Enterprise project management system and method based on big data multi-modal

Abstract

The application belongs to the technical field of enterprise project management system, and particularly relates to an enterprise project management system and method based on big data multi-modal, which implements modal hierarchical decoupling type differentiated feature mining on project multi-modal heterogeneous data, adopts feature extraction strategies for three types of modal, namely structured time series data, text data and audio-visual multimedia data, independently mines corresponding project risk correlation features of each modal, executes modal internal noise reduction filtering on original features of each modal, eliminates redundant noise and invalid fluctuation features, and obtains a multi-modal pure feature set without cross interference; a cross-modal data precise alignment mechanism is established, a project unique process identifier is used as a binding primary key, multi-type heterogeneous data under a homologous project task is associated and calibrated in a subject, scene and time dimension, and unrelated heterogeneous data is eliminated through a time series correction and space matching algorithm.

Description

Technical Field

[0001] This invention belongs to the technical field of enterprise project management systems, and particularly relates to an enterprise project management system and method based on big data and multimodal approaches. Background Technology

[0002] Currently, enterprise R&D project management generally relies on the traditional model of offline ledgers and manual archiving. Throughout the entire process of project initiation, R&D log entry, experimental record retention, and R&D expense collection, enterprises generate a large amount of structured indicator data and unstructured text materials. This data is scattered and unrelated, forcing managers to rely on manual ledger analysis and past experience to assess R&D risks. This process is not only cumbersome and costly, but also makes it difficult to uncover compliance and schedule risks hidden behind multi-dimensional data. Existing intelligent project risk control solutions mostly use single-dimensional data with fixed thresholds for anomaly detection, which cannot adapt to the dynamic conditions of different stages such as R&D initiation, iterative development, and testing. They are easily affected by data interference such as manual input errors and missing ledger entries, resulting in numerous false alarms. Furthermore, they can only identify surface-level numerical anomalies and cannot fully trace the underlying causes of these deviations, lacking a complete and interpretable logical chain for risk analysis.

[0003] Current risk analysis models for R&D scenarios generally suffer from insufficient scenario adaptability. They lack sufficient sample reserves for niche or critical risk scenarios such as abnormal R&D expense collection, experimental data deviation, R&D cycle delays, and lagging team collaboration. Furthermore, model judgment rules are often rigid and cannot be iteratively optimized to adapt to changes in enterprise R&D operations. Most management systems simply perform data correlation statistics, failing to distinguish between superficial data correlations and actual business causal logic. This easily leads to false correlations interfering with risk assessment results. Moreover, they lack a closed loop of human-machine collaborative sample accumulation, data expansion, and targeted training, making it difficult to continuously improve the accuracy of R&D anomaly identification and root cause tracing. Consequently, they fail to meet the actual business needs of enterprises for standardized R&D process construction, compliant R&D expense control, and refined management of the entire R&D project process. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a multimodal enterprise project management method based on big data, the method comprising: Modal hierarchical decoupled differential feature mining is implemented for multimodal heterogeneous data of projects. Feature extraction strategies are adopted for three types of modalities: structured time series data, text data, and audiovisual multimedia data. Project risk correlation features corresponding to each modality are independently mined. Intramodal noise reduction and filtering are performed on the original features of each modality to remove redundant noise and invalid fluctuation features, so as to obtain a set of pure multimodal features without cross interference. Establish a cross-modal data accurate alignment mechanism, using the unique process identifier of the project as the binding primary key, and perform unified association and calibration of multiple types of heterogeneous data under the same project task in terms of subject, scenario and time dimensions. Use time-series correction and spatial matching algorithms to remove irrelevant data from different sources. A two-tiered risk assessment mechanism is constructed that integrates statistical dynamic thresholds and business rigid thresholds. Based on the distribution characteristics of historical data, fluctuation thresholds that adapt to the dynamic changes of projects are generated, and fixed business thresholds are configured in combination with industry and institutional standards. Through the joint constraint of dual conditions, the accurate distinction between real risks and normal data fluctuations is achieved. Construct a causal knowledge system for project risks and introduce a causal reasoning network. Establish characteristic causal relationships based on historical risk cases. Eliminate false correlation features through causal structure learning. Combine reverse tracing reasoning to achieve hierarchical analysis of risk manifestations and essential causes. Establish a self-iterative optimization closed loop for risk assessment in human-machine collaboration, perform sample classification and labeling and backfeed training on the model assessment results, and continuously iterate and update the model inference weights, feature selection rules and risk judgment parameters by combining risk scenario sample enhancement strategies. A multimodal feature cross-validation and voting fusion judgment mechanism is adopted. Single-modal anomalies are used as suspected clues and multimodal homologous feature joint verification is used as the judgment basis. Data noise and real risks are identified through multi-dimensional feature consistency constraints.

[0005] Furthermore, embodiments of the present invention also provide an enterprise project management system based on big data and multimodal architecture, characterized in that it includes: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the above-described enterprise project management method based on big data multimodality by executing the machine-executable instructions.

[0006] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, the processor of a computer device reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the computer device to execute the above-described enterprise project management method based on big data multimodal.

[0007] Based on the above, by implementing hierarchical decoupling, cross-dimensional alignment, and dual-level threshold joint judgment on multimodal heterogeneous data, the problems of traditional single-dimensional judgment being susceptible to data noise interference and having a high false alarm rate are effectively solved. Relying on causal structure learning to eliminate false correlation features and building a four-level reverse tracing architecture, it achieves layer-by-layer analysis from surface anomalies and mid-level deviations to the underlying root causes, enabling risk assessment to have complete causal logic and interpretability. This significantly improves the accuracy of project risk identification, screening, and tracing, and can fully adapt to dynamic changes in operating conditions at different stages of a project.

[0008] This invention constructs a human-machine collaborative self-iterative closed loop integrating sample accumulation, data augmentation, hierarchical training, and parameter updates. It continuously addresses the shortcomings of samples in niche, critical, and extreme scenarios, and optimizes model inference weights, feature rules, and decision parameters in a targeted manner. Combined with multimodal cross-validation and voting fusion mechanisms, it further enhances anti-interference capabilities, enabling the system to autonomously evolve with changes in project conditions and risk types, maintaining stable risk control capabilities over the long term, and comprehensively meeting the intelligent, refined, and dynamic risk management needs of enterprises throughout the entire project process. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the execution flow of the enterprise project management method based on big data and multimodal operation provided in this embodiment of the invention.

[0010] Figure 2 This is a schematic diagram of exemplary hardware and software components of an enterprise project management system based on big data and multimodal architecture provided in an embodiment of the present invention. Detailed Implementation

[0011] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating an embodiment of the enterprise project management method based on big data and multimodal approaches provided by the present invention. The following is a detailed description of the enterprise project management method based on big data and multimodal approaches.

[0012] Step S110: Implement modal hierarchical decoupled differential feature mining for multimodal heterogeneous data of the project. For the three modalities of structured time series data, text data, and audiovisual multimedia data, feature extraction strategies are adopted respectively to independently mine the project risk association features corresponding to each modality. Intramodal noise reduction and filtering are performed on the original features of each modality to remove redundant noise and invalid fluctuation features, and obtain a set of pure multimodal features without cross interference.

[0013] To address the multimodal heterogeneous data within project scenarios, a differentiated feature mining operation with modal layering and decoupling is implemented. A unified feature extraction model is abandoned, and an appropriate extraction strategy is adopted for three core data modalities. For structured time-series data, the focus is on mining risk-related features such as fluctuations, numerical mutations, and stage shifts in project schedules, costs, resources, and quality indicators. For text data, implicit risk features such as semantic anomalies, problem descriptions, and deviation notes in project ledgers, rectification records, and work logs are analyzed. For audiovisual multimedia data, explicit risk features such as process violations, deviations in requirement implementation, and non-standard collaboration processes in R&D screen recordings and review recordings are identified. After completing the independent feature mining for each modality, intra-modal noise reduction and filtering are performed to remove equipment acquisition noise, invalid data fluctuations, redundant features, and interference information without business significance. This completely avoids the problem of cross-coupling interference between different modal features, ultimately resulting in a high-purity, interference-free multimodal pure feature set suitable for project risk assessment.

[0014] Step S120: Establish a cross-modal data accurate alignment mechanism, using the unique process identifier of the project as the binding primary key, and perform unified association calibration of multiple heterogeneous data under the same project task in terms of subject, scenario, and time dimensions. Remove irrelevant data from different sources through time-series correction and spatial matching algorithms.

[0015] A cross-modal data alignment mechanism adapted to all project scenarios is established to solve the fusion distortion problems caused by spatiotemporal misalignment, scenario confusion, and subject mismatch in multi-source heterogeneous data. Using a predefined unique process identifier for the project as the global binding primary key, all types of heterogeneous data under the same project task and work scenario are locked, and comprehensive unified correlation calibration is carried out from three core dimensions: work subject, test scenario, and collection time. To address data deviation issues caused by different collection devices and collection frequencies, mature time-series correction algorithms and spatial matching algorithms are used to accurately screen and verify all heterogeneous data. Irrelevant data from different sources across time periods, processes, and work scenarios is accurately identified and removed, preventing invalid data from interfering with subsequent analysis. This mechanism achieves accurate source attribution of multimodal heterogeneous data, ensuring that all data involved in risk assessment are from the same valid source, thus guaranteeing the accuracy of multimodal data fusion analysis and risk identification from the ground up.

[0016] Step S121: Assign a globally unique data sequence number to the multimodal heterogeneous data sources, including time-series data, text data, image and video data, operation log data, and cost ledger data, which are connected to the project management system. Simultaneously parse the collection timestamp, equipment number, collection point, and operator fields corresponding to each piece of raw data to construct a multimodal raw data metadata structure. Perform field integrity cleaning on the raw data at the data access layer to remove dirty data with missing primary keys, missing time information, and missing scene location information, and filter out valid raw multimodal data with basic association and matching attributes. The system completes the input encoding, metadata construction, and pre-cleaning of multimodal heterogeneous data, unifying access to all data sources within the project management system, including time-series monitoring, text records, images and videos, operation logs, and cost ledgers. Each piece of raw data is assigned a globally unique data sequence number, enabling full data traceability and location-based management. Core metadata fields such as collection timestamp, device number, collection point, and operator are automatically extracted from each data entry, building a standardized multimodal raw data metadata structure to fully preserve all attributes of the collected data. Strict field integrity cleaning is performed at the data access layer, batch-screening and removing various types of dirty data lacking primary keys, time information, and scene location information. Only valid raw data with complete metadata and cross-modal correlation matching is retained, controlling data quality from the source and avoiding subsequent alignment failures caused by incomplete data.

[0017] Step S122: Pre-configure the project's unique process ID as the global binding primary key, construct the process primary key index mapping table, establish a one-to-one correspondence between each R&D iteration process, version development node, and operation and maintenance task in the entire project lifecycle and the unique process ID, and mount all valid original multimodal data to the corresponding process ID node according to the business affiliation. By leveraging unique process primary keys, standardized aggregation and mounting of all valid data is achieved. A unique process ID is pre-configured as the globally unified binding primary key, and a standardized process primary key index mapping table is built. For various R&D iteration processes, R&D iteration nodes, and operation and maintenance tasks covered throughout the project lifecycle, independent and unique process IDs are assigned, establishing a one-to-one binding relationship between business tasks and process IDs. Based on the business affiliation attributes of each valid original multimodal data entry, all cleaned valid data is accurately mounted under the corresponding process ID node, forming a hierarchical data aggregation architecture centered on processes. This completely solves the industry pain points of traditional project multimodal data's scattered storage, ambiguous business affiliation, and chaotic scenario correspondence, allowing each piece of data to accurately match specific project business links, laying a unified business anchor point for subsequent precise alignment across multiple dimensions (time, space, and subject).

[0018] Step S123: Perform precision standardization processing on the original timestamps of all multimodal data mounted under the same process ID, and uniformly convert them into millisecond-level standard time format. Use dynamic time warping algorithm to correct the timing offset and acquisition delay of different devices, calculate the time offset of various modal data relative to the standard time axis of the process, and perform overall translation calibration on the video frame acquisition time, voice recording period, log filling time, and sensor sampling time based on the obtained offset. Complete the time-series standardization calibration of multimodal data within the same process. For all data attached to the same process ID, perform unified optimization processing of the time dimension. Convert non-standard raw timestamps output by various devices into a unified millisecond-level standard time format to eliminate the fundamental deviation of inconsistent time accuracy among different acquisition devices. Employ a dynamic time warping algorithm to specifically correct issues such as time series offsets, data acquisition delays, and reporting time differences existing in various devices. Accurately calculate the precise time offset of different modal data, such as video, voice, sensor, operation logs, and cost ledgers, relative to the standard time axis of the process. Based on the calculated offset, perform overall translation calibration on the video frame acquisition time, voice recording period, log filling time, and sensor sampling time to achieve complete synchronization of the time axis of all multimodal data within the same process, completely resolving the risk feature matching deviation problem caused by time series misalignment.

[0019] Step S124: Extract the spatial coordinates of each multimodal data collection, the location of the R&D workstation equipment, the R&D area number, and the R&D scene label. Construct a project R&D spatial scene topology matrix. Grid-encode the project R&D office area, R&D workstation, review scene, and maintenance point. Compare and match the spatial location information of the multimodal data with the project's standard scene grid using a scene space matching algorithm. Filter and remove cross-scene, cross-process, and non-homogeneous heterogeneous data with spatial coordinate deviations exceeding a preset threshold, scene coding mismatches, or inconsistent R&D office areas. Retain valid data that completely matches the current process scene. To achieve precise spatial matching and invalid data removal for multimodal data, spatial information such as acquisition coordinates, on-site equipment locations, R&D office area numbers, and work scenario tags are extracted from each piece of multimodal data to construct a comprehensive project spatial scenario topology matrix. All R&D office areas, workstations, R&D scenarios, and maintenance points are finely coded into a grid to form a standardized project spatial grid system. Based on a scenario spatial matching algorithm, the spatial location information of each piece of multimodal data is compared and matched against the project's standard scenario grid, and data is filtered according to preset spatial deviation standards. Cross-scenario, cross-process, and heterogeneous data with excessive spatial coordinate deviations, mismatched scenario codes, or inconsistent work areas are precisely removed, retaining only valid data that perfectly matches the current process scenario. This completes spatial data purification and alignment, ensuring the uniqueness and accuracy of data scenario matching.

