A method and system for project whole-cycle audit supervision based on synchronous prevention

Abstract

The application discloses a kind of based on synchronous prevention's engineering project whole cycle audit supervision method and system, belong to engineering construction management technical field.Method includes: through semantic data access, federal learning data cleaning and digital twin situation construction, multiple-source heterogeneous engineering data is converted into situational information assets;Based on knowledge graph, deep reinforcement learning and three-layer attention mechanism, realize the risk identification associated with engineering key node dynamically;Intelligent evidence is obtained using neural-symbolic reasoning framework, and trusted evidence is realized using blockchain technology;Based on historical cases, generate rectification strategies and monitor the whole process, feedback the results to the knowledge base, form a closed-loop optimization.The application synchronously binds audit supervision activities with engineering nodes (early warning-process-post), intelligently classifies and automatically defines responsibilities based on design, cost, construction, supervision and other functional groups;It realizes active prevention, and improves the real-time, accuracy and traceability of audit supervision.

Description

Technical Field

[0001] This invention relates to the field of engineering construction management technology, and in particular to a method and system for full-cycle audit and supervision of engineering projects based on synchronous prevention. Background Technology

[0002] As engineering projects become increasingly larger and more complex, traditional auditing and supervision models face numerous serious challenges. Firstly, regarding information collaboration: information silos exist among participating parties, leading to difficulties in communication and cooperation, unclear responsibilities, and frequent buck-passing, severely impacting project progress and investment returns. Secondly, regarding process control: on-site management relies heavily on manual monitoring and experience-based judgment, which is inefficient and prone to loopholes, resulting in frequent issues such as false approvals and irregular operations, creating hidden engineering risks and potentially causing resource waste and project delays. Thirdly, regarding audit capabilities: audit departments often lack initiative, have an unbalanced professional structure, and use relatively outdated methods and techniques, making it difficult to effectively extract audit clues from massive amounts of engineering data. Furthermore, the professional standards of external auditing firms vary widely, and their selection and management lack standardization, posing risks of fraud and threatening the impartiality and authority of audit work. Furthermore, existing audit and supervision methods are often disconnected from the specific construction milestones of engineering projects, making it impossible to achieve a closed-loop linkage between early warning triggering, process management, and post-event handling based on engineering milestones. At the same time, for problems discovered by audits, there is a lack of a mechanism for intelligent classification and responsibility delineation according to engineering management functional groups (such as design, cost, planning and construction permitting, construction, supervision, etc.), resulting in unclear division of responsibility for problems and low efficiency in rectification and coordination.

[0003] While existing technologies have improved the digitalization of engineering management to some extent, significant shortcomings remain. For example, the embedded full-coverage audit system proposed in patent application CN119809243A, although achieving multi-source data collection, does not address dynamic rule weight adjustment based on project stages or cross-entity blockchain evidence storage closed loops. Patent application CN120634481A focuses on security penetration management but lacks deep integration with audit supervision processes and closed-loop rectification mechanisms. Patent application CN119599641A provides synchronous prevention, monitoring, and early warning but does not integrate natural language processing and computer vision technologies for real-time intelligent analysis of multi-source heterogeneous data. Therefore, there is an urgent need for an intelligent audit supervision method and system that can achieve full-cycle, real-time, traceability, and adaptability. Summary of the Invention

[0004] This invention aims to systematically solve the problems mentioned in the background technology, such as information silos, passive auditing, low credibility of evidence collection, disconnect from project nodes, unclear definition of responsibilities, and lack of system adaptability. It provides a full-cycle synchronous preventive audit supervision method and system that integrates "data fusion → risk identification → intelligent evidence collection → closed-loop rectification".

[0005] To achieve the above objectives, the present invention proposes the following technical solution: A method for full-cycle audit and supervision of engineering projects based on synchronous prevention includes the following steps: S101: Intelligent data fusion and scenario construction, including: Through semantic data access, multimodal data cleaning, and context construction, multi-source heterogeneous engineering data is transformed into contextualized information assets with semantic relevance. Specifically: The semantic data access utilizes a semantic understanding gateway with a built-in engineering domain pre-trained model based on the Transformer architecture to perform field semantic parsing on data from newly added business systems and calculate its semantic similarity to concepts in the standard engineering ontology library. Based on preset thresholds, it executes automatic mapping, semi-automatic assisted confirmation, or initiates a dedicated manual processing flow to achieve cross-system semantic-level data alignment and fusion. The multimodal data cleaning employs a distributed cleaning framework based on federated learning, distributing a lightweight anomaly detection model to edge nodes of each data source for localized cleaning. For sensor data streams, runtime anomaly detection is performed and verified collaboratively with neighboring sensors; for image data streams, a lightweight convolutional neural network is used for quality assessment and to trigger re-encoders; for text data streams, syntactic analysis and semantic understanding are combined to identify logical contradictions and format errors, thereby improving the accuracy, integrity, and reliability of the data while ensuring data privacy.

[0006] The scenario construction utilizes a digital twin engine to inject cleaned and semantically aligned data into a multi-dimensional scenario model. Specifically, this includes: binding sensor data to BIM components via device IDs or coordinates to achieve spatial scenario injection; associating data with key nodes on the WBS timeline to achieve temporal scenario injection; and using knowledge graphs to attach business rules such as contract terms and technical specifications to the data to achieve business scenario injection, thereby forming contextualized information assets that can be directly used for intelligent decision-making.

[0007] S102: Dynamic risk identification and knowledge evolution, including: Based on contextualized information assets, adaptive risk identification is achieved through the collaborative use of three technologies, which are dynamically linked to engineering nodes: The knowledge graph is dynamically constructed using a hybrid strategy that combines template matching, rule and sequence labeling, and deep learning models. This strategy involves pre-defining an ontology system at the schema layer, including risk events, causal factors, control measures, project nodes, and entities categorized according to the "eight functions" of risk management, along with dynamic extraction of relation types at the instance layer. The graph integrates multi-source extraction results through an entity alignment algorithm and dynamically updates the strength of relationships between nodes based on new risk events or rectification cases, enabling continuous knowledge evolution.

[0008] Deep reinforcement learning-based adaptive optimization is employed in a virtual training platform simulating an engineering environment to train an agent that dynamically adjusts the risk identification rule base. The agent's state space is a 128-dimensional feature vector, including static project features, dynamic environment features, historical risk features, and resource status features. Its action space is defined as the continuous adjustment of the weight of a rule in the rule system, precise fine-tuning of trigger thresholds, or switching between enabled and disabled states. A multi-objective weighted reward function and a proximal policy optimization algorithm are used to train the agent, enabling it to learn to adjust its actions based on the optimal rule output according to the real-time engineering status, thus achieving adaptive optimization of the risk identification model.

