five hundred kilometers in one hour), it is determined to be abnormal; Output: generate a authenticity feature value by integrating the trajectory overlap ratio and the speed logic.

[0041] In traditional park finance or subsidy applications, enterprises often fake transactions by fabricating logistics documents. The present application penetrates the physical world through the introduction of third-party objective space-time data to verify the paper statements of enterprises; not only checking whether there are cars, but also checking whether the cars go on time and on route, making it impossible for false transactions with only financial flow and bill flow but no substantial logistics; this verification mechanism based on space-time physical facts greatly improves the authenticity of park economy data, providing strong technical support for supply chain financial risk control.

[0042] The operating logic verification includes: verifying the agent group based on the industry standard process model to obtain the theoretical input-output coefficient and energy consumption benchmark interval of the target enterprise; based on the private domain operation data, aggregating the input data and output data of the target enterprise within a preset period, calculating the conversion data between raw material procurement quantity, energy consumption quantity and product sales quantity; Perform deviation analysis on the conversion data and the theoretical input-output coefficient, including the logical mutual exclusion between raw material types and product types, and the difference between the energy consumption quantity and the corresponding theoretical energy consumption lower limit of the output value; determine the logical abnormalities of the enterprise and output the logical features.

[0043] Benchmark model acquisition: built-in or through network crawling of standard process models of various industries, obtain theoretical input-output coefficient and energy consumption benchmark interval; for example, the standard power consumption range corresponding to the production of one ton of product.

[0044] Data conversion calculation: The private domain data of the aggregated target enterprise in a preset period: total raw material procurement amount, total finished product sales amount; the regional correlation data of the period are aggregated synchronously: industrial electricity consumption, water consumption, gas consumption; Calculate the conversion data: calculate the energy consumption per unit product; the specific calculation is: construct a double-layer energy consumption model, first, use the clustering algorithm to identify the fixed benchmark energy consumption of the enterprise, which is irrelevant to the output (such as lighting, maintaining the furnace temperature, standby power consumption); second, subtract the benchmark energy consumption from the total energy consumption to obtain the effective production energy consumption; calculate the marginal energy consumption conversion rate: the effective production energy consumption divided by the finished product increment; at the same time, calculate the input-output ratio: the total raw material procurement amount divided by the total finished product sales amount.

[0045] Deviation analysis: Mutual exclusion analysis: check whether the raw material type and the finished product type have logical mutual exclusion in the process model (for example, flour is purchased but steel is sold); Energy consumption lower limit verification: compare the calculated energy consumption per unit product with the lower limit of the industry energy consumption benchmark interval. If the actual energy consumption is significantly lower than the theoretical lower limit (for example, the deviation exceeds 30%), it means that the enterprise has produced a large amount of products without consuming much energy, which is judged as a logical anomaly (may be false outsourcing or false invoices).

[0046] The present application uses the law of conservation of energy to dig deep logical loopholes in enterprise operation data. Financial statements are easy to polish, but industrial energy consumption as an objective physical quantity is difficult to fake. The present scheme forcibly couples energy consumption data and output value data for analysis, which can accurately identify abnormal enterprises with high output value but low energy consumption or with invoices but no output, helping park managers see through the appearance of financial statements and master the real production activity of enterprises, so as to realize accurate policy support and risk early warning.

[0047] Further, the operation logic verification can be further configured with a non-intrusive load fingerprint identification module; The verification intelligent agent group accesses current and voltage data by deploying edge computing gateways at the total incoming line end and key high-energy consumption branches of the enterprise, extracts current harmonic features using fast Fourier transform (FFT), and constructs a load fingerprint map; a library of industrial equipment operation characteristics is established, the load fingerprint map is compared with the characteristic library, and the start-stop state and operation cycle of the key production equipment in the enterprise are identified; the identified equipment operation cycle sequence is time-aligned and verified with the process flow (BOM route) in the project contract document; if high energy consumption but lack of corresponding key equipment characteristic waveform is monitored, or the equipment operation sequence is seriously out of order with the process flow, it is determined as a deep logical anomaly.

[0048] High-frequency acquisition and FFT transformation: Deploy high-frequency sampling modules at the back end of the household electric meter in the enterprise; perform fast Fourier transform on the collected current waveform data to obtain frequency domain signals.

[0049] Fingerprint library matching: Pre-set the load characteristic library of typical industrial equipment, use the dynamic time warping (DTW) algorithm to compare the real-time collected waveform features with the characteristic library, identify the process state, and verify the process consistency.