[0020] Step S125: Bind the main information of the person in charge, tester, R&D personnel, operation and maintenance role, and approver corresponding to the current process, automatically match the operator, recorder, and reviewer fields in the multimodal data, verify the consistency between the data behavior subject and the project task subject, and filter out external interference data and irrelevant operation data that do not match the subject. To achieve precise data filtering at the personnel level, multimodal data operation consistency verification is implemented. First, the system binds the main information corresponding to the current work process, including information on core personnel such as project managers, testers, developers, operations and maintenance personnel, and approvers, constructing a work process subject database. The system automatically matches the operator, recorder, and reviewer fields recorded in the multimodal data, performing a comprehensive consistency comparison and verification between the data behavior subject and the project's preset task subject. Data from external personnel with mismatched personnel information, data not belonging to the current work process team, and irrelevant data are uniformly identified as interference data and filtered out. Through precise verification at the subject level, external irrelevant data interference is completely eliminated, ensuring that all retained data is generated by compliant personnel currently working in the work process, further strengthening the homogeneity, authenticity, and business relevance of multimodal data.

[0021] Step S126: Based on the calibrated unified timeline, standardized scene grid, verified task subject and corresponding process event, complete the integrated binding and data fusion of the four dimensions of time, space, subject and event to generate a multimodal aligned dataset with the same source, scene, subject and time period.

[0022] By integrating multi-dimensional calibration results, a unified fusion and alignment of multimodal data is achieved. Based on the unified standard timeline, standardized scene grid, and verified task subject information calibrated in the aforementioned steps, and combined with the business events corresponding to each process, a four-dimensional fusion binding mechanism integrating time, space, subject, and event is established. Various heterogeneous multimodal data belonging to the same process, time period, workspace, and execution subject are aggregated and integrated to achieve deep binding and unified data fusion processing across four dimensions. The final result is a standardized multimodal aligned dataset that is completely homogeneous, with unified scenes, consistent subjects, and synchronized time periods. This completely solves the problems of misalignment, mixing, and low correlation of multi-source heterogeneous data. This dataset can be directly used for subsequent threshold verification, risk identification, causal tracing, and other full-process analysis work.

[0023] Step S130: Construct a two-layer risk assessment mechanism that integrates statistical dynamic thresholds and business rigid thresholds. Generate fluctuation thresholds that adapt to the dynamic changes of projects based on the distribution characteristics of historical data, and configure fixed business thresholds in combination with industry and institutional standards. Achieve accurate distinction between real risks and normal data fluctuations through dual-condition joint constraints.

[0024] A risk assessment mechanism integrating dynamic and rigid thresholds is constructed to address the shortcomings of traditional single thresholds, which cannot adapt to dynamic project conditions and have low accuracy in risk assessment. Combining the characteristics of the entire project lifecycle and relying on the distribution features of massive historical normal operation data, a statistical fluctuation threshold is fitted and generated that dynamically changes with the project's testing, development, and maintenance phases, adapting to the fluctuations in normal operating data at different stages of the project. Simultaneously, a fixed, rigid business threshold is configured in strict accordance with industry standards, corporate policies, contract terms, and specific technical specifications, defining the hard red lines for project compliance. Through a dual-condition constraint model of dynamic fluctuation tolerance and rigid compliance constraints, normal data fluctuations, minor anomalies, and genuine violation risks are differentiated in layers, enabling accurate identification of risk signals and normal operating fluctuations in complex project scenarios, significantly reducing the probability of misjudgment and omission of project risks.

[0025] Step S131: Collect project progress time-series data, cost consumption data, resource input data, and quality inspection data covering different project types, different iterative development stages, and different R&D cycles to form a complete historical operation dataset of the project. Preprocess the historical operation dataset to remove manually entered abnormal data, system fault data, and marked fault sample data. Filter and retain valid sample data under normal project operation conditions to construct a benchmark dataset of normal project operation conditions. A high-purity benchmark dataset of normal project operating conditions was constructed, extensively collecting project operation data covering various project types, different iterative development stages, and the entire R&D cycle. This data includes core indicators such as schedule timing, cost consumption, resource input, and quality inspection, forming a complete historical raw dataset of project operations. Comprehensive and meticulous preprocessing was performed on the raw dataset, employing a dual screening process combining manual and system rules to accurately remove invalid sample data, including manually entered data, data collected due to system malfunctions, equipment malfunction reports, and previously marked faults. Only valid sample data under stable and normal project operating conditions were retained, ensuring that all retained data accurately and objectively reflects the project's routine operating patterns. Finally, this data was integrated to form a comprehensive, highly accurate, and fault-free benchmark dataset of normal project operating conditions.

[0026] Step S132: Fit the data distribution characteristics of each project risk indicator based on the project's normal operating condition benchmark dataset, use the Gaussian normal distribution algorithm to calculate the mean and standard deviation of each risk indicator, combine the moving range analysis algorithm to perform dynamic interval extrapolation of the time series continuous fluctuation data, configure an adaptive sliding time window according to the differences in project stages, and statistically analyze the extreme value range of each indicator window by window to generate the upper and lower limits of normal fluctuation of the indicators that dynamically and adaptively change with the project's R&D iteration progress, version development cycle, and operation and maintenance stage, and construct a statistical dynamic threshold that adapts to the dynamic operating characteristics of the project. An adaptive statistical dynamic threshold is constructed based on a normal operating condition benchmark dataset. Distribution characteristic fitting analysis is performed on massive sample data of risk indicators for each project to accurately capture the normal operating patterns of these indicators. Combining the characteristics of time-series data, a moving range analysis algorithm is used to dynamically extrapolate continuously fluctuating time-series data, adapting to the operational differences in different testing, development, and maintenance phases of the project, and configuring an adaptively adjustable sliding time window. Within each independent time window, the normal fluctuation extreme value range of each risk indicator is statistically analyzed. Combined with the overall data distribution characteristics, the upper and lower limits of normal fluctuations in indicators at each stage are defined. Finally, a statistical dynamic threshold is generated that can autonomously adapt and update with project progress, cycle iteration, and changes in operating conditions, accurately accommodating reasonable data fluctuations at different project stages and adapting to the complex characteristics of dynamic project operation.

[0027] Step S133: Based on national industry standards, enterprise project management system, project contract constraints and project-specific technical specifications, sort out the rigid constraints and hard control boundaries corresponding to each risk indicator, configure fixed business rigid thresholds for each type of risk such as cost overrun, R&D cycle delay, quality defects, resource overrun, and collaboration timeout, build a structured business rule threshold library, bind and map each business rigid threshold with the corresponding project indicators, risk types and process scenarios, and establish a rule matching relationship that can be called in real time; A standardized, structured database of rigid business thresholds has been established. Based on current national and industry standards, corporate project management systems, project contractual constraints, and project-specific technical specifications, the database comprehensively reviews the mandatory control requirements and compliance boundaries for various project risk indicators. For all risk types, including cost overruns, R&D delays, quality defects, resource overruns, and collaboration timeouts, corresponding fixed control standards are matched one-to-one, and rigid business thresholds that cannot be arbitrarily changed are configured. All rigid thresholds are structured and archived to build a standardized business rule threshold database. Simultaneously, a one-to-one binding and mapping of rigid thresholds, monitoring indicators, risk types, and process scenarios is completed, establishing a rule association system that can be called in real-time and quickly matched. These thresholds serve as the hard red lines for compliant project operations, unaffected by fluctuations in operational data, and are used to accurately determine substantive violation risks, ensuring the compliance and rigor of risk assessment.

[0028] Step S134: Establish a two-layer progressive risk verification mechanism that combines dynamic fluctuation tolerance with rigid rule constraints. Use statistical dynamic thresholds to define the tolerance range of normal data fluctuations in the project, and use business rigid thresholds to define the boundary for judging project violation risks. Perform dual-threshold hierarchical joint verification on the real-time collected multimodal indicator data of the project. According to the different levels of the indicator values ​​in the dynamic threshold range and rigid threshold boundary, distinguish between normal operating condition fluctuations, slight abnormal fluctuations and real violation risks, so as to achieve graded judgment, differentiated processing and accurate risk identification. A two-tiered, progressive risk verification mechanism is established, integrating dynamic fluctuation tolerance with rigid rule constraints to achieve precise, tiered identification of project risks. An adaptively updated statistical dynamic threshold serves as a flexible constraint boundary, defining the data fluctuation tolerance range for normal operating conditions at each stage of the project, adapting to the dynamic operational characteristics of the project. A fixed, rigid business threshold serves as a compliance constraint red line, defining the criteria for judging project violation risks. For real-time project indicator data that has undergone multimodal alignment processing, a dual-threshold tiered, progressive verification is conducted, classifying three operating conditions based on the different intervals in which the indicator values ​​fall. This effectively distinguishes between normal operating condition fluctuations, minor abnormal fluctuations that do not touch the compliance red line, and genuine violation risks that exceed control standards. Differentiated data processing and early warning strategies are implemented for different states, achieving graded judgment, precise identification, and tiered handling of project risks.

[0029] Step S1341: Construct a two-layer nested risk verification architecture, which hierarchically binds the pre-built statistical dynamic thresholds with the business rigid thresholds. The statistical dynamic thresholds serve as the first layer of flexible fluctuation filtering boundaries, which can adapt to and accommodate normal operating data fluctuations in different iterative development stages, R&D cycles, and operation and maintenance stages of the project. The business rigid thresholds serve as the second layer of compliance risk judgment red lines, which define the compliance control standards that the project cannot cross, thus forming a two-layer risk judgment system in which flexible adaptive constraints and rigid compliance constraints are nested together. A two-layer nested risk verification architecture is constructed to achieve logical linkage and binding between flexible and rigid constraints. The first layer, flexible fluctuation filtering boundary, is set as a statistical dynamic threshold, fully adapting to the differentiated operational characteristics of different project iteration development stages, R&D cycles, and operation and maintenance phases. This accommodates reasonable data fluctuations during normal project production and operation, preventing normal operating condition fluctuations from being misjudged as risk anomalies. The second layer, compliance risk judgment red line, is set as a rigid business threshold, strictly defining inviolable control boundaries based on industry and corporate compliance standards, specifically used to identify substantive violation risks. This two-layer nested verification architecture, progressively advancing, balances the dynamic flexibility of project operation with the seriousness of compliance control, forming a two-layer nested risk judgment system adaptable to complex project scenarios.

[0030] Step S1342: Perform the first-level dynamic threshold progressive verification, and access the project multimodal standardized indicator data after the synchronization and alignment process is completed in real time. Input the single real-time indicator data into the double-layer verification link for progressive judgment. First, check whether the real-time indicator value is within the upper and lower limits of the statistical dynamic threshold corresponding to the current project stage. If the real-time indicator value falls within the dynamic fluctuation range, it is determined that the current project process is a normal working condition fluctuation. Only the data is retained and recorded, and no abnormal marking and risk warning operations are performed. The first layer of the two-layer verification architecture performs dynamic threshold progressive verification. It continuously receives multimodal indicator data from the project after four-dimensional alignment and standardization. Individual real-time indicator data is fed into the two-layer verification chain for progressive judgment. Priority is given to matching the statistical dynamic threshold range corresponding to the current project process and operational stage to verify whether the real-time indicator value is within the normal fluctuation range. If the real-time indicator value falls within the dynamic threshold range, the current process operation status is determined to be under normal operating conditions. Only this valid data is archived and time-series recorded; subsequent operations such as anomaly marking, risk warning, and root cause analysis are not performed. This completes the rapid filtering of routine operating condition data, reduces invalid judgment processes, and improves system operating efficiency.

[0031] Step S1343: Trigger the second-level business rigid threshold verification. If the real-time indicator value exceeds the normal fluctuation range of the statistical dynamic threshold, it is determined that the indicator has an abnormal deviation from the working condition. The second-level business rigid threshold verification process is automatically entered. The current real-time indicator value is matched and compared with the business rigid threshold bound to the corresponding risk type and process scenario to verify whether the indicator value reaches or exceeds the compliance control red line. The system triggers a second-level business rigid threshold compliance check, taking over any abnormal results from the first-level check. When a real-time indicator value exceeds the normal fluctuation range of the current stage's statistical dynamic threshold, it determines that the current indicator has a significant deviation from the operating condition, excluding the possibility of normal data fluctuations, and automatically jumps to the second-level rigid threshold check process. Based on the current process scenario, indicator type, and risk category, the system accurately retrieves the pre-bound business rigid threshold, precisely compares the real-time indicator value with the compliance control red line, and verifies whether the current indicator anomaly meets or exceeds the rigid control standards stipulated by the industry, system, and contract, accurately distinguishing between minor operating condition deviations and substantial compliance risks.

[0032] Step S1344: If the indicator exceeds the statistical dynamic threshold range but does not reach the business rigid threshold standard, it is judged as a slight abnormal fluctuation state. The abnormal feature is time-series marked and archived for subsequent model sample iteration and working condition data update only. It does not trigger the formal project risk warning and root cause analysis process. If the indicator value simultaneously exceeds the normal fluctuation tolerance range of the current stage statistical dynamic threshold and meets or exceeds the business rigid threshold constraint conditions of the corresponding risk scenario, it is judged as a real and valid project violation risk. The current process node, risk type, abnormality degree, and multimodal abnormal features are jointly marked to lock the risk event and automatically trigger the subsequent multimodal causal reasoning root cause localization process, completing the hierarchical identification and graded handling of normal working condition fluctuations, slight abnormal fluctuations, and real violation risks.