[0009] A three-layer attention mechanism enables context-aware risk focusing, processing three types of contextual information in parallel to calculate risk attention levels. The project phase attention layer uses a Long Short-Term Memory (LSTM) network to process WBS node sequences; the environmental state attention layer uses a Convolutional Neural Network (CNN) to process meteorological and geological raster data; and the business activity attention layer uses a Recurrent Neural Network (RNN) to process construction logs and resource scheduling records. The three feature vectors are dynamically fused through a Gated Recurrent Unit (GRU). The GRU calculates three sets of normalized attention weights, which are then weighted and fused and matched with risk prototypes in the knowledge graph. Finally, a softmax layer outputs a risk attention probability distribution vector, guiding the system to prioritize analysis resources for high-probability risk categories. This risk identification process is dynamically linked to key project nodes (such as WBS work packages), enabling pre-node warning triggering, process monitoring during nodes, and post-node auditing and evidence collection, forming a synchronous closed loop.

[0010] S103: Intelligent forensics and trusted evidence storage, including: For the identified risks, neural symbolic reasoning and blockchain technology are used to ensure that the evidence conclusions are interpretable and the evidence is tamper-proof.

[0011] Intelligent evidence collection employs a neural symbolic reasoning framework. First, it uses feature extraction networks such as temporal convolutional networks, residual networks, and BERT models to extract key features from multi-source data and convert them into symbolic factual assertions with confidence scores. Then, rule-based reasoning and case-based reasoning are executed in parallel on the knowledge graph. Finally, the results of the two reasoning paths are fused using DS evidence theory to calculate the comprehensive confidence interval for the risk hypothesis. If the interval exceeds a threshold, a final evidence collection conclusion is generated. The system automatically pushes the conclusion and structured evidence collection form to the workbench of the relevant responsibility group based on the issue-responsibility associations in the knowledge graph.

[0012] A trusted evidence storage system is built based on a consortium blockchain. Evidence packages are organized using an improved Merkel-Patricia Tree (MPT) structure, and root hashes are calculated. Consensus is quickly reached through a HotStuff-optimized BFT consensus algorithm, and hash data including timestamps and signatures is uploaded to the blockchain. Complete original evidence files are stored in distributed file systems such as IPFS. Simultaneously, dedicated audit smart contracts are deployed to automatically execute evidence solidification verification, responsible party address association, and process control (ensuring the rectification process cannot be shut down before evidence is uploaded to the blockchain), achieving end-to-end trusted evidence storage and automated management.

[0013] S104: Closed-loop rectification and system evolution, including: Intelligent rectification strategies are generated based on evidence findings and their implementation is monitored, with the system continuously evolving through experience feedback.

[0014] The intelligent rectification strategy generation process begins by calculating the similarity between the current problem and historical cases using weighted Euclidean distance, retrieving the K most similar cases and their corresponding strategy templates. Next, using template parameters (such as material usage and construction period) as decision variables, a mathematical model is constructed to minimize cost, construction period, and maximize success rate. This model is then solved using a Non-Dominated Ranking Genetic Algorithm with Elite Strategy (NSGA-II) to obtain a set of Pareto optimal solutions. Finally, a Gradient Boosting Decision Tree (GBDT) model is used to predict the success rate, cost, construction period, and confidence interval of each candidate strategy, and the final strategy is selected based on decision preferences. The system decomposes the strategy into task orders based on the responsibility mapping in the knowledge graph and pushes them to the application terminals of relevant responsible parties via a message middleware.

[0015] The system monitors and accumulates knowledge related to rectification implementation, providing comprehensive oversight of progress, quality, resources, and environment. Upon completion of rectification, the system automatically transforms cases into structured learning samples. The Apriori algorithm is periodically used for association rule mining (e.g., identifying effective combinations of rectification measures), and the PrefixSpan algorithm is used for sequence pattern analysis (e.g., identifying efficient workflows). This new knowledge is used to: update best practice nodes in the knowledge graph, optimize case retrieval similarity weights, fine-tune the GBDT prediction model, and incorporate complete disposal cycle data into the deep reinforcement learning training environment to optimize its policy network. This achieves closed-loop optimization and continuous evolution of the system across risk identification, evidence collection, and rectification stages.

[0016] Accordingly, this application also provides a synchronous prevention-based full-cycle audit and supervision system for engineering projects that implements the above method. The system adopts a microservice cloud-native architecture and includes: 1) Application layer, used to provide the user interface, including: The risk radar subsystem, developed based on WebGL technology, enables three-dimensional visualization of risk display. It marks the identified risks on the BIM model in real time, uses color coding to represent the risk level, and dynamically displays the risk change trend through particle effects. It also supports filtering and displaying risk distribution by engineering node view.

[0017] The audit workbench integrates an intelligent evidence collection assistant, supports remote auditing work involving multiple people, and displays pending issues and tasks in a group view based on the "eight functions" of responsibility.

[0018] The rectification cockpit utilizes data dashboard technology to achieve transparent management of the rectification process, displaying rectification performance dashboards for each functional group (eight functions), ensuring accountability to the group and effective supervision.

[0019] The mobile terminal supports offline operation mode and integrates QR code scanning, GPS positioning, and photo watermarking functions.

[0020] 2) Intelligent Engine Layer, which provides core intelligent services, including: Natural Language Processing Service, based on BERT models pre-trained in the engineering field, provides APIs for contract parsing, log analysis, and report generation; Computer vision services that integrate object detection, behavior recognition, and OCR functions; The rules engine service, based on the Drools rules management system, supports dynamically adjusting the priority and trigger threshold of rule execution through plugins; The graph computing service, developed based on the Neo4j graph database, provides interfaces for path analysis, community discovery, and similarity calculation.

[0021] 3) Data service layer, used to build a unified data access system, including: A streaming processing platform built on Apache Flink, supporting real-time data processing; A batch processing platform based on the Spark architecture, used for large-scale data analysis and model training; The unified data access service adopts a multi-tenant architecture and supports concurrent data access.

[0022] 4) Infrastructure layer, which provides the underlying operating environment, including: A containerized platform built on Kubernetes to enable automated service deployment and maintenance; A distributed storage system is designed to provide optimal storage solutions for time-series data, relational data, graph data, and unstructured data. The blockchain network, built on the Hyperledger Fabric framework, is a consortium blockchain that enables trusted evidence storage.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention achieves a shift from passive response to proactive prevention in engineering audit supervision by constructing a closed-loop management system encompassing "data fusion, risk identification, evidence collection and storage, and rectification and evolution." Based on Transformer-based semantic access and federated learning cleaning, it overcomes the bottlenecks of multi-source heterogeneous data fusion and privacy protection. Deep reinforcement learning-driven rule dynamic optimization and a three-layer attention mechanism for contextual awareness significantly improve the accuracy and adaptability of risk warnings. A neural symbolic reasoning framework ensures the interpretability of the intelligent evidence collection process, and the blockchain-based evidence storage system endows audit conclusions with tamper-proof legal validity. Intelligent strategy generation based on case reasoning and multi-objective optimization, along with a knowledge accumulation mechanism through association rules and sequence pattern mining, endows the system with continuous self-evolution capabilities. In particular, this invention innovatively dynamically binds the audit supervision process to key engineering nodes, achieving a synchronous closed loop of "early warning triggering - process management - post-event processing." Simultaneously, by introducing the "eight-function" principle for intelligent problem classification and automatic responsibility definition, it achieves precise risk accountability and efficient collaborative rectification, truly realizing synchronous prevention and intelligent supervision throughout the entire engineering project process. Attached Figure Description

[0024] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is the overall flowchart of the full-cycle audit supervision method provided in the embodiments of the present invention.