[0050] Verification agent checks the identified equipment sequence, and based on the logic verification of the frequency domain masking effect, the system locks the high-frequency harmonic components in a specific frequency band. Even if the current waveform appears as a sine wave in the time domain (suspected pure resistance), if the corresponding production equipment characteristic harmonic cluster is lacking in the frequency domain, the identification probability of fake production is increased, i.e., the empty burning heating furnace masks the production stop fact, and the enterprise may cheat the production value and energy consumption verification by turning on a high-power resistance load.

[0051] Further, for bus end superimposed signal equipment identification, the following decoupling algorithm is used: Constructing voltage-current trajectory images: Collect high-frequency voltage data as the horizontal coordinate and current data as the vertical coordinate, and draw voltage-current trajectory images in a time window of one power frequency period, converting one-dimensional time series into two-dimensional image features; Multi-label classification identification: Use a convolutional neural network to build a multi-label classifier. The voltage-current trajectory of different equipment has a unique geometric shape; for example, resistive load is a straight line, inductive load is an ellipse, and electronic equipment has special distortion; even if multiple equipment is running simultaneously, the superimposed trajectory image still retains the geometric features of each component.

[0052] The convolutional neural network includes an input layer, three convolutional blocks, and a fully connected output layer connected in sequence; the input layer receives binary V-I trajectory images; each convolutional block contains a 3*3 convolutional layer (step size 1), a BN layer, and a ReLU activation layer. The number of nodes in the output layer corresponds to the number of pre-set equipment types, and the output layer uses a Sigmoid activation function to output the independent existence probability of each equipment category; the loss function uses the weighted sum of binary cross-entropy loss to handle the sample imbalance problem.

[0053] State decoupling: Output the load characteristic vector distribution at the current time, use the convolutional neural network to identify the significant changes in load components, rather than absolute equipment separation, focus on identifying the proportion relationship between high-order harmonic fingerprints (for variable frequency equipment) and fundamental power (for resistive equipment) to determine whether the production line is in a full-factor operation state.

[0054] The present application utilizes non-invasive load monitoring technology to refine the verification granularity from total power consumption to device action level; the prior art can only verify how much electricity is used, and cannot verify what the electricity is used for; the present application can effectively crack the false prosperity forged by means such as idling devices and turning on irrelevant loads, and ensures that each degree of electricity corresponds to a production process link in the project contract, greatly improving the penetration and anti-fraud capability of physical verification.

[0055] The regional industry chain topological network is a weighted directed graph containing time sequence attributes, and the construction method comprises the following steps: mapping the enterprise entities in the region to network nodes, and mapping the verified real project interaction relationships to directed edges; According to the confidence weight, a connection strength value is assigned to each directed edge, and the higher the confidence of the interaction relationship, the greater the connection strength; in combination with geographic information system data, the network nodes are projected to the spatial coordinates of the industrial park; The time sequence flow attribute is integrated, the fund and material flow heat in different time windows is dynamically traced back, and the regional industry agglomeration degree and evolution trend are presented.

[0056] Node and edge mapping: each enterprise entity in the park is mapped to a node of the graph network; the real project interaction relationship that has passed the above three verifications and has a confidence exceeding a threshold is mapped to a directed edge; the direction of the edge represents the flow direction of materials or services, i.e., from the supplier to the demander.

[0057] Weighting and connection strength assignment: Weight calculation: the aforementioned trusted interaction features are used to assign weights to each edge; the specific calculation logic is: the consistency feature, the authenticity feature and the logical feature are weighted to obtain the connection strength value of the edge. The greater the connection strength, the more real and frequent the transaction chain is.

[0058] Integration of space-time attributes includes: Spatial projection: in combination with geographic information system (GIS) data, the nodes are anchored on the actual electronic map coordinates of the park; Time sequence flow backtracking: the network contains time dimension attributes, the system can dynamically trace back and present the fund flow and material flow heat in the network in different time windows on the interface; the color depth of the edge or the speed of the dynamic particle flow is changed to visualize the presentation.

[0059] Further, in order to prevent single-dimensional fraud from escaping detection, the confidence weight calculation of the trusted interaction feature adopts a multiplication veto logic, rather than a simple weighted sum.

[0060] In order to quantify the reliability of the interaction relationship and solve the problem of missing data dimensions in some business scenarios, the present system adopts a confidence calculation model containing dynamic dimension adjustment and marginal correction.