[0033] This system enables risk stratification and tiered handling, facilitating the final linkage judgment and process triggering based on dual thresholds. If an indicator only exceeds the statistical dynamic threshold range but does not reach the business rigid threshold standard, it is judged as a minor abnormal fluctuation. The system only marks and archives this abnormal feature, incorporating the data into the sample iteration library for subsequent model optimization and operational data updates, without triggering formal risk warnings or root cause analysis processes. If an indicator simultaneously exceeds both the dynamic threshold fluctuation range and the business rigid threshold red line, it is judged as a genuine and valid project violation risk. The system immediately performs joint marking of the current process node, risk type, abnormality level, and multimodal abnormal features, locking the corresponding risk event and automatically triggering the entire process of subsequent multimodal causal reasoning, reverse tracing, and root cause localization, achieving a closed-loop linkage from anomaly identification to root cause judgment.

[0034] Step S135: Construct a dual-threshold dynamic iterative update mechanism. Based on the real-time operation data of the project and the manually verified samples, continuously iterate and optimize the distribution characteristics of the benchmark data of the working conditions, periodically and adaptively update the dynamic statistical thresholds, so that the fluctuation tolerance range adapts to the real-time working condition changes of the project. At the same time, solidify the rigid threshold rule parameters of the business, keep the compliance judgment boundary constant, and rely on the collaborative judgment logic of dynamic adaptive constraints and fixed rigid constraints.

[0035] A dual-threshold differentiated iterative update mechanism is constructed to achieve dynamic adaptive optimization of the risk assessment system and solidification of compliance boundaries. Based on real-time operational data throughout the project lifecycle and manually reviewed and annotated samples, benchmark data is continuously updated to dynamically optimize the fluctuation range of statistical thresholds, allowing flexible constraint boundaries to continuously adapt to dynamic changes in project testing, development, and maintenance conditions. Simultaneously, rigid business thresholds are solidified and protected, relying on access control and data verification mechanisms to lock in compliance assessment boundaries, preventing manual or programmatic tampering and ensuring that rigid red lines remain constant in the long term. Through the synergistic cooperation of dynamically adaptive flexible thresholds and fixed compliance rigid thresholds, the system's adaptability to changes in project conditions is preserved while ensuring the uniformity, standardization, and authority of risk compliance assessments, achieving a balance between dynamic optimization and standard solidification.

[0036] Step S1251: Construct an incremental update mechanism for working condition baseline data, continuously collect real-time multimodal operation data, system risk assessment results data, and manually reviewed and labeled sample data throughout the entire project process, and merge and summarize the newly added real-time data with the historical normal working condition baseline dataset to construct an incremental update dataset. An incremental update mechanism for baseline operating data is established to enable dynamic expansion and continuous optimization of the baseline dataset. The system continuously collects real-time multimodal operational data throughout the entire project lifecycle, simultaneously aggregating multidimensional data such as automatic risk assessment results, expert review conclusions, and labeled / corrected samples. The daily added real-time operating data and verification sample data are merged and summarized with the historical normal operating condition baseline dataset, continuously expanding the dataset sample size and enriching the scenario coverage dimensions to construct a dynamically updated dataset. This continuously updated dataset accurately captures the latest operational patterns and operating condition changes of the project, compensating for the inability of static baseline datasets to adapt to iterative project changes.

[0037] Step S1352: Perform unified cleaning and filtering on the incremental update dataset to remove abnormal data caused by instantaneous equipment jitter, temporary malfunction data and invalid duplicate data, and retain valid operating condition data that can characterize the latest operating status of the project. Refined cleaning and filtering of incremental datasets ensures high-quality iterative data. For the full dataset after incremental updates, standardized cleaning and noise reduction processes are performed to accurately screen and remove various invalid and interfering data. Special attention is paid to filtering transient abnormal data caused by equipment jitter and signal fluctuations, removing invalid records resulting from temporary human errors or system malfunctions, and deleting redundant data from duplicate collections and invalid entries. Only valid data that accurately represents the latest operational status of the project and conforms to normal operating conditions is retained, thoroughly purifying the incremental dataset and preventing inferior data from interfering with the accuracy of threshold iterations, ensuring that subsequent dynamic threshold updates accurately adapt to the latest operating characteristics of the project.

[0038] Step S1353: Configure a dual-mode adaptive threshold iteration update strategy, preset a fixed-cycle iteration mechanism, automatically start the dynamic threshold update process every set fixed cycle, and add an adaptive triggering iteration mechanism for working condition fluctuations to monitor the time-series fluctuation status of project indicators in real time. When the variance of indicator fluctuations in the continuous sliding window exceeds the preset fluctuation change threshold, the emergency iteration update process is actively triggered. A dual-mode adaptive threshold iteration strategy is configured to achieve intelligent and accurate updates of dynamic thresholds. A dual triggering mechanism of fixed-cycle iteration and condition-adaptive iteration is established, forming a complementary iterative update system. With a preset fixed update cycle, the system automatically initiates a full dynamic threshold update process at predetermined intervals to complete threshold iteration optimization under normal operating conditions. Simultaneously, an adaptive emergency iteration mechanism is added to monitor the temporal fluctuation status of various project indicators in real time, continuously statistically analyzing indicator fluctuations within a continuous sliding window. When the overall fluctuation amplitude exceeds the preset range or a significant change in operating conditions occurs, an emergency iteration update process is proactively triggered to quickly adapt to sudden changes in operating conditions. The dual-mode iteration strategy balances regular updates and emergency optimization, significantly improving the accuracy of threshold adaptation.

[0039] Step S1354: Based on the iteratively updated normal operating condition benchmark dataset, refit the Gaussian distribution characteristics and time series change patterns of each risk indicator. Combine the moving range analysis algorithm and the adaptive sliding time window to recalculate the upper and lower limits of normal fluctuations of the indicators corresponding to each project stage. Complete the adaptive correction of the statistical dynamic threshold so that the dynamic fluctuation tolerance range can be adapted in real time to the dynamic change characteristics of project version iteration progress, R&D iteration, and operation and maintenance conditions. The adaptive correction and optimization of the statistical dynamic thresholds were completed. Based on the cleaned incremental benchmark dataset, the data distribution characteristics and temporal variation patterns of each risk indicator were refitted. Combining the moving range analysis algorithm and the adaptive sliding time window mechanism, the normal fluctuation extreme value ranges of the indicators corresponding to each testing, R&D, and operation and maintenance phase of the project were recalculated, and the upper and lower limits of the dynamic thresholds for each phase were updated. Through continuous iterative correction, the dynamic fluctuation tolerance range can adapt to the data changes brought about by the progress of the project, the iteration of working conditions, and the adjustment of work modes in real time. It always keeps in line with the latest normalized operating characteristics of the project, effectively solving the problems of poor adaptability and high false judgment rate of traditional static thresholds, and continuously improving the accuracy of flexible risk filtering.

[0040] Step S1355: In the entire process of statistical dynamic threshold iteration and update, a parameter anti-tampering protection mechanism is implemented for the rule database of rigid thresholds for storage business. All rigid threshold parameters for business are fixed and constrained through permission encryption locking, rule read-only configuration, and parameter consistency verification. This restricts the addition, modification, and deletion of rigid thresholds by automated iteration programs and manual operations, so that the compliance judgment boundary established based on industry standards, corporate management systems, project contract constraints, and special technical specifications remains constant and does not change with fluctuations and iteration updates of project working data. Implement end-to-end tamper-proof and fixed protection for rigid business thresholds. Throughout the entire process of dynamic threshold iteration and updates, the rule database storing rigid business thresholds is securely locked and controlled. Through a multi-layered protection mechanism including tiered access control with encryption, read-only rule configuration, and cyclical parameter consistency checks, all rigid threshold parameters are fixed and constrained. Strict restrictions are placed on automated iteration programs and manual operations to add, modify, delete, or adjust rigid thresholds, preventing arbitrary alterations to compliance judgment boundaries. Ensure that the rigid red lines, based on national industry standards, corporate management systems, project contract constraints, and specific technical specifications, remain constant and consistent, unaffected by project fluctuations or data iterations, guaranteeing the authority and consistency of risk compliance judgment standards.

[0041] Step S1356: After each iteration and correction of the dynamic statistical threshold, reconstruct the linkage verification mapping relationship between the updated dynamic statistical threshold and the fixed business rigid threshold, and synchronously update the matching parameters of the two-layer progressive risk verification architecture to form a judgment system that works in concert with dynamic adaptive flexible fluctuation constraints and long-term constant rigid compliance constraints.

[0042] The system has completed the reconstruction and iterative upgrade of the dual-threshold linkage relationship. After each round of dynamic statistical threshold correction, the system automatically reconstructs the linkage verification mapping relationship between the updated dynamic flexible threshold and the fixed rigid threshold. The hierarchical matching parameters, interval judgment rules, and progressive verification logic of the two-layer progressive risk verification architecture are updated simultaneously, optimizing the dual-threshold collaborative working mechanism. Ultimately, a collaborative judgment system is formed, where dynamic adaptive flexible fluctuation constraints adapt to changes in operating conditions, while long-term constant rigid compliance constraints safeguard the risk control red line. This ensures that the dual-layer risk judgment mechanism always adapts to the latest operational status of the project, while maintaining unchanged compliance judgment standards, continuously improving the comprehensive performance of project risk stratification and accurate judgment.

[0043] Step S140: Construct a project risk causal knowledge system and introduce a causal reasoning network. Establish characteristic causal relationships based on historical risk cases. Eliminate false association features through causal structure learning. Combine reverse tracing reasoning to achieve hierarchical analysis of risk manifestations and essential causes. This project constructs a risk causal knowledge system and intelligent reasoning mechanism adapted to project management scenarios, breaking through the limitations of traditional data correlation analysis that can only capture superficial correlations. It relies on real project risk cases to precipitate standardized causal logic. By collecting historical risk data from multiple scenarios, it systematically sorts out the real business relationships between multimodal features, risk manifestations, process deviations, and underlying root causes, forming a structured and reusable project risk causal knowledge chain. Simultaneously, it introduces a multimodal causal reasoning network, using causal structure learning algorithms to deeply train massive feature data, automatically identifying and eliminating pseudo-correlation features, noisy features, and redundant features without business support, and uncovering stable and reliable risk causal dependencies. Combined with a four-level hierarchical reverse tracing reasoning logic, it achieves layer-by-layer decomposition and analysis from surface-level abnormal signals, mid-level process deviations to underlying essential causes, completing a layered, accurate, and interpretable assessment of project risk manifestations and deep-seated causes, and constructing a complete intelligent causal analysis system for project risks.

[0044] Step S141: Construct a standardized project risk causal case knowledge base, which collects historical project risk cases, risk rectification records, anomaly ledgers, and multimodal risk feature samples covering multiple scenarios such as R&D cycle delays, cost overruns, quality defects, resource overruns, and collaboration timeouts. All risk cases are structured and labeled, and each case is bound to the corresponding multimodal original features, risk manifestations, root cause types, and rectification and handling results to form a project risk original case library. A comprehensive and structured original knowledge base of project risk causal cases has been established, fully collecting typical risk data throughout the entire project lifecycle, covering mainstream risk scenarios such as R&D cycle delays, cost overruns, quality defects, resource overruns, and collaboration timeouts. Various historical risk cases, on-site anomaly logs, problem rectification records, fault handling files, and corresponding multimodal risk feature samples are uniformly compiled to achieve full coverage of risk data across multiple scenarios. Each risk case in the database is meticulously and structurally labeled, binding the original multimodal feature data, external manifestations of the risk, the type of root cause of the risk, the process scenario in which the problem occurred, and the final rectification and handling results to each case. This establishes a full-dimensional correlation record of case characteristics, risk phenomena, root causes, and solutions, forming a complete, scenario-wide, traceable, and reusable original knowledge base of project risk cases.

[0045] Step S142: Perform cleaning and normalization on the original case database data, remove invalid cases with incomplete records, ambiguous causal relationships, and untraceable sources, retain high-quality risk samples with complete features, clear root causes, and specific scenarios, build a project risk causal knowledge system based on high-quality risk samples, sort out the characteristic manifestations, intermediate influencing factors and causal chains corresponding to various risks, and establish the correlation mapping relationship between multimodal features and the root causes of project risks. The process involved data purification and the construction of a causal knowledge system for the original case library. Strict data cleaning and normalization were performed on the massive amount of original risk cases entering the library, eliminating low-quality and invalid cases with incomplete information, ambiguous causal relationships, or inability to trace the source. Only high-quality risk samples with complete feature information, clear risk scenarios, well-defined causal chains, and accurate root cause identification were retained, ensuring the purity and reliability of the knowledge base data. Based on these selected high-quality samples, a systematic project risk causal knowledge system was built, comprehensively analyzing the front-end characteristics, intermediate influencing factors, and underlying core causes of various risks, and fully dissecting the transmission chains of different risk types. Finally, a precise mapping relationship was established between multimodal risk characteristics and the intermediate-level deviations and underlying risk root causes in the project.

[0046] Step S143: Construct a multimodal causal reasoning network model structure. Use the preprocessed multimodal pure feature vector as the network input layer, various risk events of the project as observed variables, and various potential risk factors of the project as latent variables. Construct a causal structure graph adapted to the project management scenario. Use the causal structure learning algorithm to traverse and train massive multimodal feature data, mine the conditional independence relationship and real causal edge between each feature variable, calculate the causal correlation degree of each feature relative to the risk event, screen out highly correlated causal features, and remove pseudo-correlation features, noise features, and redundant interference features that only have statistical correlation and no actual business causal logic. A multimodal causal reasoning network model adapted to project risk assessment is constructed. This model input system is built upon pre-refined multimodal clean feature vectors, and a complete variable system and causal topology are defined based on the project's risk business characteristics. Various multimodal clean features are used as observable inputs, explicit project risk events are used as output variables, and implicit factors such as personnel, equipment, materials, plans, processes, and management are used as latent variables, constructing a causal structure diagram that fits the project management scenario. A causal structure learning algorithm is used to train the model batch-by-batch on massive multimodal feature samples, continuously examining the conditional independence relationships between variables and accurately identifying the true causal edges between features and risks / causes. By quantitatively calculating the causal correlation and contribution weight of each feature to risk events, highly correlated core causal features are selected, and pseudo-correlation, noise, and redundant features with only statistical correlation and no actual business logic are completely eliminated, achieving precise refinement of the causal dimension of risk features.