[0027] Figure 2 This is a detailed flowchart of intelligent data fusion and context construction provided in the embodiments of the present invention.

[0028] Figure 3 This is a schematic diagram of the architecture for dynamic risk identification and knowledge evolution provided in an embodiment of the present invention.

[0029] Figure 4 This is a schematic diagram of the intelligent evidence collection and trusted evidence storage process provided in the embodiments of the present invention.

[0030] Figure 5 This is a schematic diagram of the closed-loop rectification and system evolution process provided in the embodiments of the present invention.

[0031] Figure 6 This is an overall architecture block diagram of the full-cycle audit and supervision system provided in this embodiment of the invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0033] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0034] Furthermore, the use of terms such as "first" and "second" in this invention is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" and "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but only on the basis of being achievable by those skilled in the art. When the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed by this invention.

[0035] Example 1 like Figure 1 As shown in the flowchart of the full-cycle audit supervision method of the present invention, the present invention provides a full-cycle audit supervision method for engineering projects based on synchronous prevention. This method forms a complete closed loop from data foundation to continuous improvement through four closely linked technical stages: "intelligent data fusion and scenario construction (step S101)," "dynamic risk identification and knowledge evolution (step S102)," "intelligent evidence collection and credible evidence storage (step S103)," and "closed-loop rectification and system evolution (step S104)." Each stage is built upon the results of the previous stage, forming a complete full-cycle audit supervision system. Specifically, the full-cycle audit supervision method for engineering projects based on synchronous prevention includes the following steps: S101: Intelligent Data Fusion and Context Construction The core task of this phase is to transform scattered, multi-source, heterogeneous raw data into contextualized information assets with rich semantic relationships, providing a solid data foundation for subsequent risk identification and intelligent decision-making. The multi-source, heterogeneous raw data refers to engineering data from different business systems; the contextualized information assets are engineering data that, after semantic alignment, data cleaning, and context construction, possess spatial, temporal, and business contexts. This data can be directly used for risk identification, intelligent decision-making, and audit supervision, and is a high-quality, interpretable, and associative information product.

[0036] like Figure 2 As shown, the specific implementation process includes three progressively layered sub-stages: First, the system achieves intelligent access and initial alignment of multi-source heterogeneous raw data through a deep learning-based semantic understanding gateway. This gateway incorporates a pre-trained engineering domain model based on the Transformer architecture. This model, trained on hundreds of thousands of professional documents such as engineering contracts, technical specifications, and construction plans, is capable of deeply understanding engineering terminology and semantic relationships. When data from a new business system is accessed, the semantic understanding gateway executes the following processes: 1) Field semantic parsing: Extracting field names, sample data, and data dictionary descriptions from the accessed data source and inputting them into the pre-trained model. The model analyzes the field context through its attention mechanism, mapping it into a high-dimensional semantic vector. 2) Semantic similarity calculation and alignment: Calculating the similarity between this semantic vector and concept vectors in a standard engineering ontology library (which predefines standardized concepts such as contract amount, planned start date, and concrete strength grade, and their vector representations). 3) Mapping Decision and Structure Standardization: Decisions are made based on preset thresholds: If the similarity is higher than the threshold (e.g., higher than 0.85), automatic mapping is performed, establishing a reference to a standard concept for the field within the system, completing preliminary semantic alignment; if the similarity is in the middle range (e.g., between 0.6 and 0.85), a semi-automatic auxiliary confirmation interface is activated, providing candidate mapping suggestions to the data administrator. The administrator's confirmation result will serve as feedback signals to fine-tune the pre-trained model in real time and update the mapping rules for the field; if the similarity is lower than the lower limit (e.g., 0.6), the field is marked as semantically abnormal, a special processing flow is initiated, its data structure is temporarily stored in an isolated buffer, and a manual processing work order is generated for domain experts to define the mapping. The expert's definition will be used to update the ontology library and training model simultaneously. Through the above process, raw data from different systems are given a unified semantic understanding framework during the access phase, laying the foundation for subsequent deep cleaning and fusion.

[0037] Secondly, the system employs a federated learning-based distributed cleaning framework to perform deep cleaning on the semantically aligned data already accessed in the preceding stages. The innovation of this framework lies in the fact that cleaning logic (such as anomaly detection models) is distributed to the edge computing nodes corresponding to each data source. Data cleaning is performed locally after the data leaves its original system and before entering the central fusion layer, avoiding centralized exposure of the original data and ensuring data privacy. Each edge node runs a lightweight anomaly detection model, which is collaboratively trained through federated learning while protecting data privacy. For example, for accessed sensor data streams (such as vibration data), the local cleaning module runs a sequential anomaly detection algorithm to identify anomalies caused by noise or equipment malfunctions, and can perform collaborative verification with data from neighboring sensors, uploading only the verified valid data. For accessed image data streams, the local cleaning module uses a lightweight convolutional neural network for quality assessment (such as blur and occlusion recognition), automatically triggering retake requests or marking low-confidence images that do not meet quality standards. For accessed text data streams (such as construction logs), the local cleaning module combines syntactic analysis and preliminary semantic understanding to identify obvious logical contradictions, missing key information, or format errors. After this round of cleaning, the data has been further improved in terms of accuracy, completeness, and reliability, while maintaining semantic consistency.

[0038] Finally, the system uses a digital twin engine to inject the cleaned, semantically aligned, high-quality data from the preceding stages into a multi-dimensional context model for context construction. The specific construction process is as follows: 1) Spatial context injection: The system reads the BIM model and parses it into a component-level spatial object library. When cleaned sensor data (such as displacement monitoring data) flows in, it performs rapid retrieval and matching in the spatial object library using the attached device ID or spatial coordinate information, binding the data to specific BIM components (such as foundation pit support piles - number A12), achieving precise spatial mapping between data and physical entities. 2) Temporal context injection: The system parses the project work breakdown structure (WBS) and schedule, generating a sequence of key nodes with a timeline as the axis (e.g., 2025-12-11: foundation pit excavation completed). The system automatically adds a processing timestamp to each data point and associates it with the WBS timeline. For example, concrete strength test data collected and cleaned on 2025-12-10 is associated with the second-floor slab pouring node under the main structure construction stage, giving the data a clear time stage affiliation. 3) Business Context Injection: The system utilizes the knowledge graph under construction to associate business rules with the data. For example, when the cleaned material arrival and acceptance form data flows in, the system automatically queries the knowledge graph for the corresponding contract terms and technical specifications based on the aligned semantic information (material type, quantity), and attaches these business rules as metadata to the data. Through the pipeline processing of semantic access → data cleaning → context injection, the original multi-source heterogeneous data is gradually transformed into contextualized information assets rich in accurate semantics, high quality, and closely related to spatial, temporal, and business contexts.