[0061] Single-dimensional score mapping: define the mapping rule of three dimensions to map the original feature value to the score between 0 and 1: Information consistency score: based on semantic vector cosine similarity; adopt piecewise linear mapping: when the similarity is very high, the score is one; when the similarity is very low, it is directly zero; the intermediate interval is linearly mapped in proportion; Physical authenticity score: based on trajectory coincidence degree index. Considering the positioning drift error, an S-shaped nonlinear function is used for mapping; this function makes the score change rate maximum when the coincidence degree is near the center threshold, so as to realize sensitive identification of physical deviation and smooth the influence of noise data; Operational logic score: based on unit energy consumption deviation rate; adopt reverse linear penalty mechanism, that is, the larger the energy consumption deviation rate, the lower the score; when the deviation rate exceeds the allowed upper limit, the dimension score is zero.

[0062] Adaptive processing of missing data: for missing dimensions (such as pure service projects without logistics data), the system introduces an effective dimension counter. If the data of a certain dimension is missing, the dimension will not participate in subsequent calculation, and the value of the effective dimension counter will decrease accordingly; if all dimensions are missing, the confidence is directly set to zero.

[0063] Adaptive geometric mean algorithm is used to calculate the basic confidence, and a time decay factor is introduced as a marginal adjustment: Extract the scores of all effective dimensions with data; multiply the scores of all effective dimensions; take the square root of the multiplication result, and the number of square roots is equal to the number of effective dimensions, to get the basic confidence; calculate the time decay coefficient according to the time length from the data generation time to the current time; the longer the time length, the smaller the coefficient; multiply the basic confidence and the time decay coefficient to get the final confidence weight.

[0064] This calculation method ensures that if the score of any effective dimension is zero, the final weight is zero, realizing strict risk control. At the same time, by dynamically adjusting the number of square roots according to the number of effective dimensions, the system can be compatible with various business scenarios with and without logistics, avoiding misjudgment.

[0065] The present application converts abstract economic data into intuitive and dynamic industry chain panoramic map; traditional statistical methods can only provide static list or summary chart, and cannot show the complex correlation between enterprises. The weighted directed graph constructed by the present application not only shows who is doing business with whom, but also shows the real degree and closeness of the business through the weight. The integrated space-time attribute enables managers to dynamically trace the industry evolution process and intuitively see the formation of industrial agglomeration area, the radiation range of core enterprises and the context of capital flow, providing accurate digital tools for park macro decision-making.

[0066] Further, the analysis agent group integrates a space-time graph convolutional neural network prediction model (ST-GCN) when constructing the regional industrial chain topology network; the prediction model takes the regional industrial chain topology network in a historical time window as input, wherein the enterprise node features include capital flow density and production capacity load, and the edge features include service interaction frequency; the space-time topology dependence of the industrial chain is captured by using the graph convolutional layer, and the time sequence evolution law of the service flow is captured by using the time convolutional layer; by model inference, a link prediction matrix of a preset time step in the future is output, which is used to predict potential industrial chain breaking risk points and emerging industrial cooperation relationships.

[0067] The present embodiment aims to upgrade from post-event analysis to pre-event warning.

[0068] Graph sequencing: the analysis agent group slices the regional industrial chain topology network into graph sequences on a daily basis.

[0069] ST-GCN model architecture: Spatial dimension: the graph convolutional layer is used to process the graph at each time step to extract the spatial dependence features between enterprises, such as the influence of core enterprises on upstream and downstream enterprises; Temporal dimension: the output of the GCN is input into the gated recurrent unit (GRU) or one-dimensional convolutional layer to capture the evolution law of the industrial chain over time (such as seasonal order fluctuations).

[0070] Link prediction: the model outputs the adjacency matrix probability distribution at the future time; if the probability of an existing edge (supply chain relationship) in the prediction matrix is lower than a threshold value, the system warns of the risk of chain breaking; if the connection probability between two unconnected nodes surges, the system suggests potential cooperation opportunities.