[0047] Step S1431: Construct a standardized model input feature matrix, perform unified vector quantization on the multimodal clean features obtained after hierarchical decoupling and noise reduction, perform dimension normalization and feature alignment encoding on structured numerical features, text semantic features, and audiovisual visual features with different dimensions and types, and convert all heterogeneous features into standardized multimodal feature vectors with fixed dimensions. Construct the model input feature matrix based on the standardized multimodal feature vectors to complete the standardized construction of input samples for the causal inference network. The standardization of input samples for the causal inference network was completed. Unified vector quantization was applied to the clean multimodal features obtained after hierarchical decoupling and denoising, addressing the issues of inconsistent dimensionality, significant type differences, and the inability to model heterogeneous features uniformly. Dimensional normalization and cross-modal feature alignment encoding were performed on structured numerical index features, textual semantic features, and on-site audiovisual features to eliminate dimensional and dimensional differences between different modalities. All heterogeneous and multidimensional multimodal features were uniformly converted into standardized feature vectors of fixed dimension, uniform specifications, and uniform computation, ensuring that each feature vector completely retains the risk representation information of its corresponding modality. Based on the batch standardized feature vectors, a regular and dense model input feature matrix was constructed, achieving standardized and unified encapsulation of all training and inference samples.

[0048] Step S1432: Define the hierarchical variable system of the causal inference network, set all standardized multimodal feature vectors as observable input variables of the model, set various project risk events such as schedule anomalies, cost anomalies, quality anomalies, and collaboration anomalies as observable output variables, and set the potential influencing factors of personnel, materials, equipment, plans, processes, and management that cause project risks as latent variables, and complete the definition of the variable space of the causal inference model adapted to the project management scenario. The scenario-based definition of the hierarchical variable system of the causal inference network was completed, constructing a variable space adapted to project risk assessment. All normalized and aligned coded standardized multimodal feature vectors were uniformly set as observable input variables for the model, serving as the front-end data foundation for risk assessment. Explicit risk events that can be directly identified during project operation, such as schedule anomalies, cost anomalies, quality anomalies, and coordination anomalies, were uniformly set as observed output variables, corresponding to the dimensions of the model's final risk assessment results. Simultaneously, implicit factors that cannot be directly observed but can induce various risks, including potential triggers such as personnel operation, material quality, equipment status, testing plans, control processes, and on-site management, were uniformly defined as latent variables, covering all dimensions of project risk causes. Through the three-level variable system, the complete transmission logic of project risk from feature input, risk manifestation to implicit causes was accurately matched, constructing a causal modeling foundation adapted to the project scenario.

[0049] Step S1433: Initialize the project risk causal topology network structure. Based on the business relationship logic between observable input variables, observed output variables and latent variables, initialize the project risk scenario causal structure topology graph, construct an initial causal network structure containing feature nodes, risk event nodes and risk cause nodes, and form a multi-dimensional variable relationship topology architecture. The initial construction of the project risk causal topology network is completed. Based on the project business logic between the previously defined observable input variables, observable output variables, and latent causal variables, an initial causal structure topology diagram of project risks is generated. Using multimodal feature nodes as the bottom-level input nodes, various project risk events as the middle-level output nodes, and multiple potential risk triggers as the top-level latent variable nodes, a multi-level, multi-associative initial causal network topology is constructed. Each node establishes preliminary association edges according to real project business rules, forming an initial transmission structure where features trigger risks, triggers dominate deviations, and deviations generate risks. This topology fully covers the association dimensions of project multimodal features, explicit risks, and latent root causes.

[0050] Step S1434: Achieve accurate feature selection and network optimization based on causal structure learning. The causal structure learning algorithm is used to train all multimodal feature samples in batches. Conditional independence tests are performed on the variables of each feature node to verify the direct causal dependency between different multimodal features and risk events one by one, and to eliminate indirect transitive associations and random statistical associations between variables. Leveraging causal structure learning algorithms, network topology optimization and correlation identification are achieved. Batch-by-batch iterative training is performed on all multimodal feature samples in the database, enabling autonomous model learning and logical optimization. During training, systematic conditional independence checks are conducted on each feature node variable, verifying and judging the correlation between each multimodal feature and various risk events. Direct causal dependencies, indirect transmission correlations, and random statistical correlations between variables are accurately distinguished. False correlation edges caused by data transmission and random fluctuations are selectively eliminated, retaining valid correlation structures that conform to business transmission logic. Through batch iterative training, the network topology connection method is continuously corrected, overcoming the mismatch defects of traditional correlation modeling and gradually bringing the network topology structure closer to the actual risk transmission patterns of the project.

[0051] Step S1435: During the training process, discover the true causal edges between feature variables, risk events and risk triggers, calculate the causal association score and marginal contribution of each feature corresponding to the risk event, and generate a quantitative causal association weight matrix. This method achieves quantitative modeling of causal relationships and generation of weight matrices. During the iterative training of causal structure learning, it continuously uncovers the true causal edges between multimodal feature variables, explicit risk events, and implicit risk triggers, constructing a complete causal transmission chain. Through quantitative calculations, it solves for the causal relationship score of each type of multimodal feature on different project risk events, accurately assessing the marginal contribution and influence of each feature in the risk generation process. Based on the quantified relationship scores and contribution ratios, a unified and standardized multidimensional causal relationship weight matrix is ​​generated, clearly recording the corresponding influence weights of different features and triggers on various risks. This weight matrix can intuitively reflect the differences in the causal strength of each element.

[0052] Step S1436: Based on the preset causal correlation threshold, multimodal features are stratified and filtered, retaining core causal features that meet the causal correlation standard and have actual business causal support, and eliminating pseudo-correlation features that only have numerical statistical correlation, no actual business causal logic, collection noise features, and duplicate and redundant features. This completes the causal dimension noise reduction and purification of multimodal features, and outputs a concise, highly reliable multimodal feature set with real causal logic.

[0053] The causal dimension of multimodal features is purified and precisely screened. Based on a preset causal correlation threshold, all multimodal features participating in the modeling are subjected to hierarchical screening and identification. Features with causal correlation scores higher than the set threshold, consistent with project business logic, and possessing a genuine risk transmission mechanism are identified as core causal features, prioritized for retention, and included in the final feature set. Simultaneously, pseudo-correlation features that only exhibit statistical correlation at the data level and lack actual business causal support are batch-removed. Noise features generated by equipment acquisition noise and environmental interference, as well as interfering features with overlapping functions and redundant representations, are also filtered out. Through precise noise reduction and purification of causal dimensions, invalid feature interference is completely eliminated, ultimately outputting a highly concise, reliable, and causally logically sound multimodal core feature set, significantly improving the accuracy of subsequent risk assessment and root cause tracing.

[0054] Step S144: Build a multi-level reverse tracing reasoning link. Based on the completed causal structure and the purified causal characteristics, construct a four-level reverse reasoning architecture including the observation layer, phenomenon layer, inducement layer, and root cause layer. Take the identified abnormal characteristics of project risks as the surface phenomena of the observation layer, and reverse-deduce the influencing factors at each level layer by layer. First, identify the direct abnormal factors on the surface corresponding to the risk, and then dissect the intermediate inducement factors behind the surface abnormalities, such as process deviation, resource deviation, management deviation, and execution deviation. Finally, locate the underlying essential risk root cause corresponding to the system process, personnel collaboration, materials and equipment, solution design, and R&D process operation. A multi-level risk reverse tracing reasoning link is constructed, from the surface to the core. Based on a trained, precise causal topology and refined core causal features, a four-level progressive reasoning architecture is built: observation layer, phenomenon layer, trigger layer, and root cause layer. Multimodal anomaly features identified by the system's dual-threshold mechanism serve as the starting signal for tracing. The four-layer architecture progressively deconstructs the causes of risk, achieving penetrating analysis from surface anomalies to underlying root causes. First, at the observation layer, all original anomaly signals are collected to anchor risk clues. Then, at the phenomenon layer, discrete features are aggregated to condense explicit risk events. Next, at the trigger layer, mid-level deviation triggers such as schedule, cost, resources, execution, and management are deconstructed. Finally, at the root cause layer, the underlying essential causes such as systems, personnel, equipment, plans, and testing are located. The entire process relies on real causal edges for targeted tracing, avoiding misjudgments caused by statistical correlations, and achieving layered, accurate, and explainable root cause localization of project risks.

[0055] Step S1441: Construct a four-level progressive reverse tracing reasoning architecture. Based on the trained risk causal topology and the purified multimodal causal feature set, initialize the four-level reverse reasoning architecture consisting of the observation layer, phenomenon layer, inducement layer, and root cause layer. The four-layer architecture forms a risk reverse reasoning logical link that goes from the surface to the core, decomposes layer by layer, and penetrates level by level. A four-level progressive reverse attribution reasoning architecture was initialized, relying on the completed and converged project risk causal topology and a highly reliable multimodal core feature set after causal purification. A standardized four-layer reverse attribution system was built, consisting of the observation layer, phenomenon layer, trigger layer, and root cause layer from top to bottom. This four-layer architecture forms a clear, logically progressive, and layer-by-layer penetrating reverse attribution chain, following the reasoning logic from the collection of original abnormal signals, the condensation of explicit risks, the decomposition of mid-level deviations, to the location of the root cause at the bottom level. This fully adapts to the transmission law of project risks from manifestation to cause. The architecture as a whole achieves layered decoupling processing of shallow risk signal collection, mid-level analysis, and deep attribution, enabling the progressive refinement of risk influencing factors and avoiding the one-sidedness of analysis caused by single-dimensional attribution. It provides stable architectural support for refined and hierarchical risk root cause tracing.

[0056] Step S1442: The observation layer completes the collection and anchoring of original risk anomaly signals. It inputs various multimodal pure features corresponding to project risk anomalies identified by the system through the dual-threshold risk judgment mechanism into the observation layer. Through the unified collection of time series numerical fluctuation anomalies, text semantic risk features, and audiovisual site violation anomalies, the observation layer completes the summarization, collection and precise anchoring of project risk original anomaly data. The observation layer collects and precisely anchors the original risk anomaly signals of the project. It integrates the multimodal clean features corresponding to various risk anomalies identified and filtered by the system through a dual-layer threshold risk judgment mechanism into the observation layer of the inference architecture. The observation layer centrally collects and integrates all modal original observation signals, including abnormal fluctuations in time-series indicators, textual semantic risk anomalies, and abnormal violations in on-site audiovisual footage, achieving unified aggregation and collection of multi-source anomaly clues. All integrated anomaly features are time-series tagged, scenario-bound, and process-anchored to accurately pinpoint the specific time period, work scenario, and process node where the anomaly occurred, completing the comprehensive collection and precise location of the project's original risk anomaly data. This layer is only responsible for the retention and anchoring of original anomaly signals and does not perform risk attribution judgment.

[0057] Step S1443: The phenomenon layer realizes the aggregation and transformation of discrete abnormal features into explicit risk phenomena. Based on the original multimodal abnormal features collected by the observation layer, the discrete abnormal features are associated, clustered and fused. Combined with the project risk business type, specific risk phenomena such as R&D cycle delays, cost overruns, quality defects, and collaboration conflicts are matched. The noise of simple data fluctuations without business support is stripped away, and a set of surface risk events and intuitive abnormal manifestations that are currently actually occurring in the project is condensed, completing the accurate transformation from scattered abnormal features to complete explicit risk phenomena. By amplifying discrete anomaly features into explicit risk events at the phenomenon layer, the system performs feature association clustering and fusion refinement on the scattered, discrete, multimodal anomaly features collected at the observation layer. This process aggregates and matches various anomaly features from the same source, time period, and process. Combined with project risk business classification standards, the aggregated set of anomaly features is precisely matched to standardized risk phenomena such as R&D cycle delays, cost overruns, quality defects, collaboration conflicts, and resource overruns. During feature aggregation, the system proactively removes random data fluctuation noise without business logic support and filters out invalid and weak interference signals, refining the currently occurring, objectively existing surface-level risk events and intuitive anomalies to form a structured set of risk phenomena. This completes the precise transformation from scattered feature signals to complete business risk phenomena.

[0058] Step S1444: The inducing layer completes the reverse decomposition and screening of intermediate deviation factors. Taking the explicit risk phenomena output by the phenomenon layer as the starting node of reasoning, relying on the feature-deviation association rules built into the causal network, it traverses the real causal edge relationship of the causal topology layer by layer in reverse, deeply decomposes various intermediate deviation factors that cause surface risk phenomena, accurately identifies intermediate inducing factors such as iteration progress control deviation, R&D cost consumption deviation, R&D human / tool ​​resource allocation deviation, R&D execution process deviation, and team collaboration management deviation in the project operation process, quantifies the contribution weight of each intermediate deviation to the surface risk phenomenon, screens high-matching core intermediate inducing factors, and eliminates irrelevant interference deviation factors. By employing a causal layer, the process involves the reverse decomposition and precise screening of mid-level deviation factors. Using the explicit risk phenomena output from the phenomenological layer as the starting point for reverse reasoning, and relying on the built-in features and deviation association rules of the causal network, the process traverses the true causal edge chains in the causal topology layer by layer. It deeply decomposes various intermediate inducing factors that trigger surface-level risk phenomena, accurately identifying multiple types of mid-level deviations during project operation, such as schedule control deviations, cost consumption deviations, material resource allocation deviations, on-site operation execution deviations, and team collaboration management deviations. Through causal weighting, the contribution ratio and impact of various deviations on surface-level risk phenomena are quantified, selecting core mid-level inducing factors with high matching and strong correlation, while eliminating irrelevant interference deviations with weak influence and no direct causal relationship, thus accurately pinpointing the core intermediate issues that trigger risks.

[0059] Step S1445: The root cause layer achieves the fundamental risk tracing and positioning. For the mid-level deviation causes identified, the ultimate attribution matching is carried out by combining the business logic of the entire project process with the risk causal knowledge base. Various mid-level process deviations are accurately mapped to the fundamental root causes such as defects in systems and processes, unreasonable product solution design, non-standard R&D operation processes, substandard configuration of development tools / testing environments, poor personnel collaboration, and lack of version control. A complete causal tracing chain is constructed, which includes multimodal abnormal characteristics, surface risk phenomena, mid-level process deviations, and fundamental root causes. The entire process is based on the real causal edge mining based on causal structure learning and is executed for targeted tracing.