[0039] S102: Dynamic Risk Identification and Knowledge Evolution Building upon the contextualized information assets established in the first phase, this phase focuses on constructing a dynamic risk identification and knowledge evolution system with continuous evolution capabilities. For example... Figure 3 As shown, this stage utilizes three technologies in synergy: dynamic knowledge graph construction, adaptive tuning through deep reinforcement learning, and three-layer attention contextual focusing, to achieve risk identification dynamically associated with engineering nodes. The specific implementation process is as follows: The knowledge graph is dynamically constructed using a hybrid strategy of predefined schema layers and dynamic extraction at the instance layer. The schema layer predefines the ontology system and relationship types (such as triggering, controlling, responsibility, temporal, and spatial inclusion) including risk events, causal factors, control measures, project milestones, and entities categorized according to the "eight functions" principle (e.g., design, cost estimation, planning and approval). The instance layer is constructed through multi-strategy information extraction. For structured business data (e.g., contracts, schedules), pre-defined extraction templates are used to automatically extract entities and relationships. For semi-structured log reports, a rule engine and sequence labeling model are combined to identify entities and relationships. For unstructured text and images, a BERT-based named entity recognition model and a Faster R-CNN-based visual entity detection model are used for extraction, respectively. All extraction results are disambiguated and fused using a graph neural network-based entity alignment algorithm to form a unified knowledge graph. The knowledge graph is not static but continuously evolves with project progress: when the system identifies a new risk or completes a rectification, the graph is automatically updated. During updates, the system extracts new entities and relationships from relevant reports, matches existing nodes or creates new nodes using entity linking technology, and recalculates the association strength between related nodes based on the event context using a graph attention network, dynamically updating edge weights. For example, if a strong correlation is repeatedly found between rainfall and foundation pit displacement under specific geological conditions, the weight of the aggravating relationship between the two will be automatically increased. This mechanism enables the knowledge graph to accumulate project experience and achieve dynamic evolution.

[0040] Deep reinforcement learning adaptive tuning overcomes the problem of rigid static rules. The system constructs a virtual training environment based on deep reinforcement learning, the core of which is training an agent to learn how to adjust risk identification rule parameters under different engineering states. The specific mechanism is as follows: 1) State space representation: the state at each simulation time step. It is a 128-dimensional feature vector, which is composed of four parts: static features of the project (such as type, scale, complexity, etc.), dynamic environmental features (such as meteorology, geology, surrounding environment, etc.), historical risk features (such as type distribution, frequency, severity, etc.), and resource status features (such as personnel, equipment, materials, etc.). 2) Action space definition: the actions of the intelligent agent. The adjustment operations for a rule in the risk identification rule base are defined as follows: ① Continuous adjustment of rule weight: adjusting within the range of [0,1] to affect the importance of the rule in the comprehensive judgment; ② Fine-tuning of rule trigger threshold: such as adjusting the concrete temperature alarm threshold from 30°C to 28°C; ③ Switching rule activation status: enabling or disabling a rule based on context. 3) Reward function design: A multi-objective weighted reward is adopted, considering four dimensions: the accuracy of risk identification, the timeliness of early warning, the control level of false alarms, and the system resource consumption. The Pareto optimal solution is found through weighted summation. Among them, Accuracy is the identification accuracy rate (based on matching real risks in a simulated environment); Timeliness is the early warning time. 1) False positive rate; 2) Instability is the frequency of rule adjustments (penalty for unnecessary changes); 3) Weights α, β, γ, δ are determined through hyperparameter tuning. 4) Training and Decision-Making: The agent is trained using the Proximal Policy Optimization (PPO) algorithm. Over millions of simulated engineering cycles, the agent learns the optimal policy through exploration and utilization. In practical applications, the system collects project status data in real time. The trained policy network π outputs actions. It automatically fine-tunes the rule base to ensure that the risk identification model can adapt to changes in project phases, environment, and resources.

[0041] A three-layer attention mechanism enables context-aware risk focus. This mechanism processes three types of contextual information in parallel and calculates the overall risk focus. 1) Information Encoding: The project-stage attention layer uses a Long Short-Term Memory (LSTM) network to process the current and historical WBS node sequences, outputting vectors representing stage characteristics and trends. The environmental state attention layer uses a convolutional neural network (CNN) to process spatial data such as meteorological grids and ground-penetrating radar maps, and outputs an environmental feature vector. The business activity attention layer uses a recurrent neural network (RNN) to process construction flow logs and resource scheduling records, outputting business activity feature vectors. 2) Attention Fusion: Three Feature Vectors The data is fed into a gated recurrent unit. Inside the gated recurrent unit, three sets of normalized attention scores are calculated using learnable gating weights. These represent the weights of the project stage, environmental conditions, and business activities at the current moment on the overall risk profile. For example, during the foundation pit excavation stage, The weight may increase; during heavy rain, 3) Risk Attention Output: The weighted and fused comprehensive context vector is matched with the predefined risk category prototype vector in the knowledge graph for similarity, and a probability distribution vector is output through a softmax layer. ,in This represents the probability that the i-th type of risk needs to be given priority attention under the current overall situation. This distribution directly guides the system to allocate monitoring resources and analytical computing power towards high-probability risk categories, achieving precise focus.

[0042] S103: Intelligent Evidence Collection and Trusted Evidence Storage Regarding the risk alerts output by S102, this stage uses alert details and related multi-source contextualized data as input. It achieves interpretable intelligent evidence collection through neural symbolic reasoning and establishes tamper-proof and trustworthy evidence storage based on blockchain technology, outputting structured audit evidence collection forms and on-chain evidence storage records. For example... Figure 4 As shown, this stage ensures the accuracy and legal validity of audit evidence through two parallel technical processes.