[0071] Specifically, when constructing the space-time graph convolutional neural network prediction model, the specific definition of the input data is as follows: Construction of node feature vector: for each enterprise node in the network, its feature vector consists of three dimensions; the first dimension is the capital flow density, which is calculated as the normalized total amount of money in and out of the enterprise in the time window; the second dimension is the production capacity load rate, which is calculated as the ratio of the actual electricity consumption of the enterprise to the theoretical full-production electricity consumption; the third dimension is the credit score, which is the mean value of the confidence weight calculated by the system in history; Construction of adjacency matrix: a weighted adjacency matrix is constructed to represent the connection relationship between nodes; if there is a transaction between two enterprise nodes, the element value at the corresponding position in the matrix is the normalized transaction frequency, i.e., the number of transactions in the time window divided by the maximum number of transactions in the entire network; if there is no transaction, the element value is zero.

[0072] Space-time convolution operation: the model first uses a graph convolution layer to aggregate the feature vectors of the nodes with those of their neighboring nodes, and the aggregation weights are determined by the adjacency matrix; then, the aggregated feature sequence is input into a time convolution layer or a gated recurrent unit to extract the trend of the features over time.

[0073] The present application breaks through the limitation of the prior art that can only present static or historical data; by introducing a space-time graph convolutional neural network, the system can not only see the current industrial chain topology, but also predict the future evolution trend; this enables park managers to identify impending industrial chain disruptions (such as the imminent exhaustion of a key supplier's capacity) or discover potential industrial clusters, thereby intervening before problems occur, and achieving a leap from passive governance to proactive predictive maintenance.

[0074] Further, the regional industrial chain topology network also uses high-order correlation modeling based on hypergraphs; The analysis agent group not only constructs binary connections between nodes, but also constructs hyperedges containing three or more nodes, which represent complex symbiotic relationships such as "a logistics service provider serving multiple manufacturing enterprises" or "a single actual controller controlling multiple associated shell companies"; a hyperedge centrality index is defined to identify the hidden control hubs behind scattered transactions; spectral clustering analysis is performed using the hypergraph Laplacian operator to mine hidden related gangs that share key production factors on the surface but have no actual transactions, thereby correcting the vulnerability assessment model of the industrial chain.

[0075] Hyperedge definition: an ordinary graph is point-to-point, and a hypergraph allows a single edge to connect N nodes.

[0076] Define hyperedges, such as three companies sharing the same logistics provider, or two enterprise subjects sharing the same IP address for data reporting.

[0077] Spectral clustering analysis: construct the association matrix H of the hypergraph, calculate the hypergraph Laplacian matrix L, perform eigenvalue decomposition on L, extract the eigenvectors for K-means clustering.

[0078] Hidden gang identification: if seemingly independent enterprises are found to appear together in logistics hyperedges, IP hyperedges, and guarantee chain hyperedges, they are determined to be a hidden related gang; the analysis agent will recalculate the systemic risk of the gang to prevent it from artificially inflating production value through mutual false transactions.

[0079] The present application breaks through the limitation of traditional binary graph networks that cannot express complex multi-element relationships. In an industrial park, many risks, such as mutual insurance circle disruption and related transaction fraud, are often transmitted through complex shared elements. Hypergraph technology can sensitively capture these non-direct transaction hidden connections, helping managers identify seemingly unrelated but actually integrated risk syndicates, significantly improving the depth of industrial chain safety monitoring.

[0080] The complementary chain path planning module includes supply and demand coupling index calculation and breakpoint detection, which comprises: monitoring the flow flux of each node in the regional industry chain topological network in real time, calculating the regional supply proportion and the matching degree of supply and demand growth rate based on the confidence weighted flow flux, and generating the service supply and demand coupling index between upstream and downstream nodes; according to the service supply and demand coupling index, the key nodes of supply and demand imbalance are identified and marked as industry chain breakpoints, and the targeted complementary chain path planning is generated.

[0081] Flow flux monitoring: The in-degree and out-degree flow of each node in the topological network is monitored in real time; the flow flux of the node is calculated, specifically the sum of the weight of all input edges and the sum of the weight of output edges of the node.

[0082] Service supply and demand coupling index calculation: For a specific industry chain, the system calculates the service supply and demand coupling index between upstream and downstream nodes. Index calculation logic: first, calculate the regional supply proportion, that is, the total capacity in the region divided by the total demand; second, calculate the growth rate matching degree, that is, the absolute value of the difference between the upstream capacity growth rate and the downstream demand growth rate; after normalization, the two indicators are combined by weighting to obtain the service supply and demand coupling index.

[0083] Breakpoint detection and complementary chain planning: Breakpoint identification: if the supply and demand coupling index of a node is lower than the preset warning line, or the supply of a certain key link completely depends on outside the region (i.e. there is no corresponding upstream node in the network), the system will mark it as an industry chain breakpoint.