[0060] By establishing a root cause layer, the ultimate source of project risks is located. For various core mid-level deviations identified at the causal layer, precise attribution matching is performed by combining the project's full-process business logic with a risk causal knowledge base. Mid-level deviations such as schedule, cost, resources, execution, and management are mapped one by one to underlying root causes such as incomplete systems and processes, flawed work plan design, non-standard R&D processes, substandard material and equipment quality, poor personnel collaboration, and lack of operation and maintenance control. A complete closed-loop causal traceability chain is constructed, encompassing multimodal anomaly characteristics, surface-level risk phenomena, mid-level process deviations, and underlying root causes. Targeted tracing is conducted throughout the process, relying on real causal edges discovered through causal structure learning, avoiding tracing errors caused by statistical correlation, and achieving precise, penetrating location of project risks from surface anomalies to deep-seated root causes.

[0061] Step S145: In the process of multi-level reverse tracing reasoning, the confidence level, influence weight and contribution ratio of each causal reasoning chain are calculated in real time. The results of multi-level tracing are quantitatively scored and prioritized. The analysis results of risk manifestations, intermediate causes and essential root causes are output in layers. A visual and interpretable causal reasoning path is generated to complete the accurate layer-by-layer tracing from multi-modal abnormal features and risk manifestations to causes.

[0062] This system enables quantitative scoring, prioritization, and visual interpretation of causal attribution results. Throughout the multi-level reverse attribution reasoning process, it calculates the credibility, impact weight, and risk contribution ratio of each causal reasoning chain in real time. It quantifies and ranks all participating feature chains, deviation factors, and underlying root causes. Based on the quantitative results, it differentiates the strength, severity, and probability of different risk factors, automatically prioritizes root causes, and outputs structured analytical results layered from surface-level risk manifestations to mid-level deviation factors and underlying essential root causes. Simultaneously, it generates a visualized and interpretable causal reasoning path map, intuitively displaying the complete path of risk propagation from characteristic anomalies to the final root cause, achieving precise and transparent attribution and analysis of the entire chain from multimodal anomalies to risk manifestations and then to core causes.

[0063] Step S150: Establish a self-iterative optimization closed loop for risk assessment in human-machine collaboration, perform sample classification and labeling and backfeed training on the model assessment results, and continuously iterate and update the model inference weights, feature selection rules and risk judgment parameters in combination with risk scenario sample enhancement strategies. A self-iterative optimization closed-loop system for project risk assessment, based on human-machine collaboration, is established to address the technical shortcomings of traditional risk models, such as fixed parameters, poor scenario adaptability, and inability to autonomously iterate and optimize. Relying on a two-way linkage mechanism between intelligent model assessment and expert review, risk assessment sample resources from real-world project scenarios are continuously accumulated. Each risk assessment output from the model is systematically classified and labeled, and multi-scenario sample augmentation techniques are used to supplement scarce risk samples, improving the scenario coverage and feature diversity of the sample library. Through a hierarchical priority sample backfeeding training mode, targeted optimization iterations are performed on the risk assessment model and causal inference network, dynamically correcting the model's core operating parameters. Multimodal feature selection rules, risk causal inference weights, and dual-threshold risk judgment parameters are continuously and adaptively updated, allowing the model's assessment logic to continuously adapt to dynamic project conditions, new risk scenarios, and changes in business models, achieving a closed-loop iterative capability of unmanned intervention, autonomous learning, and continuous optimization for the risk assessment system.

[0064] Step S151: Construct a human-machine collaborative two-way linkage sample update architecture, build a closed-loop iterative system for risk assessment, retain full-process data for each project risk assessment event, fully record the abnormal feature screening results, causal tracing reasoning paths, risk level judgment results and risk root cause reasoning conclusions automatically output by the model, and simultaneously collect the manual review content corresponding to project managers and industry experts, risk result correction records, model misjudgment reason annotations and root cause correction labels, forming a pairwise comparison sample data with the automatic assessment results of the model and the manual expert review results. A dynamic sample update architecture with two-way human-machine interaction is constructed, establishing a fully traceable, iterative closed-loop system for risk assessment, enabling bidirectional data accumulation and complementary optimization between machine assessment and human verification. The system retains data across all dimensions and throughout the entire process for every project risk assessment event, comprehensively recording the multimodal anomaly feature screening results generated by the automatic model calculation, the complete causal reasoning path, risk level determination conclusions, and underlying root cause reasoning results. Simultaneously, it collects human intervention data such as the content of manual review by project managers and senior industry experts, risk result correction records, model misjudgment cause annotations, and root cause correction labels. Automatic machine assessment data and precise human review data are paired and bound one-to-one, forming paired standard sample data to ensure that each sample includes both machine defect manifestations and human correction basis.

[0065] Step S152: Perform fine-grained classification and labeling of all risk assessment samples accumulated during the closed-loop iteration process, dividing them into samples correctly judged by the model, samples missed by the model, samples misjudged by the model, samples with ambiguous abnormal boundaries, and newly added unknown scenarios. Match corresponding structured labels to each type of sample, complete the labeling and archiving of risk scenario type, abnormal feature combination, judgment parameter deviation, and causal reasoning defect dimensions, and build a dynamically expandable human-machine collaborative iterative sample library. The process of refining and labeling iterative samples and standardizing the database was completed. For the massive risk assessment samples accumulated during the closed-loop iteration process, multi-dimensional refined classification and identification were implemented. Based on the model's assessment performance, samples were uniformly categorized into five types: correctly assessed samples, frequently misjudged samples, missed samples, samples with ambiguous boundaries, and newly added unknown scenarios. A structured labeling system was configured for each type of sample, accurately labeling core information such as the risk scenario type, abnormal feature combination patterns, model judgment parameter bias, and causal reasoning defect dimensions. All labeled samples were uniformly archived and managed, building a human-machine collaborative iterative sample library that supports dynamic incremental expansion, achieving standardized management with clear sample classification, well-defined defects, and full scenario coverage.

[0066] Step S153: Adopt a multi-dimensional scenario-based sample enhancement strategy, supplement various scarce sample scenarios through feature space transformation, modality recombination, scenario migration and temporal fine-tuning, and broaden the coverage dimension of risk samples. This approach employs a multi-dimensional, scenario-based fusion sample augmentation strategy to specifically address industry pain points in the iterative sample library, such as the scarcity of niche risk, extreme risk, and critical boundary risk samples, uneven data distribution, and weak model generalization ability. It abandons traditional single-data augmentation methods, integrating four differentiated sample augmentation techniques to form a diversified augmentation system, comprehensively addressing the shortcomings of various scarce scenario samples. Single-sample multi-dimensional mutation augmentation is achieved through feature space perturbation transformation; feature combination forms are enriched through multimodal feature cross-recombination; gaps in niche scenarios are filled using mature scenario transfer sampling; and massive boundary anomaly samples are generated through temporal trajectory fine-tuning. This multi-method collaborative augmentation effectively enriches the scenario coverage dimensions, feature diversity, and boundary integrity of the sample library, thoroughly improving the sample distribution imbalance problem and significantly enhancing the model's ability to identify and adapt to marginal, novel, and extreme risk scenarios.

[0067] Step S1531: Targeted sample optimization of the target scenario. In response to the shortcomings of insufficient sample quantity, uneven sample distribution, and lack of sample in the human-machine collaborative iterative sample bank for niche risk scenarios, critical anomaly scenarios, and extreme risk scenarios, a multi-dimensional scenario-based sample enhancement and expansion process is launched to specifically make up for the sample data shortcomings of various scarce risk scenarios. By precisely identifying core optimization scenarios for sample enhancement and leveraging the statistical results of existing data in the human-machine collaborative iterative sample library, we accurately pinpoint scenarios where the model's judgment is weak. We focus on targeted sample expansion for niche and less common risk scenarios, borderline ambiguous and abnormal scenarios, historically high-frequency misjudgment scenarios, and extreme and sudden risk scenarios. This addresses the long-standing issues of scarce sample quantity, uneven data distribution, lack of boundary samples, and low model recognition accuracy in these scenarios. We have launched a multi-dimensional, scenario-based sample enhancement process, with high-incidence scenarios of model misjudgment and missed judgment as the core optimization target, and real-world project business mechanisms as constraints, preventing the generation of invalid samples that are detached from actual working conditions. Through targeted scenario enhancement, we accurately compensate for structural data deficiencies in the sample library and specifically address weaknesses in model recognition.

[0068] Step S1532: Implement feature space perturbation transformation expansion, select original multimodal feature samples from scarce risk scenarios and historically high-incidence scenarios, and apply small Gaussian perturbation, data offset fine-tuning and controllable noise superposition processing to the sample numerical features under the premise of conforming to the actual business logic of the project and conforming to the fluctuation range of working conditions, perform multi-dimensional feature mutation expansion on a single original sample, generate compliant and effective mutated feature samples in batches, and expand the basic sample volume of scarce scenarios; The system implements feature space perturbation transformation sample expansion processing, prioritizing original multimodal valid samples from scarce risk scenarios and historically high-risk scenarios as the basic expansion material. It strictly adheres to the actual business logic and operational data fluctuation range of the project, avoiding the generation of invalid abnormal data that violates business rules. Fine-grained micro-Gaussian perturbations, data offset fine-tuning, and controllable noise superposition processing are applied to various numerical features of the samples. While preserving the core risk characteristics and business attributes of the original samples, it achieves multi-dimensional feature variation expansion of a single original sample. It generates compliant, valid, and realistic variant feature samples in batches, rapidly expanding the basic sample volume of scarce scenarios, enriching the feature fluctuation patterns under similar risk scenarios, solving the problems of insufficient sample base and inadequate model learning in niche scenarios, and strengthening the model's ability to capture similar weak abnormal features.

[0069] Step S1533: Conduct multimodal heterogeneous feature cross-recombination, extract numerical indicator features, text record features, on-site audiovisual features, and time series fluctuation features corresponding to project risks, and perform random and effective cross-combination and flexible pairing of various heterogeneous modal features; This approach expands and recombines multimodal heterogeneous features, comprehensively extracting four core categories of features corresponding to project risks: numerical quantitative features, semantic features of project text records, on-site audiovisual features, and process temporal fluctuation features. Breaking away from the traditional fixed modal pairing and single feature combination model, it randomly, compliantly, and effectively cross-combines and flexibly pairs and reconstructs various heterogeneous modal features. While ensuring that feature combinations conform to the project's business logic, it generates a large number of new multimodal feature combination samples, simulating the complex risk patterns of simultaneous occurrence and cross-coupling of multi-dimensional anomalies during actual project operation. This effectively enriches the diversity of risk feature combinations, solves the problems of fixed single-modal features and insufficient samples of complex coupled risk risks, and improves the model's ability to identify multi-feature linked anomaly risks.

[0070] Step S1534: Perform similar scenario migration sampling expansion, mine and extract general data feature patterns and operating condition fluctuation patterns from mature risk scenarios with sufficient sample stock, and based on the common features of scenarios with the same type of projects, the same testing process, and the same management logic, migrate and adapt the effective feature patterns of mature scenarios to niche and scarce risk scenarios to generate new scarce scenario samples that fit the real business mechanism. A similar scenario migration sampling expansion strategy is implemented to fully leverage the abundant and well-established data resources of conventional risk scenarios in the sample library. This involves deeply extracting the common data change patterns, operational fluctuation modes, and risk characteristic representation logic of mature scenarios. Based on the common characteristics of scenarios with consistent project types, similar testing processes, shared management logic, and common risk mechanisms, effective feature patterns from mature scenarios are intelligently migrated and adapted to niche, novel, and scarce risk scenarios. Adaptation adjustments are made based on the business characteristics of scarce scenarios to generate new scarce scenario samples that fully align with real business mechanisms and scenario operating rules. This quickly fills the sample data gaps for niche and novel risks, resolving the model recognition blind spots caused by a lack of or weak samples in new scenarios.

[0071] Step S1535: Conduct refined trajectory optimization processing for time-series risk samples, adaptively fine-tune the fluctuation rhythm, numerical range, and abrupt change inflection point of the original indicator time-series fluctuation curve, simulate the real dynamic working condition fluctuation state of the entire process of project version iteration, R&D, and operation and maintenance, and generate diversified time-series anomaly boundary samples in batches. This system employs refined trajectory fine-tuning and expansion processing for time-series risk samples, addressing the pain points of complex project time-series operating condition fluctuations and difficulty in identifying boundary anomalies. Adaptive and refined adjustments are made to the time-series fluctuation curves of various original indicators, including multi-dimensional optimization operations such as fine-tuning of fluctuation rhythm, numerical range offsetting, and reconstruction of abrupt inflection points. It realistically simulates dynamic operating condition changes, phased fluctuations, and sudden small deviations throughout the entire project testing, R&D iteration, and operation and maintenance management process. It generates batches of diverse, richly defined, and realistically aligned time-series anomaly boundary samples, significantly increasing the number of samples in critical ambiguity scenarios. This addresses the problem of model inconsistencies and frequent critical misjudgments caused by a lack of boundary samples, improving the model's accuracy in distinguishing time-series boundary anomaly risks.

[0072] Step S1536: Through a diversified sample enhancement strategy that integrates multiple methods, the number of samples in scarce, marginal, and extreme risk scenarios is expanded in all aspects, enriching the scenario coverage dimensions and feature diversity of the sample library.

[0073] Leveraging a diversified sample enhancement system that integrates multiple methods, the system achieves comprehensive sample expansion and quality upgrades. It integrates four enhancement techniques—feature perturbation transformation, multimodal cross-recombination, similar scene migration sampling, and temporal trajectory fine-tuning—to achieve complementary and comprehensive sample expansion. This comprehensively increases the number of samples in scenarios with scarce risks, marginal risks, extreme emergencies, and critical ambiguity risks, thoroughly addressing the structural deficiencies of uneven data distribution, incomplete scene coverage, and missing boundary samples in the sample library. It significantly enriches the scene coverage dimensions, feature combination forms, and temporal fluctuation diversity of the sample library, enabling it to fully cover normal operating conditions, abnormal operating conditions, critical operating conditions, and novel operating conditions.