[0043] The intelligent evidence collection process employs a neural symbolic reasoning framework, which integrates the perceptual capabilities of deep learning with the interpretability of symbolic reasoning. The specific steps are as follows: 1) Multi-source feature extraction and symbolic assertion generation: The system extracts key features from data associated with risk alerts. For example, regarding the risk of temperature cracks in large-volume concrete, anomaly features of temperature rise rate are extracted from time-series temperature sensor data using a temporal convolutional network (TCN); anomaly features of surface temperature gradient are extracted from infrared thermal imaging images using ResNet; and features of maintenance interruption records are extracted from construction log text using a BERT model. Each feature is converted into a feature vector and accompanied by a confidence score CF (0≤CF≤1). A pre-trained symbolic encoder maps these vectors to logical assertions, such as: Assert1: Temperature rise rate exceeds limit (CF=0.93); Assert2: Surface temperature gradient exceeds limit (CF=0.88); Assert3: Maintenance record missing (CF=0.75). 2) Dual-path parallel reasoning based on knowledge graph: The system simultaneously performs rule-based and case-based reasoning on the knowledge graph. Rule-based reasoning retrieves matching rule chains (e.g., "IF temperature rise rate exceeds limit AND insufficient maintenance THEN high risk of temperature cracks"), substitutes the assertions into the calculation, and obtains the reasoning result R1 and its confidence level. Case-based reasoning retrieves similar historical cases from the case library (e.g., similar temperature rise and maintenance, ultimately resulting in cracks), and obtains the treatment conclusions of similar cases as the reasoning result R2 and its similarity weight. 3) DS evidence theory fusion and conclusion generation: R1 and R2 are treated as two independent sources of evidence and fused using DS evidence theory. First, the basic probability assignment for each assertion is calculated based on its confidence level CF and the confidence level of the rule / case. Then, the DS combination rule is used to perform an orthogonal sum operation on the basic probability assignments of the two sources of evidence to obtain the comprehensive confidence interval for hypothesis H (high risk of temperature cracks). .like If the preset threshold is exceeded, a final evidence conclusion is generated. The system automatically traces the typical responsibility chain of "concrete temperature cracking problem" in the knowledge graph and pushes the conclusion and structured audit evidence form to the workbench of the construction and supervision responsibility groups.

[0044] The blockchain-based evidence storage system constructs a hierarchical trusted evidence storage architecture. To ensure the authenticity, integrity, and non-repudiation of the evidence collection process and results, the system builds an evidence storage system based on a consortium blockchain: 1) Evidence organization and hashing on-chain: The system organizes the evidence packages (evidence conclusions, evidence collection forms, feature data summaries, etc.) generated by intelligent evidence collection using an improved Merkel-Patricia Tree (MPT) structure and calculates their root hash. Key metadata such as the root hash, timestamp, and relevant party digital signatures are packaged into transactions, and consensus is quickly reached among consortium blockchain nodes through a HotStuff-optimized BFT consensus algorithm, and then written to the blockchain. 2) Distributed storage of original files: Complete original evidence files (such as original sensor data streams and high-resolution images) are stored in a distributed file system based on IPFS, and the content identifier generated in IPFS serves as a digital fingerprint corresponding to the on-chain hash. 3) Automated smart contract management: Dedicated audit smart contracts deployed on the blockchain execute automatically: ① Evidence solidification contract: Verifies data format, signature validity, and timestamp logic; ② Liability association contract: Permanently records the association between the relevant responsible party's address and the event hash in the ledger based on the liability determination in the evidence collection conclusion; ③ Process control contract: Ensures that the downstream rectification process cannot be marked as complete before the corresponding evidence is verified and uploaded to the blockchain. 4) Query and verification: Authorized users can query the evidence storage records using transaction ID, time range, etc., through the blockchain explorer, and can use Merkel path proofs and IPFS content identifiers to verify the integrity of the original files stored offline.

[0045] S104: Closed-Loop Rectification and System Evolution This phase uses the audit evidence collection form and on-chain evidence records output by S103 as triggers, combines historical case libraries with the current project context to generate intelligent rectification strategies and monitor their execution, ultimately feeding the rectification results back to the knowledge base to achieve closed-loop optimization and continuous evolution of the system. For example... Figure 5 As shown, the specific implementation is as follows: The generation of intelligent rectification strategies is a decision-making process based on case-based reasoning and multi-objective optimization. The first step, case retrieval and context matching, involves the system retrieving similar cases from a historical case database based on the characteristics of the current risk issue (type, location, responsible party) and the project context (stage, environment). Similarity is calculated using weighted Euclidean distance. ω is an adjustable weight. The system returns the K most similar cases and their successful rectification strategy templates (e.g., strategy T: add cooling water pipes + cover with insulation layer + adjust pouring time). The second step is to optimize the strategy parameters in multiple objectives, using the adjustable parameters in the strategy template as decision variables (e.g., cooling water pipe spacing X1, insulation layer thickness X2, pouring time window X3), to construct a multi-objective optimization model: ; ; .

[0046] The system employs a non-dominated ranking genetic algorithm (NSGA-II) with an elitist strategy to solve the model: a new population is generated through selection, crossover, and mutation, and the next generation is selected based on non-dominated ranking and crowding. After iteration, a set of Pareto optimal solutions (i.e., solutions that cannot improve one objective without harming others) is obtained. The third step is effect prediction and strategy selection. For each candidate strategy in the Pareto solution set, a pre-trained gradient boosting decision tree (GBDT) model is used to predict its effect. This model takes strategy parameters and current risk situation characteristics as input and outputs the predicted success rate, expected cost, and project duration, along with their 90% confidence intervals. The system combines the decision-maker's preferred weights (e.g., cost priority or project duration priority) to select an optimal solution from the Pareto front as the final recommended strategy. The fourth step is responsibility decomposition and task assignment. Based on the problem-responsibility mapping relationship in the knowledge graph, the system automatically decomposes the rectification strategy into specific task work orders and assigns them to the main responsible group (such as the construction party), the supervision responsible group (such as the supervision party), and the collaborating party (such as the design party that needs to confirm the plan). The tasks are then pushed to the rectification dashboard and mobile field application of each responsible party in real time through a message middleware (such as RabbitMQ).

[0047] The rectification process is comprehensively monitored using the Internet of Things (IoT) and mobile devices. Progress monitoring compares the completion percentage reported by mobile devices with the plan, automatically calculating deviations and issuing warnings. Quality verification combines automated sensor data comparison (such as thermometer data) with standardized checklists on mobile devices. Resource consumption is synchronized in real time via IoT RFID tag scanning and electronic material requisition forms. Environmental monitoring continuously collects meteorological data; if adverse conditions are detected (such as impending rainfall), the system automatically sends protective alerts. All data is aggregated in real time in the rectification dashboard, forming a visualized management panel.