[0084] Path generation: the system generates planning based on the breakpoint type.

[0085] If it is a missing type breakpoint, the system retrieves the external enterprise database and recommends to introduce complementary enterprises.

[0086] If it is a congestion type breakpoint (shortage of supply), the system generates resource allocation optimization suggestions, such as increasing the transportation capacity of a specific logistics line.

[0087] The present application realizes the leap from passive management to active service; existing systems mostly stay at the data display level and lack problem solving ability; the present scheme can sensitively find weak links and supply and demand imbalance points in the industry chain by calculating the supply and demand coupling index, and the complementary chain path planning generated based on this is not a general suggestion, but an executable scheme calculated based on accurate data; not only optimizes the resource allocation of the park, but also directly assists the business flow of enterprises, truly realizing the fine management and strong chain complementary chain of the industry chain.

[0088] The complementary chain path planning includes matching optimization based on full network data: based on the complementary chain demand list, full network retrieval is performed in the external enterprise database connected in the regional public data space; candidate enterprises with technical patent reserves and capacity scale are screened, and based on a graph simulation algorithm, the flow change and output value increment of the candidate enterprises after accessing the local industry chain topology network are simulated; according to the simulation result, the complementary chain path planning including the target enterprise portrait, the recommended park location and the expected economic benefit prediction is generated, so as to complete the resource optimization configuration of enterprise project service.

[0089] Embodiment two: The application provides an enterprise project service management system based on an intelligent agent, which comprises: A multi-source data acquisition module establishes a mapping channel for enterprise private domain operation data and regional associated data, accesses business voucher data flow, project contract documents and fund flow records of enterprises as private domain operation data, and synchronously accesses logistics transportation space-time trajectories and industrial energy consumption data in the regional associated data; the business voucher data flow comprises invoice documents and logistics documents; A distributed intelligent agent processing module constructs a verification intelligent agent group for data verification and an analysis intelligent agent group for network modeling; the verification intelligent agent group calls logistics and energy consumption data in the regional associated data based on enterprise private domain operation data, performs multi-dimensional physical space-time and logical consistency matching, generates trusted interaction features including confidence weights, and the analysis intelligent agent group receives the trusted interaction features, maps micro interaction relationships to regional macro dimensions, and constructs a regional industry chain topology network comprising enterprise nodes and weighted service flow transfer edges; A complementary chain path planning module dynamically calculates service supply and demand coupling indexes between enterprise nodes based on the regional industry chain topology network, identifies industry chain breakpoints hindering project flow transfer, and generates a complementary chain path planning for optimizing resource allocation.

[0090] In the embodiment, in order to solve the problems of cross-domain data association difficulty and heterogeneous data unification, the multi-source data acquisition module adopts specific quantitative standards and hardware configurations: Hardware sampling parameters are specified: for non-intrusive load monitoring, the verification intelligent agent group is deployed at the edge computing gateway at the key node of the enterprise, such as the total incoming line, and the sampling frequency of the current and voltage signals is set to 10kHz to 50kHz; this frequency band setting follows the Nyquist sampling theorem, ensuring that it can completely cover and analyze the current harmonic components within 50 times (i.e. below 2500Hz), thereby effectively extracting the characteristic fingerprints of frequency conversion equipment and overcoming the technical obstacle that low-frequency sampling cannot perform effective FFT transformation.

[0091] Structured definition of address data: In order to overcome the problem of unknown output format in small sample learning, the standardized context data index defined in this embodiment adopts a hierarchical Key-Value JSON structure; specific fields include: {RegionID: administrative division code, RoadName: standard road name, Number: door number accurate matching bit, POI_Coordinates: [longitude, latitude]}; when the model outputs, the non-standard text is mapped to the above unique structured object, ensuring accurate projection of subsequent geographic information systems.

[0092] Inference rules for entity alignment: For the fuzzy description of the multi-hop path connection, this embodiment sets specific graph reasoning constraints; Path depth limit: only calculate the associated path with a jump number less than or equal to 3; Strong association edge type limit: only include the three types of strong association edges "same legal representative", "same controlling shareholder" and "completely consistent registration address" to construct sub-graphs; Determination logic: if two paths meet the above conditions and the name text similarity of the end point entity is greater than 0.6 using Jaccard similarity coefficient, it is determined as the same main body. This effectively prevents false aggregation caused by too long associated paths.