[0074] Step S154: Establish a differentiated hierarchical priority sample backfeeding iterative training mechanism, set training priorities for enhanced samples of different defect types, implement targeted optimization of risk judgment model and causal inference network, and modify feature screening logic, supplement causal relationship constraints, and optimize risk judgment threshold parameters through classified samples. A differentiated, hierarchical, and priority-based sample backfeeding iterative training mechanism is established, abandoning the traditional, indiscriminate, and uniform mixed training approach. Instead, a problem-oriented, targeted, and refined training logic is adopted. For high-quality defect samples that have undergone classification, cleaning, and scenario enhancement, different training priorities are assigned based on the differences in model optimization needs, constructing a gradient-based training system. Backfeeding channels are built for both the two-layer threshold risk assessment system and the multimodal causal inference network, enabling parallel targeted optimization of both models. For misjudged samples, missed samples, and samples with ambiguous boundaries, corresponding model optimization objectives are matched, precisely correcting the feature selection logic, supplementing missing causal relationship constraints, and optimizing critical judgment threshold parameters. This achieves targeted repair of different model defects, comprehensively improving the accuracy of risk assessment and the reliability of causal attribution.

[0075] Step S1541: Construct a defect-oriented hierarchical priority queue for samples. Based on the sample library annotation system, classify and grade high-quality iterative samples after classification, cleaning and scene enhancement. According to the degree of model optimization necessity, divide the training priority from high to low. Set historical high-frequency misjudgment samples as first-level optimization samples, extreme risk missed judgment samples as second-level optimization samples, and critical boundary ambiguous samples as third-level optimization samples. Construct a hierarchical and graded differentiated sample iterative training queue. A defect-oriented hierarchical priority training queue is constructed. Based on a standardized defect annotation system in the sample library, all high-quality iterative samples that have completed classification, cleaning, and scene enhancement are graded according to their defect dimensions. Combining the statistical results of historical model judgment errors, the training priority queues are divided into three levels according to the urgency and necessity of model optimization, from highest to lowest. High-frequency, repeatedly misjudged samples are set as the highest priority level (Level 1) samples, prioritizing the resolution of core model judgment errors; samples with extreme risk omissions are set as Level 2 priority samples, supplementing the ability to identify hidden risks; and samples with ambiguous critical boundaries are set as Level 3 priority samples, optimizing boundary judgment accuracy. Through this hierarchical queue arrangement, training resources are precisely allocated, forming a differentiated training arrangement mechanism with clear priorities and targeted repair.

[0076] Step S1542: Build a dual-system targeted backfeeding training channel. Construct independent sample backfeeding adaptation channels for the dual-layer threshold risk judgment system and the multimodal causal inference network respectively. Complete sample loading and iterative training in batches according to sample priority. Prioritize the correction of model weights through high-priority defect samples, and then rely on secondary priority samples to achieve fine-tuning of boundary parameters and enhancement of model generalization ability, forming a gradient-type targeted training mechanism. A dual-system independent targeted backfeeding training channel was established, with separate data backfeeding and training links constructed for the two-layer threshold risk judgment system and the multimodal causal inference network, enabling parallel, independent, and accurate iterative optimization of the two types of models. Training samples were loaded in batches according to a strict pre-set priority, prioritizing the use of first-level high-priority misjudged samples to correct and iterate the core weights and rules of the model, quickly resolving core model defects. Then, second-level missed samples and third-level boundary samples were used to enhance the implicit risk identification capability and fine-tune the boundary parameters, forming a gradient-based targeted training mechanism of first correcting errors, then filling blind spots, and finally fine-tuning. This effectively avoids the shortcomings of traditional unified training, such as incomplete core problem repair and inadequate boundary optimization, significantly improving model iteration efficiency and optimization accuracy.

[0077] Step S1543: Based on the first-level misjudged sample correction feature screening and causal weight, high-frequency misjudged samples are input into the network in batches to perform reverse gradient update, recalibrate the effective screening threshold of various multimodal features, optimize the causal association weight matrix between features and risk events, weaken the weight of pseudo-association features, and strengthen the contribution ratio of real causal features. Based on high-frequency misjudgment samples from the first level, feature selection logic and causal weight correction are completed, specifically addressing high-frequency misjudgment issues caused by feature selection bias, distorted causal correlation weights, and feature mismatch. A batch of high-frequency misjudgment samples are input into the causal inference network and risk assessment model, and inverse gradient update iterative training is performed. The selection thresholds and effective judgment intervals for various multimodal features are recalibrated, and the causal correlation weight matrix between features and various risk events is reconstructed. The weights of pseudo-correlation features without business basis and invalid noise features are weakened, while the contribution ratio of true causal features and core risk features is significantly strengthened. From the bottom layer of feature selection and causal matching, the model's judgment defects are corrected, completely resolving the high-frequency misjudgment problem caused by feature mismatch and weight distortion.

[0078] Step S1544: Based on the second-level missed samples, supplement the causal constraints and risk identification logic. Using the abnormal feature combination and real cause chain of the extreme risk missed samples as the training basis, add effective risk feature causal association edges in the causal topology network and supplement the causal constraints corresponding to various niche and new risk scenarios. This paper addresses the shortcomings of the model in identifying niche, novel, and weak-signal risks by supplementing the causal constraints and implicit risk identification logic of the model with missed samples of secondary extreme risks. Optimization is carried out based on these missed samples, using the abnormal feature combinations, real risk transmission chains, and underlying causal relationships as the core training basis. Effective feature causal association edges corresponding to niche and novel risks are added to the original causal topology network structure. The causal constraints and risk transmission rules of various scarce scenarios are supplemented, improving the model's logic for capturing implicit risks, weak anomalies, and new operating condition risks. This fills the risk identification blind spots of traditional models and significantly enhances the model's ability to identify and cover extreme, niche, and newly added risk scenarios.

[0079] Step S1545: Optimize the critical parameters for risk judgment based on three-level boundary fuzzy samples, focus on the accuracy optimization requirements for critical scenarios, use massive critical abnormal samples to carry out iterative training, fine-tune the risk judgment critical parameters of the fluctuation tolerance range of the statistical dynamic threshold of each indicator and the rigid threshold of the business, and converge the judgment error of critical working conditions. By leveraging three levels of boundary fuzzy samples to optimize the critical judgment parameters of the model, this study focuses on addressing the issues of inconsistent and large errors in model judgments under critical and fuzzy conditions in projects. It utilizes a massive amount of enhanced critical anomalies and boundary fuzzy samples for iterative training, continuously fine-tuning the tolerance range of statistical dynamic threshold fluctuations for each monitoring indicator, and optimizing the critical judgment node parameters for rigid business thresholds. Through multiple rounds of boundary sample iterations, the model converges boundary judgment errors, correcting the judgment bias between normal fluctuations and minor anomalies, and between minor anomalies and true risks. This resolves the problems of misjudgment, missed judgment, and inconsistent judgments under critical conditions, significantly improving the model's accuracy and stability in distinguishing fuzzy boundary scenarios.

[0080] Step S1546: After each round of hierarchical and targeted backfeeding training is completed, the multimodal feature screening threshold, global causal association weight matrix, four-level reverse tracing inference path weight, dynamic fluctuation threshold parameter and rigid compliance judgment parameter are updated synchronously to complete the collaborative iterative optimization of the entire link of risk feature screening, causal tracing inference and double-layer threshold verification.

[0081] The entire model's parameters are updated iteratively and collaboratively across the entire chain. After each round of hierarchical and targeted backfeeding training, the core operating parameters of the entire model are updated in batches simultaneously. The multimodal feature selection threshold standard, global causal association weight matrix, and weight parameters of the four-level reverse tracing inference path are uniformly corrected, and the dynamic threshold fluctuation range and rigid compliance judgment critical parameters are iterated synchronously. The entire chain mechanism of front-end feature selection, mid-level causal inference, and back-end two-layer threshold verification is optimized collaboratively, ensuring that the model's feature extraction logic, risk causal judgment logic, and risk classification judgment logic are iterated and upgraded synchronously. This allows the model to adapt to the optimized sample features and scenario patterns, comprehensively reducing model misjudgments, omissions, and tracing biases, and achieving an overall quality improvement and upgrade of the risk judgment system.

[0082] Step S155: Construct a fully automated closed-loop self-iterative optimization system. Through continuous loop execution of automatic risk assessment, manual verification and correction, sample structured annotation, scenario-based data augmentation, hierarchical backfeeding training, and model parameter iterative update process, a complete human-machine collaborative iterative closed loop is formed, so that risk assessment parameters, feature selection rules, and causal inference weights can continuously adapt to new working conditions, new risks, and dynamic business changes in the project.

[0083] A fully automated, closed-loop, self-iterative optimization system is constructed, forming a self-evolving human-machine collaborative risk assessment loop mechanism. This system continuously executes six core processes: automatic model risk assessment, human expert verification and correction, structured sample annotation, scenario-based data augmentation, hierarchical priority retraining, and iterative updates of all model parameters, forming a seamless, self-driven, closed-loop optimization system. As project conditions continuously update, business models iterate, and new risks emerge, the system leverages this closed-loop mechanism to continuously accumulate new scenario samples and optimize model judgment rules and causal weights. This allows risk assessment parameters, feature selection logic, and causal reasoning capabilities to continuously adapt to dynamic project changes, enabling the model to learn, correct errors, and upgrade autonomously without human intervention, maintaining the accuracy and scenario adaptability of risk assessment in the long term.

[0084] Step S160: Employ a multimodal feature cross-validation and voting fusion judgment mechanism, using single-modal anomalies as suspected clues and multimodal homologous feature joint verification as the judgment basis, and identify data noise and real risks through multi-dimensional feature consistency constraints.

[0085] A highly reliable risk assessment mechanism integrating multimodal feature cross-validation and voting is constructed to address the shortcomings of traditional single-modal assessment, such as strong bias, high false alarm rate, and difficulty in distinguishing between genuine and false risks. Single-modal independent anomaly capture is used as a preliminary screening method for suspected clues, ensuring full retention of weak risk signals and preventing the omission of hidden risks. Joint cross-validation of homogeneous multimodal features serves as the core judgment criterion, relying on multi-dimensional feature consistency constraints based on spatiotemporal, scenario, and subject homogeneity to distinguish between random data noise and genuine business risks. Multimodal voting fusion judgment logic enables complementary verification of multi-source information, effectively avoiding false anomalies caused by single data fluctuations and single-point collection errors. This significantly improves the anti-interference capability and accuracy of project risk assessment, achieving intelligent risk identification with low false alarms and high reliability.

[0086] Step S161: Conduct initial screening and capture of independent anomalies in multimodal features. For various single-modal features of numerical index modes, text record modes, on-site audiovisual modes, and time-series fluctuation modes collected in real time by the project, conduct independent anomaly screening according to the threshold rules and feature discrimination criteria of each mode. Signals that deviate from the normal operating range or exhibit abnormal features in any single mode are uniformly marked as suspected risk clues, and all single-modal anomaly signs are fully retained. A comprehensive initial screening of independent anomalies across multiple modalities was completed, ensuring no potential risk clues were missed. For the four core modalities of data collected in real-time from the project—numerical indicator time-series modality, project text record modality, on-site audiovisual multimedia modality, and operating condition time-series fluctuation modality—dedicated threshold rules and anomaly identification criteria were configured for each. Independent and parallel anomaly screening and analysis were conducted for each modality, without relying on other modalities or prematurely filtering weak anomaly signals. Any single modal indicator deviating from the normal operating range or exhibiting abnormal characteristics was immediately marked as a potential risk clue and fully retained. This comprehensive capture of all subtle, hidden, and early-stage risk indicators prevented weak risk signals from being filtered by the system or covered by the overall mean, providing a comprehensive and complete reserve of potential clues for subsequent multimodal cross-validation.

[0087] Step S162: Construct a multimodal homogeneous feature cross-validation mechanism. For the multimodal single-modal suspected abnormal clues obtained from the initial screening, carry out spatiotemporal alignment and homogeneous association matching of multimodal features of the same process, the same time period, and the same risk object. Collect all modal feature data belonging to the same risk event and remove invalid interference features across scenarios and time periods. A precise cross-validation mechanism for multimodal homogeneous features was established to conduct precise homogeneous association calibration and invalid data removal for the massive amount of suspected anomaly clues obtained from the initial screening. Using the same process, time period, and work object as homogeneous constraints, the dispersed multimodal anomaly clues were spatiotemporally aligned, scenario-matched, and subject-associated. Various modal anomaly features belonging to the same risk event, the same work scenario, and the same operating period were aggregated and integrated to form a homogeneous multimodal feature set. Simultaneously, irrelevant interfering data from different sources across scenarios, time periods, and processes were accurately identified and removed, and invalid anomaly clues caused by misplaced data collection or mixed scenarios were eliminated. This ensures that all features participating in the fusion analysis originate from the same real risk event, guaranteeing the rigor and accuracy of cross-validation.

[0088] Step S163: Establish a multimodal voting fusion consistency judgment logic, set the independent anomaly judgment results of each modality as independent voting units, preset multimodal feature consistency constraint judgment rules, and each voting unit outputs three types of judgment results: normal, slightly abnormal, and significantly abnormal. Calculate the voting ratio of each type of judgment result and the consistency degree of multimodal feature matching. Judgment results with highly consistent multimodal feature performance and pointing to the same risk type are judged as real and valid risks. Abnormal signals with contradictory judgment results of each modality, only a single modality being abnormal, and the other modalities being in the normal operating range are judged as false anomalies caused by equipment acquisition jitter, data transmission error, or local noise. A multimodal voting fusion consistency assessment logic is established to achieve quantitative identification and judgment of risk authenticity. The independent assessment results of each modality are set as equal and independent voting units, with pre-defined standardized multimodal consistency constraint judgment rules. Each voting unit outputs three judgment results: normal, slightly abnormal, and significantly abnormal. The system automatically calculates the voting percentage of each type of result and the consistency matching degree of multimodal features. When multiple modal features are highly consistent and point to the same risk type, it is judged as a real and valid risk event; when only a single modality is abnormal and all other modalities are normal, or when the judgment results of each modality are contradictory, it is judged as a false anomaly caused by acquisition jitter, transmission error, or local noise, thus distinguishing real risks from random noise interference at the mechanism level.