[0048] Knowledge accumulation and system evolution mechanisms ensure experience consolidation and capability enhancement. Each closed rectification case is automatically transformed into a structured learning sample, including: risk scenario, adopted strategy, and actual execution results (success / failure, actual cost, actual construction period). The system regularly mines and analyzes these samples: 1) Association rule mining: using the Apriori algorithm to mine strong association rules between the combination of measures and the successful results, such as {covering the insulation layer, adjusting the pouring time} → {high success rate}. 2) Sequence pattern analysis: using the PrefixSpan algorithm to discover efficient rectification workflow patterns. This new knowledge is used for: ① Updating the knowledge graph: adding strong association rules as new best practice nodes to the graph and strengthening the association between related nodes; ② Optimizing the retrieval and prediction model: adjusting the similarity weight ω of case retrieval to prioritize cases more relevant to the successful pattern; and fine-tuning the GBDT effect prediction model with new samples to improve its prediction accuracy; ③ Strengthening DRL training: adding the complete risk handling cycle (from identification to closure) as a new training sample to the DRL virtual environment for subsequent retraining of the policy network to optimize its decision-making ability. Through this closed-loop feedback, the system achieves continuous self-evolution and performance improvement in all aspects of risk identification, evidence collection, and rectification.

[0049] Example 2 This invention also provides a full-cycle audit and supervision system for engineering projects based on synchronous prevention. Applying the method described in Embodiment 1, this system adopts a microservice cloud-native architecture design, with each layer communicating through standard interfaces to jointly support the implementation of the above method. Figure 6 As shown, the system architecture is divided into an application layer, an intelligent engine layer, a data service layer, and an infrastructure layer from top to bottom, forming a complete technology stack.

[0050] The system specifically includes: 1) Application Layer The application layer, serving as the interface between the system and users, employs a micro-frontend architecture to achieve modularization and independent deployment of functions. The Risk Radar subsystem, developed based on WebGL technology, enables 3D visualized risk display, marking identified risks in real-time on the BIM model and supporting multi-granularity viewing from the overall project to individual components. The system uses color coding to represent risk levels (red for high risk, yellow for medium risk, and green for low risk) and dynamically displays risk trends through particle effects. This subsystem can receive and intuitively display the risk attention distribution generated by the intelligent engine layer based on a three-layer attention mechanism, achieving contextualized risk awareness. The audit workbench integrates an intelligent evidence collection assistant, supporting remote auditing work involving multiple people. The evidence collection assistant provides standardized evidence collection templates, guiding auditors to complete the evidence collection process in a standardized manner and automatically generating evidence documents that meet legal requirements. The workbench can receive targeted pushes of evidence collection conclusions and audit evidence collection forms from the intelligent engine layer and supports task filtering and processing according to the "eight functions" responsibility group view. The system also provides video conferencing integration, supporting real-time communication and remote inspection between auditors and on-site personnel.

[0051] The rectification dashboard achieves transparent management of the rectification process through data dashboard technology. The dashboard design employs a multi-layered drill-down structure. The first layer displays the overall rectification KPIs of the project, including core indicators such as rectification completion rate, average rectification time, and first-time acceptance pass rate. The second layer displays a comparison of the rectification performance of each responsible unit. The third layer displays the detailed execution status of specific rectification tasks. The system uses various visualization components (bar charts, line charts, radar charts, etc.) to display data from multiple dimensions, supporting filtering and statistical analysis by time, responsible unit, risk type, and other dimensions. The mobile field terminal is optimized for the special environment of engineering sites and supports offline operation mode. When the network signal is interrupted, the system automatically caches operation data and intelligently synchronizes it after the network is restored. The mobile terminal also integrates functions such as QR code scanning, GPS positioning, and photo watermarking, and can receive and execute rectification task work orders issued by the rectification dashboard, supporting on-site data collection and task progress feedback to ensure the authenticity and traceability of on-site data.

[0052] 2) Intelligent Engine Layer The intelligent engine layer is the core technology of the system, employing containerized deployment and achieving fine-grained service governance through a service mesh. The natural language processing service, based on a BERT model pre-trained on engineering domain corpora, provides multiple dedicated API interfaces. The contract parsing API automatically identifies key clauses and liability divisions in contracts; the log analysis API extracts key events and anomalies from construction logs; and the report generation API automatically generates compliant audit reports. The computer vision service integrates multiple advanced visual analysis models. The object detection API, based on the YOLOv5 architecture and specifically optimized for engineering scenarios, accurately identifies objects such as safety equipment, construction machinery, and material stacks; the behavior recognition API uses 3D convolutional networks to analyze video sequences, identifying violations and unsafe behaviors; and the OCR API is specifically optimized for text recognition in engineering scenarios, accurately recognizing instrument readings, equipment nameplates, and material labels.

[0053] The rules engine service supports dynamic feature expansion through a plug-in architecture. The core engine is based on the Drools rules management system and utilizes a custom-developed plug-in that supports dynamic weight adjustment to achieve real-time optimization of the rules system. This plug-in can receive weight signals from the attention mechanism and dynamically adjust the priority and trigger threshold of rule execution. The graph computation service is developed based on the Neo4j graph database and provides rich graph algorithm interfaces. The path analysis API can trace the propagation path and impact range of risks; the community discovery API can automatically identify risk clustering patterns; and the similarity calculation API can quickly find similar historical cases. All intelligent services are provided externally through a unified API gateway, which implements functions such as load balancing, traffic control, and access authentication to ensure high availability and security of the services.

[0054] 3) Data Service Layer The data service layer constructs a unified data access system, adopting a hybrid architecture to accommodate both real-time processing and batch analysis needs. The stream processing platform, built on Apache Flink, provides complete stream data processing capabilities. The platform supports multiple data source accesses, including Kafka, MQTT, and WebSocket, and can process tens of thousands of real-time data streams per second. The platform offers a rich set of stream processing operators, supporting advanced features such as window aggregation, multi-stream association, and complex event processing. The batch processing platform, based on the Spark architecture, is used for large-scale data analysis and model training. The platform provides distributed data storage and computing capabilities, supporting various computing modes such as SQL queries, machine learning, and graph computing. The platform implements resource management through YARN, enabling dynamic allocation of computing resources based on task requirements.

[0055] The unified data access service adopts a multi-tenant architecture, supporting concurrent data access needs. The service provides a standardized RESTful API, supporting complex queries and transaction operations. The query optimizer selects the optimal execution plan based on a cost model, supporting index recommendation and query rewriting. The data governance service establishes a full lifecycle data management system, including data quality management, metadata management, and data lineage analysis. The service automatically detects data quality issues through a data quality rule engine and tracks the history and impact of data changes through data lineage analysis.

[0056] 4) Infrastructure layer The infrastructure layer provides a stable operating environment, employing cloud-native technology stacks to ensure system elasticity and reliability. The containerized platform, built on Kubernetes, enables automated service deployment and maintenance. The platform offers complete application lifecycle management capabilities, including automatic scaling, rolling upgrades, and fault self-healing. The platform achieves fine-grained traffic management through a service mesh, supporting advanced features such as canary deployments, fault injection, and circuit breaker degradation. The distributed storage system is designed with optimal storage solutions for different data types. Time-series data uses an InfluxDB cluster for storage, ensuring high-concurrency read / write performance through sharding and replication mechanisms. Relational data uses a MySQL cluster for storage, improving concurrency processing capabilities through read / write separation and connection pool optimization. Graph data uses a Neo4j cluster for storage, optimizing complex query performance through graph partitioning algorithms. Unstructured data uses MinIO object storage, ensuring data reliability through erasure coding technology.