[0093] The verification agent group includes three levels of verification logic for verifying the true performance status of project services, specifically including: The first level is information consistency verification, which uses multi-modal semantic alignment technology to verify the logical closed loop of the surface of the single document, filters format abnormal records, and obtains consistency features; The second level is physical reality verification, which introduces spatiotemporal data to verify the physical fact of logistics transportation for the records verified in the first level, and obtains reality features; The third level is business logic verification, which introduces energy consumption and material data to verify the rationality of production capacity support for the records verified in the second level, and obtains logic features; The verification agent group performs comprehensive identification according to the consistency features, reality features and logic features to obtain credible interaction features containing confidence weights.

[0094] The information consistency verification process includes: Multi-modal data preprocessing and OCR extraction: inputting unstructured PDF contracts and picture format logistics documents, first perform image enhancement processing, including denoising, binarization and tilt correction.

[0095] OCR engine configuration: DBNet is used as the text detection network to extract the coordinate frame of the text area; CRNN is used as the text recognition network with CTC loss function; Confidence ranking processing mechanism: the system sets the recognition confidence threshold to 0.85; High confidence (≥0.85): directly output the recognized text and coordinates; Low confidence (<0.85): trigger the semantic error correction mechanism. Use the BERT mask language model to predict and replace ambiguous characters based on the context; if the confidence after correction is still not up to standard, mark the document as "manual review" status and temporarily do not enter the subsequent automatic comparison process to prevent dirty data pollution.

[0096] Visual-semantic fusion multi-modal alignment: for the common table misplacement and key-value pair separation problems in the document, pure NLP cannot handle it, so the system uses the LayoutLM multi-modal model for entity extraction. Input layer construction: concatenate the text token embedding vector, one-dimensional position embedding vector, and two-dimensional layout embedding vector extracted by OCR.

[0097] Feature extraction and alignment: the model learns the text semantics and spatial layout relationship simultaneously through the Self-Attention mechanism, so as to accurately identify the specific values of key entities such as "Party A", "Party B", and "Total Amount", even if the table line is missing or the layout is complex, it can still achieve accurate alignment.

[0098] Semantic vector space construction and logical comparison: Vectorization: use the RoBERTa pre-training model to convert the extracted contract subject matter description and invoice detail description into 768-dimensional high-dimensional semantic vectors.

[0099] Similarity calculation and determination: calculate the cosine similarity of the two vectors.

[0100] Complete match: similarity > 0.95, determined to be consistent; Fuzzy match: 0.8 < similarity ≤ 0.95, the system further retrieves the synonym library, if in the library, it is determined to be consistent; Semantic conflict: similarity ≤ 0.8, determined to be inconsistent, and a risk warning is generated.

[0101] Numerical logic verification: compare the contract amount, invoice amount, and bank flow amount; the system allows a small calculation tail difference, if it exceeds this range, or the timestamp logic violates the standard time sequence of "contract signing -> delivery -> invoicing -> payment", it is determined to be abnormal.

[0102] Further, the business logic verification can be configured with a non-invasive load fingerprint recognition module; The verification agent group accesses current and voltage data by deploying edge computing gateways at the total incoming line end of the enterprise and key high-energy consumption branches (such as injection molding workshop branches), extracts current harmonic features using fast Fourier transform (FFT), and constructs a load fingerprint atlas; an industrial equipment operation feature library is established, the load fingerprint atlas is compared with the feature library, and the start-stop state and operation cycle of key production equipment in the enterprise are identified; the identified equipment operation cycle sequence is time-aligned and verified with the process flow (BOM route) in the project contract document; if high-energy consumption but lack of corresponding key equipment feature waveform is monitored, or the equipment operation sequence is seriously out of order with the process flow, it is determined as a deep logic anomaly.

[0103] High-frequency acquisition and FFT transformation: deploy a high-frequency sampling module at the back end of the enterprise's household electricity meter; perform fast Fourier transform on the collected current waveform data to obtain frequency domain signals.

[0104] Fingerprint library matching: prestore the load characteristics library of typical industrial equipment, compare the real-time collected waveform features with the feature library using dynamic time warping (DTW) algorithm, identify the process state, and perform process consistency verification.