[0089] Step S164: Relying on the multimodal cross-validation and voting fusion discrimination system, overcome the one-sidedness of traditional single-modal single-risk judgment, and accurately distinguish between noisy data, pseudo-anomaly signals and real project risks through multi-dimensional feature consistency constraints. It retains real risk events jointly corroborated by multimodal features, filters out isolated, sporadic, and noise anomalies without supporting features from the same source, and outputs accurate judgment results after multi-source information fusion verification.

[0090] This system achieves highly reliable and accurate risk assessment through multimodal fusion, completely overcoming the shortcomings of traditional single-modal, one-time judgments, such as incompleteness, poor stability, and high false alarm rates. Relying on a multimodal cross-validation and voting fusion discrimination system, and using the consistency of multi-dimensional features as the core identification criterion, it accurately distinguishes between false anomalies caused by equipment noise, data jitter, and transmission errors, and genuine project risks. The system proactively filters isolated, sporadic, and false anomaly signals lacking multi-source feature support, fully preserving risk events corroborated by multimodal features and possessing genuine business mechanisms. Finally, it outputs accurate assessment results after multi-source information fusion and multi-layered verification and filtering, achieving intelligent risk assessment with low false alarms, low false negatives, and high reliability, comprehensively improving the stability and reliability of risk assessment in complex project scenarios.

[0091] Based on the same inventive concept, please refer to Figure 2 This document shows a schematic block diagram of a big data-based multimodal enterprise project management system 100 provided in this application embodiment for executing the above-described big data-based multimodal enterprise project management method. The big data-based multimodal enterprise project management system 100 may include a communication unit 110, a machine-readable storage medium 120, and a processor 130.

[0092] In this embodiment, both the machine-readable storage medium 120 and the processor 130 are located within the big data multimodal enterprise project management system 100 and are separately configured. However, it should be understood that the machine-readable storage medium 120 may also be independent of the big data multimodal enterprise project management system 100 and may be accessed by the processor 130 via a bus interface. Alternatively, the machine-readable storage medium 120 may also be integrated into the processor 130 and may communicate and interact with external systems through the communication unit 110.

[0093] The processor 130 is the control center of the big data-based multimodal enterprise project management system 100. It connects various parts of the system via various interfaces and lines, and executes software programs and / or modules stored in the machine-readable storage medium 120, as well as calling data stored in the medium, to perform various functions and process data, thereby providing overall monitoring of the system. Optionally, the processor 130 may include one or more processing cores; for example, it may integrate an application processor and a modem processor, where the application processor primarily handles the operating system, user interface, and applications, and the modem processor primarily handles wireless communication. It is understood that the modem processor may not be integrated into the processor. The machine-readable storage medium 120 stores machine-executable instructions for implementing the scheme of this application, and the processor 130 executes the machine-executable instructions stored in the machine-readable storage medium 120 to implement the big data-based multimodal enterprise project management method provided in the aforementioned method embodiments.

[0094] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

[0095] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A multimodal enterprise project management method based on big data, characterized in that: Includes the following steps: Modal hierarchical decoupled differential feature mining is implemented for multimodal heterogeneous data of projects. Feature extraction strategies are adopted for three types of modalities: structured time series data, text data, and audiovisual multimedia data. Project risk correlation features corresponding to each modality are independently mined. Intramodal noise reduction and filtering are performed on the original features of each modality to remove redundant noise and invalid fluctuation features, so as to obtain a set of pure multimodal features without cross interference. Establish a cross-modal data accurate alignment mechanism, using the unique process identifier of the project as the binding primary key, and perform unified association and calibration of multiple types of heterogeneous data under the same project task in terms of subject, scenario and time dimensions. Use time-series correction and spatial matching algorithms to remove irrelevant data from different sources. A two-tiered risk assessment mechanism is constructed that integrates statistical dynamic thresholds and business rigid thresholds. Based on the distribution characteristics of historical data, fluctuation thresholds that adapt to the dynamic changes of projects are generated, and fixed business thresholds are configured in combination with industry and institutional standards. Through the joint constraint of dual conditions, the accurate distinction between real risks and normal data fluctuations is achieved. Construct a causal knowledge system for project risks and introduce a causal reasoning network. Establish characteristic causal relationships based on historical risk cases. Eliminate false correlation features through causal structure learning. Combine reverse tracing reasoning to achieve hierarchical analysis of risk manifestations and essential causes. Establish a self-iterative optimization closed loop for risk assessment in human-machine collaboration, perform sample classification and labeling and backfeed training on the model assessment results, and continuously iterate and update the model inference weights, feature selection rules and risk judgment parameters by combining risk scenario sample enhancement strategies. A multimodal feature cross-validation and voting fusion judgment mechanism is adopted. Single-modal anomalies are used as suspected clues and multimodal homologous feature joint verification is used as the judgment basis. Data noise and real risks are identified through multi-dimensional feature consistency constraints.

2. The enterprise project management method based on big data and multimodal approaches according to claim 1, characterized in that: The establishment of a cross-modal data accurate alignment mechanism uses the unique process identifier of the project as the binding primary key to uniformly associate and calibrate multiple heterogeneous data under the same project task in terms of subject, scene, and time dimensions. It also removes irrelevant heterogeneous data through time-series correction and spatial matching algorithms, including: For multimodal heterogeneous data sources such as time-series data, text data, image and video data, operation log data, and cost ledger data connected to the project management system, a globally unique data sequence number is uniformly assigned. The collection timestamp, equipment number, collection point, and operator fields corresponding to each piece of raw data are parsed synchronously to construct a multimodal raw data metadata structure. At the data access layer, the raw data is cleaned for field integrity, and dirty data with missing primary keys, missing time information, and missing scene location information are removed to filter out valid raw multimodal data with basic attributes of association and matching. Pre-configure a unique process ID for the project as a global binding primary key, build a process primary key index mapping table, establish a one-to-one correspondence between each R&D iteration process, version development node, and operation and maintenance task in the entire project lifecycle and the unique process ID, and mount all valid original multimodal data to the corresponding process ID node according to the business affiliation. The original timestamps of all multimodal data mounted under the same process ID are processed for precision standardization and uniformly converted into millisecond-level standard time format. Dynamic time warping algorithm is used to correct the timing offset and acquisition delay of different devices. The time offset of various modal data relative to the standard time axis of the process is calculated. Based on the obtained offset, the video frame acquisition time, voice recording time, log filling time and sensor sampling time are generally shifted and calibrated. Extract the spatial coordinates of each multimodal data collection, the location of R&D workstation equipment, the R&D area number, and the R&D scene label to construct a project R&D spatial scene topology matrix. Grid-encode the project R&D office area, R&D workstation, review scene, and maintenance point. Use a scene space matching algorithm to compare and match the spatial location information of the multimodal data with the project's standard scene grid. Filter and remove cross-scene, cross-process, and non-homogeneous heterogeneous data with spatial coordinate deviations exceeding preset thresholds, scene coding mismatches, or inconsistent R&D office areas. Retain valid data that completely matches the current process scene. Bind the main information of the person in charge, tester, R&D personnel, operation and maintenance role, and approver corresponding to the current process, automatically match the operator, recorder, and reviewer fields in multimodal data, verify the consistency between the data behavior subject and the project task subject, and filter out external interference data and irrelevant operation data that do not match the subject; Based on the calibrated unified timeline, standardized scene grid, verified task subjects and corresponding process events, the integrated binding and data fusion of the four dimensions of time, space, subject and event are completed to generate a multimodal aligned dataset with the same source, scene, subject and time period.

3. The enterprise project management method based on big data and multimodal approaches according to claim 1, characterized in that: The proposed dual-layer risk assessment mechanism, which integrates statistical dynamic thresholds and rigid business thresholds, generates fluctuation thresholds adapted to the dynamic changes of projects based on historical data distribution characteristics, and configures fixed business thresholds in conjunction with industry and regulatory standards. Through joint constraints under two conditions, it accurately distinguishes between real risks and normal data fluctuations, including: The project's full historical operation dataset is constructed by collecting project progress time-series data, cost consumption data, resource input data, and quality inspection data covering different project types, different iterative development stages, and different R&D cycles. The historical operation dataset is preprocessed to remove manually entered abnormal data, system fault data, and marked fault sample data. Valid sample data under normal project operation is selected and retained to construct the project's normal operating condition benchmark dataset. Based on the benchmark dataset of normal operating conditions of the project, the data distribution characteristics of risk indicators of each project are fitted. The mean and standard deviation of each risk indicator are calculated by the Gaussian normal distribution algorithm. The moving range analysis algorithm is combined to perform dynamic interval extrapolation of the continuous fluctuation data of time series. An adaptive sliding time window is configured according to the differences of project stages. The extreme value range of fluctuation of each indicator is statistically analyzed window by window. The upper and lower limits of normal fluctuation of indicators are generated dynamically and adaptively changing with the project R&D iteration progress, version development cycle, and operation and maintenance stage. A statistical dynamic threshold adapted to the dynamic operating characteristics of the project is constructed. Based on national industry standards, enterprise project management systems, project contract constraints, and project-specific technical specifications, we sort out the rigid constraints and hard control boundaries corresponding to each risk indicator. For various risk types such as cost overruns, R&D cycle delays, quality defects, resource overruns, and collaboration timeouts, we configure fixed and unchanging business rigid thresholds, build a structured business rule threshold library, and bind and map each business rigid threshold with corresponding project indicators, risk types, and process scenarios to establish a rule matching relationship that can be called in real time. A two-tiered progressive risk verification mechanism combining dynamic fluctuation tolerance and rigid rule constraints is established. The statistical dynamic threshold defines the tolerance range of normal data fluctuations in the project, while the business rigid threshold defines the boundary for judging project violation risks. The multimodal indicator data of the project collected in real time is jointly verified by the two thresholds. Based on the different levels of the indicator values, which are in the dynamic threshold range and the rigid threshold boundary, normal operating condition fluctuations, slight abnormal fluctuations and real violation risks are distinguished, so as to achieve graded judgment, differentiated processing and accurate risk identification. A dual-threshold dynamic iterative update mechanism is constructed. Based on real-time project operation data and manually verified samples, the distribution characteristics of the benchmark data of the working conditions are continuously optimized iteratively. The dynamic statistical threshold is periodically and adaptively updated to adapt the fluctuation tolerance range to the real-time changes of the project's working conditions. At the same time, the rigid threshold rule parameters of the business are solidified to keep the compliance judgment boundary constant. The mechanism relies on the collaborative judgment logic of dynamic adaptive constraints and fixed rigid constraints.

4. The enterprise project management method based on big data and multimodal approaches according to claim 3, characterized in that: The proposed dual-layer progressive risk verification mechanism, combining dynamic fluctuation tolerance with rigid rule constraints, uses statistical dynamic thresholds to define the tolerance range of normal data fluctuations in a project, and business rigid thresholds to define the boundary for determining project violation risks. It performs dual-threshold hierarchical joint verification on real-time collected multimodal indicator data of the project. Based on the different levels of indicator values ​​within the dynamic threshold range and rigid threshold boundary, it distinguishes between normal operating condition fluctuations, minor abnormal fluctuations, and genuine violation risks, achieving graded judgment, differentiated processing, and accurate risk identification, including: A two-layer nested risk verification architecture is constructed, which hierarchically binds the pre-built statistical dynamic thresholds with the rigid business thresholds. The statistical dynamic thresholds serve as the first layer of flexible fluctuation filtering boundaries, which can adapt to and accommodate normal operating data fluctuations in different iterative development stages, R&D cycles, and operation and maintenance stages of the project. The rigid business thresholds serve as the second layer of compliance risk judgment red lines, which define the compliance control standards that the project cannot cross, thus forming a two-layer risk judgment system in which flexible adaptive constraints and rigid compliance constraints are nested together. The first-level dynamic threshold progressive verification is performed. The multimodal standardized indicator data of the project after synchronization and alignment are accessed in real time. The single real-time indicator data is sequentially input into the two-level verification link for progressive judgment. The real-time indicator value is first checked to see if it is within the upper and lower limits of the statistical dynamic threshold corresponding to the current project stage. If the real-time indicator value falls within the dynamic fluctuation range, the current project process is determined to be a normal working condition fluctuation. Only the data is retained and recorded, and no abnormal marking and risk warning operations are performed. Triggering the second-level business rigid threshold verification, if the real-time indicator value exceeds the normal fluctuation range of the statistical dynamic threshold, it is determined that the indicator has an abnormal deviation from the working condition, and automatically enters the second-level business rigid threshold verification process. The current real-time indicator value is matched and compared with the business rigid threshold bound to the corresponding risk type and process scenario to verify whether the indicator value reaches or exceeds the compliance control red line. If an indicator exceeds the statistical dynamic threshold range but does not reach the business rigid threshold standard, it is judged as a slight abnormal fluctuation state. This abnormal feature is marked and archived for subsequent model sample iteration and working condition data update only. It does not trigger the formal project risk warning and root cause analysis process. If the indicator value simultaneously exceeds the normal fluctuation tolerance range of the current stage statistical dynamic threshold and meets or exceeds the business rigid threshold constraint conditions of the corresponding risk scenario, it is judged as a real and valid project violation risk. The current process node, risk type, abnormality degree, and multimodal abnormal features are jointly marked to lock the risk event and automatically trigger the subsequent multimodal causal reasoning root cause location process, thus completing the hierarchical identification and graded handling of normal working condition fluctuations, slight abnormal fluctuations, and real violation risks.