[0057] The blockchain network is built using the Hyperledger Fabric framework to construct a consortium blockchain. The network features a multi-channel architecture, with different business scenarios using independent channels to ensure data isolation and performance optimization. Each participant (construction unit, construction unit, supervision unit, design unit, cost consulting unit, etc.) runs a peer node to jointly maintain the distributed ledger. The sorting service uses a Kafka cluster to ensure the orderliness and reliability of transactions. Smart contracts are developed using Go to implement complete notarization business logic. The monitoring system is built on Prometheus and Grafana to collect operational metrics at various levels. The system uses machine learning algorithms to provide intelligent early warnings, enabling early detection of potential performance issues and failure risks. The self-healing system automatically handles common faults through pre-set repair plans, ensuring the continuous and stable operation of the system.

[0058] Example 3 This embodiment uses a municipal water system comprehensive management project as a specific application scenario to explain in detail the specific implementation process and effects of Embodiments 1 and 2 of the present invention in practice. The project includes multiple sub-projects such as river dredging, sewage interception pipeline laying, and ecological bank protection construction. Through the implementation of this system, intelligent auditing and supervision of the entire process and all aspects are achieved.

[0059] 1) Application scenarios for construction machinery supervision During the river dredging construction phase, the system achieved intelligent monitoring of construction machinery throughout the entire process through the integration of multiple technologies. First, the system uses OCR technology to automatically process the machinery approval forms submitted by the construction unit. A specially trained OCR model accurately identifies key information such as machinery model, equipment number, and inspection date, and applies an attention mechanism to focus on key fields to ensure accuracy. After recognition, the system automatically compares the information with the machinery management database and establishes a digital ledger through an entity alignment algorithm to ensure the compliance of machinery approvals.

[0060] At the construction site, the system, based on a federated learning framework, monitors the machinery's operating status in real time through IoT sensors and high-definition smart cameras. Sensors, including positioning modules and vibration sensors, continuously collect data and transmit it to the data processing center. Simultaneously, computer vision technology, based on deep learning models, identifies and tracks the machinery's type, operating status, and location information, enabling comprehensive monitoring. This multi-technology integrated monitoring approach ensures the real-time nature and accuracy of the machinery's operating data.

[0061] The system also features automatic early warning for abnormal behavior, such as detecting situations where machinery with a future approval date is detected on a construction day. Through a multi-source verification process involving retrieving trajectory data and analyzing video surveillance and sensor data, the system automatically generates an "unapproved machinery in use" warning and solidifies the evidence via blockchain. Warning information is pushed to the responsible person's mobile terminal in real time, and a rectification notice and accountability traceability procedure are automatically generated to ensure timely handling of violations.

[0062] 2) Application scenarios of intelligent monitoring of construction quality In the construction of sewage interception pipelines, the system achieves precise quality control through data fusion analysis. The system first intelligently analyzes the construction drawings, using improved computer vision algorithms to automatically identify key parameters such as pipe specifications, laying slope, and backfill requirements, and converts the design requirements into structured data stored in a quality standard database. Simultaneously, the system establishes a digital model of the pipeline laying, linking technical requirements with three-dimensional spatial location, providing a digital benchmark for quality monitoring. Real-time quality data acquisition is achieved through the deployment of various intelligent monitoring devices, including thickness gauges, compaction gauges, and high-precision sensors, continuously collecting data on paving thickness, compaction degree, and slope. This data is aggregated through IoT gateways to edge computing nodes for cleaning and standardization, and then uploaded to a cloud analysis platform to ensure data reliability and real-time performance. When the system detects quality deviations, such as a laying thickness lower than design requirements, it initiates multi-dimensional verification, including analysis of compaction data, aerial imagery, and construction records. Based on the comprehensive analysis results, the system automatically generates a quality rectification order, which is pushed to the responsible party and the supervision unit for verification via smart contracts, triggering control measures. The rectification process is tracked in real time through video and sensor data until the problem is resolved. Once the problem is solved, the system automatically lifts the control measures, ensuring a closed-loop process for quality issues.

[0063] 3) Application scenarios for verifying the authenticity of personnel performance During the ecological revetment construction phase, the system uses digital identity technology to ensure the authenticity of key personnel's performance. The system establishes a digital identity for each participant based on a digital certificate system, including information such as digital certificates, biometrics, professional qualifications, and job permissions. Personnel log in through terminals integrating biometrics and multi-factor authentication. Each authentication generates a digital signature, which is then packaged and stored with timestamps and geographic location information, forming a reliable identity verification mechanism.

[0064] For the signing of critical documents, such as the submission of construction organization designs and the approval of special plans, the system requires the responsible persons to use digital signatures. The signing process is standardized, including document drafting, automatic identification of the signing order, sequential authentication of signatures by each responsible person, and storage of the entire evidence chain on the blockchain to generate electronic documents with trusted timestamps, ensuring the legality and immutability of the documents.

[0065] The system can also prevent fraudulent signatures. For example, during the approval process of special projects, it uses technologies such as biometric comparison, behavioral feature analysis, device fingerprint verification, and geolocation verification to identify anomalies. Once a risk is detected, the system automatically refuses to sign, generates a security alert, freezes the business process, requires identity verification, and records the entire event on the blockchain, effectively ensuring the authenticity of project documents.

[0066] 4) Implementation effect analysis Through practical application, this system has achieved significant improvements in management efficiency in municipal water system projects. For example, the time required to detect violations of machinery usage, the timeliness of identifying quality issues, and the time spent on document approval processes have all been significantly reduced. These improvements have reduced human intervention and increased the level of automation.

[0067] In terms of risk control, the system effectively prevented incidents of unauthorized use of machinery, promptly addressed quality deviations, and successfully prevented violations such as forged signatures. Through the automatic execution of control measures via smart contracts, early detection and resolution of risks were achieved, reducing project risks.

[0068] In terms of overall benefits, the system reduces rework losses caused by quality issues, avoids safety problems that may be caused by violations of regulations, saves manpower costs for on-site supervision, and improves the standardization and transparency of project management.

[0069] By deeply integrating audit and oversight activities with WBS nodes (such as "completion of foundation pit excavation" and "pipeline laying acceptance"), the system achieves automated linkage between risk warning, process tracking, and post-event verification, reducing the average problem discovery and response time by approximately 60%. Simultaneously, the intelligent problem classification and responsibility assignment function based on the "eight functions" principle enables over 95% of audit findings to be accurately attributed to specific functional groups such as design, construction, and supervision. The system automatically assigns rectification tasks, significantly reducing blame-shifting and communication costs, and improving the efficiency of cross-departmental collaborative rectification.