[0105] Verification agent checks the identified equipment sequence, performs logic verification based on the frequency domain masking effect, and locks the high-frequency harmonic components in a specific frequency band in view of the situation that the strong fundamental signal generated by the high-power resistance load may mask the motor load signal. Even if the current waveform appears as a sine wave in the time domain (suspected pure resistance), if the corresponding production equipment characteristic harmonic cluster is lacking in the frequency domain, the identification probability of fake production is increased.

[0106] Specifically, the quantitative calculation logic of the service supply-demand coupling index is as follows: Calculation of regional supply proportion index: divide the total production capacity of a certain product in a region by the total demand of this product in the whole network, and perform maximum-minimum normalization on the result.

[0107] Calculation of supply-demand growth rate matching degree index: calculate the same period growth rate of upstream production capacity and the same period growth rate of downstream demand, take the absolute value of the difference between the two; take the inverse of the absolute value and normalize it, so that the smaller the difference is, the higher the index value is.

[0108] Weighted synthesis: set the weight coefficient of the regional supply proportion index and the weight coefficient of the supply-demand growth rate matching degree index; multiply the two indexes by their corresponding weight coefficients and sum them up to get the final service supply-demand coupling index. The higher the index is, the more balanced and close the supply-demand relationship is.

[0109] Specifically, breakpoint detection and chain planning include: executing differentiated solution logic based on the nature of the breakpoint; Automatic identification of breakpoint types: The system determines based on the sub-item indicators of the service supply-demand coupling index: Structural breakpoint identification: If the regional supply proportion of a certain industry chain link is below the extremely low threshold (e.g., close to zero), it means that this link is empty in the region, and the system will mark it as a structural breakpoint, i.e., "missing link"; Functional breakpoint identification: If the regional supply proportion is normal, but the supply-demand growth rate matching degree is severely unbalanced, and the node flow flux appears congestion, it means that the existing capacity cannot meet the rapidly growing demand, and the system will mark it as a functional breakpoint, i.e., "weak link".

[0110] Chain supplement path for structural breakpoints: For missing links, the system starts targeted investment matching in the entire network; Analyze the upstream and downstream demand characteristics of the breakpoint location, generate a chain supplement demand list containing technical keywords, required capacity size, and qualification requirements; connect the external enterprise database through the regional public data space docking, retrieve candidate enterprises with high matching degree; perform virtual node embedding simulation. Create a virtual node representing the candidate enterprise in the current topology network, and establish virtual connection edges between it and existing upstream and downstream enterprises; simulate the injection of capital flow and material flow, predict the changes in network connectivity after the activation of the virtual node, and select the candidate enterprise that can best repair the breakpoint as the recommended introduction object.

[0111] Strong chain path for functional breakpoints: For congested links, the system starts optimization and scheduling of existing resources; Capacity sharing path: Scan other enterprise nodes of the same type in the region that are in a low load state, generate capacity sharing or order diversion suggestions, and balance the load; Logistics speed-up path: If the congestion is caused by low logistics transportation efficiency, the system analyzes the congested section and generates logistics scheduling instructions to increase transportation capacity or optimize transportation routes.

[0112] After the implementation of the prediction scheme, all candidate schemes are weighted and comprehensively scored according to the improvement amplitude of the service supply-demand coupling index, and ranked from high to low according to the score.

[0113] Although embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and alterations can be made to the embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. An enterprise project service management system based on intelligent agents, characterized in that, include: The multi-source data acquisition module establishes a mapping channel between enterprise private domain operation data and regional related data. It accesses the enterprise's business voucher data stream, project contract documents, and fund transfer records as private domain operation data, and simultaneously accesses the logistics transportation spatiotemporal trajectory and industrial energy consumption data in the regional related data. The business voucher data stream includes invoices and logistics documents. The distributed intelligent agent processing module constructs a verification intelligent agent group for data verification and an analysis intelligent agent group for network modeling. The verification intelligent agent group, based on enterprise private domain operation data, calls logistics and energy consumption data from regional correlation data to perform multi-dimensional physical-temporal and logical consistency matching, generating trusted interaction features containing confidence weights. The analysis intelligent agent group receives the trusted interaction features, maps micro-interaction relationships to regional macro-dimensional dimensions, and constructs a regional industrial chain topology network containing enterprise nodes and weighted service flow edges. The supply chain path planning module, based on the regional industrial chain topology network, dynamically calculates the service supply and demand coupling index between enterprise nodes, identifies industrial chain breakpoints that hinder project flow, and generates supply chain path plans that optimize resource allocation.