5. The enterprise project management system and method based on big data and multimodal operation as described in claim 3, characterized in that: The aforementioned dual-threshold dynamic iterative update mechanism continuously optimizes the distribution characteristics of benchmark data based on real-time project operation data and manually verified samples, periodically and adaptively updating dynamic statistical thresholds to adapt the fluctuation tolerance range to real-time changes in project operating conditions. Simultaneously, it solidifies rigid threshold rule parameters to maintain a constant compliance judgment boundary. This mechanism relies on the collaborative judgment logic of dynamic adaptive constraints and fixed rigid constraints, including: Establish an incremental update mechanism for benchmark data of operating conditions, continuously collect real-time multimodal operation data, system risk assessment results data, and manually reviewed and labeled sample data throughout the entire project process, and integrate and summarize the newly added real-time data with the historical normal operating condition benchmark dataset to construct an incremental update dataset; A unified cleaning and filtering process is carried out on the incremental update dataset to remove abnormal data caused by instantaneous equipment jitter, temporary misoperation data, and invalid duplicate data, while retaining valid operating condition data that can characterize the latest operating status of the project. Configure a dual-mode adaptive threshold iteration update strategy, preset a fixed-cycle iteration mechanism, automatically start the dynamic threshold update process every set fixed cycle, and add an adaptive trigger iteration mechanism for working condition fluctuations to monitor the time-series fluctuation status of project indicators in real time. When the variance of indicator fluctuations within the continuous sliding window exceeds the preset fluctuation change threshold, the emergency iteration update process is actively triggered. Based on the updated normal operating condition benchmark dataset, the Gaussian distribution characteristics and time series variation patterns of each risk indicator are refitted. Combining the moving range analysis algorithm and the adaptive sliding time window, the upper and lower limits of normal fluctuations of the indicators corresponding to each project stage are recalculated, and the adaptive correction of the statistical dynamic threshold is completed, so that the dynamic fluctuation tolerance range can be adapted in real time to the dynamic change characteristics of project version iteration progress, R&D iteration, and operation and maintenance conditions. Throughout the entire process of statistical dynamic threshold iteration and update, a parameter anti-tampering protection mechanism is implemented for the rule database of rigid thresholds for storage business. By using permission encryption locking, rule read-only configuration, and parameter consistency verification, all rigid threshold parameters for business are solidified and constrained, restricting the addition, modification, and deletion of rigid thresholds by automated iteration programs and manual operations. This ensures that the compliance judgment boundary established based on industry standards, corporate management systems, project contract constraints, and special technical specifications remains constant and does not change with fluctuations and iteration updates of project operating data. After each iteration and correction of the dynamic statistical threshold, the linkage verification mapping relationship between the updated dynamic statistical threshold and the fixed business rigid threshold is reconstructed, and the matching parameters of the two-layer progressive risk verification architecture are updated simultaneously to form a judgment system that works in conjunction with dynamic adaptive flexible fluctuation constraints and long-term constant rigid compliance constraints.

6. The enterprise project management method based on big data and multimodal approaches according to claim 1, characterized in that: The project constructs a causal knowledge system for project risks and introduces a causal reasoning network. Based on historical risk cases, it establishes characteristic causal relationships, eliminates false associations through causal structure learning, and combines reverse attribution reasoning to achieve a hierarchical analysis of risk manifestations and underlying causes. This includes: A standardized project risk causal case knowledge base is constructed, which collects historical project risk cases, risk rectification records, anomaly ledgers and multimodal risk feature samples covering multiple scenarios such as R&D cycle delays, cost overruns, quality defects, resource overruns, and collaboration timeouts. All risk cases are structured and labeled, and each case is bound to the corresponding multimodal original features, risk manifestations, root cause types and rectification and handling results to form a project risk original case library. The original case database data is cleaned and normalized to remove invalid cases with incomplete records, ambiguous causal relationships, and untraceable sources. High-quality risk samples with complete features, clear root causes, and specific scenarios are retained. Based on these high-quality risk samples, a project risk causal knowledge system is built, and the characteristic manifestations, intermediate influencing factors, and causal chains of various risks are sorted out. The correlation mapping relationship between multimodal features and the root causes of project risks is established. A multimodal causal reasoning network model structure is constructed. The preprocessed multimodal pure feature vectors are used as the network input layer. Various risk events of the project are used as observed variables, and various potential risk factors of the project are used as latent variables. A causal structure graph adapted to the project management scenario is constructed. The massive multimodal feature data is traversed and trained through the causal structure learning algorithm to mine the conditional independence relationship and real causal edge between each feature variable. The causal correlation degree of each feature relative to the risk event is calculated. Highly correlated causal features are selected, and pseudo-correlation features, noisy features, and redundant interference features that only have statistical correlation and no actual business causal logic are removed. A multi-level reverse tracing reasoning link is established. Based on the completed causal structure and the purified causal characteristics, a four-level reverse reasoning architecture is constructed, including the observation layer, phenomenon layer, inducement layer, and root cause layer. The identified abnormal characteristics of project risks are taken as the surface phenomena of the observation layer. The influencing factors at each level are deduced layer by layer. First, the surface direct abnormal factors corresponding to the risks are identified. Then, the intermediate inducement factors such as process deviation, resource deviation, management deviation, and execution deviation behind the surface abnormalities are dissected. Finally, the underlying essential risk root causes corresponding to the system and process, personnel collaboration, materials and equipment, solution design, and R&D process operation are located. In the process of multi-level reverse tracing reasoning, the confidence level, influence weight and contribution ratio of each causal reasoning chain are calculated in real time. The results of multi-level tracing are quantitatively scored and prioritized. The analysis results of risk manifestations, intermediate causes and essential root causes are output in layers. A visual and interpretable causal reasoning path is generated, and accurate layer-by-layer tracing from multi-modal abnormal characteristics and risk manifestations to causes is completed.

7. The enterprise project management method based on big data and multimodal approaches according to claim 6, characterized in that: The constructed multimodal causal reasoning network model structure uses preprocessed multimodal pure feature vectors as the network input layer, various project risk events as observed variables, and various potential risk factors as latent variables. It constructs a causal structure graph adapted to the project management scenario, and trains it through massive multimodal feature data using a causal structure learning algorithm. This process uncovers the conditional independence relationships and true causal edges between feature variables, calculates the causal correlation between each feature and risk event, and filters out highly correlated causal features. It also removes pseudo-correlation features, noisy features, and redundant interference features that only have statistical correlation but lack actual business causal logic. These features include: A standardized model input feature matrix is ​​constructed. Unified vector quantization is applied to the multimodal clean features obtained after hierarchical decoupling and noise reduction. Dimensional normalization and feature alignment encoding are performed on the structured numerical features, text semantic features, and audiovisual visual features with different dimensions and types. All heterogeneous features are uniformly converted into standardized multimodal feature vectors with fixed dimensions. The model input feature matrix is ​​constructed based on the standardized multimodal feature vectors to complete the standardized construction of input samples for the causal inference network. Define a hierarchical variable system for causal inference network, set all standardized multimodal feature vectors as observable input variables of the model, set various project risk events such as schedule anomalies, cost anomalies, quality anomalies, and coordination anomalies as observable output variables, and set potential influencing factors such as personnel, materials, equipment, plans, processes, and management that cause project risks as latent variables, thus completing the definition of the variable space of the causal inference model adapted to project management scenarios. Initialize the project risk causal topology network structure. Based on the business relationship logic between observable input variables, observed output variables and latent variables, initialize the project risk scenario causal structure topology graph, construct an initial causal network structure containing feature nodes, risk event nodes and risk cause nodes, and form a multi-dimensional variable relationship topology architecture. Based on causal structure learning, we achieve accurate feature selection and network optimization. We use the causal structure learning algorithm to train all multimodal feature samples in batches. We perform conditional independence tests on the variables of each feature node, verify the direct causal dependence between different multimodal features and risk events, and eliminate indirect transitive associations and random statistical associations between variables. During training, the true causal edges between feature variables, risk events, and risk triggers are discovered, and the causal association scores and marginal contributions of each feature corresponding to the risk event are calculated to generate a quantitative causal association weight matrix. Based on a preset causal correlation threshold, multimodal features are stratified and filtered. Core causal features that meet the causal correlation standard and have actual business causal support are retained, while pseudo-correlation features that only have numerical statistical correlation and no actual business causal logic, as well as noise features and redundant features, are removed. This completes the causal dimension denoising and purification of multimodal features, and outputs a concise, highly reliable multimodal feature set with real causal logic.

8. The enterprise project management system and method based on big data and multimodal operation as described in claim 6, characterized in that: The aforementioned multi-level reverse tracing reasoning link, based on the constructed causal structure and refined causal characteristics, builds a four-level reverse reasoning architecture including an observation layer, a phenomenon layer, a trigger layer, and a root cause layer. The identified abnormal characteristics of project risks are treated as surface phenomena in the observation layer, and the influencing factors at each level are deduced layer by layer. First, the direct surface-level abnormal factors corresponding to the risks are identified, then the intermediate-level triggering factors behind the surface abnormalities, such as process deviations, resource deviations, management deviations, and execution deviations, are dissected. Finally, the underlying essential risk root causes corresponding to systems and processes, personnel collaboration, materials and equipment, solution design, and R&D process operations are located, including: A four-level progressive reverse reasoning architecture is constructed. Based on the trained risk causal topology and the purified multimodal causal feature set, the four-level reverse reasoning architecture consisting of observation layer, phenomenon layer, inducement layer and root cause layer is initialized. The four-layer architecture forms a risk reverse reasoning logical link that goes from the surface to the core, decomposes layer by layer and penetrates level by level. The observation layer completes the collection and anchoring of original risk anomaly signals. It inputs various multimodal pure features corresponding to project risk anomalies identified by the system through the dual-threshold risk judgment mechanism into the observation layer. Through the unified collection of time series numerical fluctuation anomalies, text semantic risk features, and audiovisual site violation anomalies, the observation layer completes the summarization, collection and precise anchoring of project risk original anomaly data. The phenomenon layer realizes the aggregation and transformation of discrete abnormal features into explicit risk phenomena. Based on the original multimodal abnormal features collected by the observation layer, it performs correlation clustering and fusion on various discrete abnormal features. Combined with the project risk business type, it matches specific risk phenomena such as R&D cycle delays, cost overruns, quality defects, and collaboration conflicts. It strips away the noise of simple data fluctuations without business support and condenses into a set of surface risk events and intuitive abnormal manifestations that are currently happening in the project, thus completing the accurate transformation from scattered abnormal features to complete explicit risk phenomena. The inducement layer completes the reverse decomposition and screening of mid-level deviation factors. Taking the explicit risk phenomena output by the phenomenon layer as the starting node for reasoning, and relying on the feature-deviation association rules built into the causal network, it traverses the real causal edge relationships of the causal topology layer by layer in reverse, deeply decomposes various intermediate deviation factors that cause surface risk phenomena, accurately identifies mid-level inducing factors such as iteration progress control deviation, R&D cost consumption deviation, R&D human / tool ​​resource allocation deviation, R&D execution process deviation, and team collaboration management deviation during project operation, quantifies the contribution weight of each mid-level deviation to the surface risk phenomenon, screens high-matching core mid-level inducements, and eliminates irrelevant interference deviation factors. The root cause layer enables the identification and positioning of underlying essential risks. For the mid-level deviation causes identified through screening, it conducts ultimate attribution matching by combining the business logic of the entire project process with the correlation of the risk causal knowledge base. This accurately maps various mid-level process deviations to underlying essential root causes such as defects in systems and processes, unreasonable product solution design, non-standard R&D operation processes, inadequate configuration of development tools / testing environments, poor personnel collaboration, and lack of version control. This constructs a complete causal tracing chain that includes multimodal anomaly characteristics, surface risk phenomena, mid-level process deviations, and underlying essential root causes. The entire process is based on the real causal edge mining based on causal structure learning and targeted tracing.

9. The enterprise project management method based on big data and multimodal approaches according to claim 1, characterized in that: The establishment of a human-machine collaborative risk assessment self-iterative optimization closed loop involves classifying and labeling samples and retraining the model's assessment results, combining risk scenario sample augmentation strategies, and continuously iteratively updating the model's inference weights, feature selection rules, and risk judgment parameters, including: A human-machine collaborative two-way linkage sample update architecture is constructed, and a closed-loop iterative system for risk assessment is built. Data is retained throughout the entire process for each project risk assessment event. The abnormal feature screening results, causal reasoning paths, risk level judgment results, and risk root cause reasoning conclusions automatically output by the model are fully recorded. The manual review content, risk result correction records, model misjudgment reason annotations, and root cause correction labels are collected simultaneously from project managers and industry experts, forming a pairwise comparison sample data where the automatic assessment results of the model and the manual expert review results correspond one-to-one. All risk assessment samples accumulated during the closed-loop iteration process are finely classified and labeled, dividing them into samples correctly judged by the model, samples missed by the model, samples misjudged by the model, samples with ambiguous abnormal boundaries, and samples of newly added unknown scenarios. Corresponding structured labels are matched for each type of sample, and the labeling and archiving of risk scenario type, abnormal feature combination, judgment parameter deviation, and causal reasoning defect dimensions are completed, and a human-machine collaborative iterative sample library that can be dynamically and incrementally expanded is constructed. A multi-dimensional scenario-based sample enhancement strategy is adopted, which supplements various scarce sample scenarios and broadens the coverage dimension of risk samples through multiple expansion methods such as feature space transformation, modality recombination, scenario migration and temporal fine-tuning. Establish a differentiated, hierarchical, priority-based sample reinjection and iterative training mechanism, set training priorities for enhanced samples of different defect types, implement targeted optimization of the risk assessment model and causal inference network, and modify feature screening logic, supplement causal relationship constraints, and optimize risk assessment threshold parameters through classified samples. A fully automated closed-loop self-iterative optimization system is constructed. Through continuous loop execution of automatic risk assessment, manual verification and correction, sample structure labeling, scenario-based data augmentation, hierarchical backfeeding training, and model parameter iterative update process, a complete human-machine collaborative iterative closed loop is formed, which enables risk assessment parameters, feature selection rules, and causal inference weights to continuously adapt to new working conditions, new risks, and dynamic business changes in the project.

10. A multimodal enterprise project management system based on big data, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the enterprise project management method based on big data multimodality as described in any one of claims 1 to 9 by executing the machine-executable instructions.

Images