[0070] This application example fully demonstrates the effectiveness and practicality of the intelligent audit and supervision system constructed by this invention through multi-technology integration and closed-loop management in engineering practice, providing strong support for the smooth implementation of engineering projects. This method and system have broad application value and are also applicable to other types of engineering projects.

[0071] Example 4 The present invention also provides an electronic device, including: a processor, a transmitting device, an input device, an output device, and a memory. The processor may be implemented using a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit, or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory may be implemented using a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM), and is used to store computer program code. The computer program code includes computer instructions. When the processor executes the computer instructions, the electronic device executes a method as described in any of the above possible implementation methods.

[0072] Example 5 The present invention also provides a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor of an electronic device, cause the processor to perform a method as described in any of the above possible implementations.

[0073] In the description of this specification, the references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0074] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for full-cycle auditing and supervision of engineering projects based on synchronous prevention, characterized in that, Includes the following steps: S101, Intelligent Data Fusion and Context Construction, including: intelligent access and semantic alignment of multi-source heterogeneous raw data through a deep learning semantic understanding gateway; deep cleaning of semantically aligned data using a federated learning-based distributed cleaning framework; and injection of cleaned, semantically aligned high-quality data into spatial, temporal, and business multi-dimensional context models through a digital twin engine to output contextualized information assets. S102, Dynamic Risk Identification and Knowledge Evolution, including: based on the contextualized information assets, dynamically constructing and evolving a knowledge graph, and adaptively optimizing risk identification rules through deep reinforcement learning; and performing context-aware risk focusing based on a three-layer attention mechanism of project stage, environmental state, and business activities to achieve dynamic risk identification; the entity types of the knowledge graph include responsibility group entities divided according to engineering management functions; the dynamic risk identification is dynamically associated with key engineering nodes, triggering early warnings before nodes, monitoring processes during nodes, and initiating audit and evidence collection after nodes, forming a synchronous closed loop; S103, Intelligent Evidence Collection and Trusted Evidence Preservation, including: for identified risks, multi-source feature extraction and symbolic assertion generation are performed through a neural symbolic reasoning framework, dual-path parallel reasoning based on knowledge graphs and DS evidence theory fusion to achieve intelligent evidence collection and analysis, and evidence collection conclusions are pushed to relevant responsible groups based on the associations in the knowledge graph, and evidence is organized, hashed on the chain, distributed stored and managed by smart contracts using blockchain technology to implement trusted evidence preservation; S104. Closed-loop rectification and system evolution: Based on historical case database retrieval and context matching, multi-objective optimization and adjustment of strategy parameters, effect prediction and selection, intelligent rectification strategies are generated. The generation and push of the strategies are associated with the responsibility group to which the problem is classified. The rectification execution process is monitored in all aspects, and the rectification results and experience are fed back to the knowledge base to realize the closed-loop optimization and continuous evolution of the system.

2. The method according to claim 1, characterized in that, The responsible entities include at least one of the following: design, cost estimation, planning and approval, construction, supervision, procurement, operation and maintenance, and internal management of the construction unit.

3. The method according to claim 1, characterized in that, The semantic data access in step S101 specifically includes: Using a pre-trained model based on the Transformer architecture in the engineering domain, the semantic similarity between data source fields and concepts in standard ontology libraries is calculated. Based on the comparison results between the semantic similarity and the preset threshold, automatic mapping, semi-automatic assisted confirmation, or special processing procedures are executed to achieve semantic-level data fusion across systems.

4. The method according to claim 1, characterized in that, The dynamic risk identification in step S102 specifically includes: The deep reinforcement learning operates in a virtual training platform simulating an engineering environment. Its state space feature vectors include project static features, dynamic environment features, historical risk features, and resource status features. Its action space is defined as the adjustment operations of rule weights, trigger thresholds, and activation status in the risk identification rule system. The agent is trained to find the optimal strategy through a multi-objective reward function that includes accuracy, timeliness, false alarm rate, and adjustment frequency. The three-layer attention mechanism processes project phase, environmental state, and business activity information through a long short-term memory network, a convolutional neural network, and a recurrent neural network, respectively, and performs dynamic fusion and weight calculation through a gated recurrent unit to output the risk attention distribution.

5. The method according to claim 1, characterized in that, The intelligent forensics process in step S103 specifically includes: Using a neural symbolic reasoning framework, features are first extracted from multi-source data through a feature extraction network and symbolic factual assertions with confidence scores are generated. Subsequently, rule-based reasoning and case-based reasoning are performed in parallel on the knowledge graph, and the reasoning results are fused using DS evidence theory to generate the final risk assessment conclusion and audit evidence report.

6. The method according to claim 1, characterized in that, The trusted evidence storage in step S103 specifically includes: Audit evidence is organized using an improved Merkle tree structure, and fast consensus is achieved through an improved consensus algorithm based on HotStuff. Evidence is solidified, liability is determined, and process control is achieved through a dedicated smart contract for auditing. Evidence hash values ​​are uploaded to the blockchain in real time using an asynchronous parallel approach, while the original evidence files are stored in a distributed file system.

7. A full-cycle audit and supervision system for engineering projects based on synchronous prevention, wherein the system applies the method described in any one of claims 1 to 6, characterized in that, The system adopts a microservice cloud-native architecture, including: The application layer provides a user interface, including a risk radar subsystem for 3D visualization of risk, an audit workbench for supporting remote collaborative auditing, a rectification dashboard for transparent rectification process management, and a mobile field terminal for on-site data collection. The intelligent engine layer provides core intelligent services, including natural language processing services, computer vision services, rule engine services, and graph computing services. The data service layer is used to build a unified data access system, including a streaming processing platform for real-time data processing and a batch processing platform for batch analysis. The infrastructure layer provides the underlying operating environment, including containerized platforms, distributed storage systems, and blockchain networks.

8. The system according to claim 7, characterized in that, The risk radar subsystem in the application layer is developed based on WebGL technology. It marks the identified risks on the BIM model in real time and dynamically displays the risk level and trend through color coding and particle effects. The audit workbench and rectification dashboard support information filtering and display by key project node view and by responsibility group view. The mobile site terminal supports offline operation and integrates QR code scanning, GPS positioning and photo watermarking functions to ensure the authenticity and traceability of site data. The rule engine service in the intelligent engine layer supports dynamic adjustment of rule execution priority and trigger threshold through plugins. The graph calculation service provides interfaces for path analysis, community discovery and similarity calculation.

9. An electronic device comprising a processor, a memory, and a computer program stored in the memory, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1-6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-6.

Images