2. The enterprise project service management system based on intelligent agents according to claim 1, characterized in that: The multi-source data acquisition module is equipped with an entity alignment and semantic standardization unit. This unit establishes a unified metadata standard for heterogeneous data from different regions. It utilizes a knowledge graph reasoning algorithm to bind unique identifiers to the same entity scattered across the financial and tax management system, the park asset management system, and the judicial publicity system, achieving cross-domain data association. A few-shot learning model is used to clean and normalize non-standardized address text and company abbreviations, constructing a standardized contextual data index. The cleaned data stream is then transmitted to the distributed intelligent agent processing module via an encrypted channel.

3. The enterprise project service management system based on intelligent agents according to claim 1, characterized in that: The verification intelligent agent group includes three levels of verification logic to verify the true performance status of the project services, specifically including: The first level is information consistency verification, which uses multimodal semantic alignment technology to verify the logical closed loop on the surface of the document, filter out abnormal records in the format, and obtain consistency characteristics. The second level is physical authenticity verification. For the records verified in the first level, spatiotemporal data is introduced to verify the physical facts of the logistics transportation and obtain authenticity characteristics. The third level is the operational logic verification. For the records verified in the second level, energy consumption and material data are introduced to verify the rationality of capacity support and obtain logical characteristics. The verification agent group performs comprehensive identification based on consistency features, authenticity features, and logical features to obtain credible interaction features that include confidence weights.

4. The enterprise project service management system based on intelligent agents according to claim 3, characterized in that: The information consistency verification includes: verifying that the intelligent agent group uses optical character recognition and natural language processing technology to extract semantic key information from unstructured project contract documents and logistics documents. The semantic key information includes the description of the project subject matter, service amount and performance terms; constructing a semantic vector space and calculating the cosine similarity between the semantic key information and structured VAT invoice details and bank transaction records. Based on similarity and key semantic information, the consistency of fund flow, bill flow and contract flow is verified; for records with semantic conflicts and missing key elements, risk warning markers are automatically generated and output as the consistency features.

5. The enterprise project service management system based on intelligent agents according to claim 3, characterized in that: The physical authenticity verification includes: verifying that the intelligent agent group parses the business voucher data that has been verified at the first level, and extracts structured features including timestamps, vehicle identification information, shipping location coordinates, receiving location coordinates, and cargo type; Retrieve traffic and logistics data from the region-related data, which includes vehicle passage records and satellite positioning trajectories; The overlap index between the logistics path shown on the business voucher and the actual vehicle trajectory in the time and space dimensions is calculated as the authenticity feature to verify the physical existence of the project's material flow.

6. The enterprise project service management system based on intelligent agents according to claim 3, characterized in that: The operational logic verification includes: verifying that the intelligent agent group obtains the theoretical input-output coefficient and energy consumption benchmark range of the target enterprise based on the industry standard process model; and based on private domain operation data, aggregating the target enterprise's input and sales data within a preset period to calculate the conversion data between raw material procurement volume, energy consumption and finished product sales volume. The conversion data is compared with the theoretical input-output coefficients to perform a deviation analysis, including the logical mutual exclusion between raw material types and finished product types, and the difference between energy consumption and the theoretical lower limit of energy consumption corresponding to output value; logical anomalies in the enterprise are identified, and logical characteristics are output.

7. The enterprise project service management system based on intelligent agents according to claim 1, characterized in that: The regional industrial chain topology network is a weighted directed graph containing temporal attributes. The construction method includes: mapping enterprise entities within the region to network nodes and mapping verified real project interaction relationships to directed edges. Each directed edge is assigned a connection strength value based on the confidence weight, with higher confidence values ​​indicating stronger connection strengths. Combined with geographic information system data, network nodes are projected onto the spatial coordinates of the industrial park. By integrating time-series flow attributes, the flow of funds and materials in the industrial chain can be dynamically traced back to different time windows, presenting the regional industrial agglomeration and evolution trend.

8. The enterprise project service management system based on intelligent agents according to claim 1, characterized in that: The supply chain path planning module includes supply and demand coupling index calculation and breakpoint detection, including: real-time monitoring of the flow throughput of each node in the regional industrial chain topology network, calculating the regional supply ratio and the matching degree of supply and demand growth based on confidence-weighted flow throughput, generating a service supply and demand coupling index between upstream and downstream nodes; identifying key nodes of supply and demand imbalance based on the service supply and demand coupling index, marking them as industrial chain breakpoints, and generating targeted supply chain path planning.

Images