Project performance evaluation and risk early warning method for multi-source heterogeneous scientific research data

Abstract

The invention discloses a project performance evaluation and risk early warning method for multi-source heterogeneous scientific research data, and relates to the technical field of scientific research project management, and the method comprises the steps: collecting the multi-source heterogeneous scientific research data, cleaning, converting, filling missing values, and storing in a unified format; extracting performance and risk features, and fusing to obtain a fused feature set; combining an analytic hierarchy process, an entropy weight method and the like to construct a dynamic weight performance evaluation model; constructing a multi-modal risk early warning model by adopting multiple algorithms and optimizing a threshold value; collecting data in real time to update a feature set, and dynamically adjusting the model; and visually outputting a result, generating an improvement suggestion, and forming closed-loop feedback. According to the method, the multi-source data processing efficiency and evaluation accuracy are improved, the risk early warning timeliness and adaptability are enhanced, a management closed loop is formed, and the problems of difficult data integration, evaluation lagging, insufficient early warning and the like of a traditional method are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of scientific research project management, and in particular to a project performance evaluation and risk early warning method for multi-source heterogeneous scientific research data. BACKGROUND

[0002] Modern scientific research projects have the characteristics of multi-disciplinary intersection, long research period, and large resource investment. The scientific research data generated in the whole life cycle of the project has significant multi-source heterogeneous characteristics, covering structured table data such as fund account and personnel attendance, semi-structured XML / JSON data such as project proposal and interim report, and unstructured text and multimedia data such as academic papers, experimental images, and conference recordings. These data are scattered in different systems, have large format differences and low correlation, and traditional data processing methods cannot effectively integrate them, which leads to the problem that the data value cannot be fully tapped, and the project performance evaluation and risk early warning are difficult at the basic level.

[0003] The existing project performance evaluation method has obvious limitations. The evaluation indicators are mostly focused on static results such as scientific research output, ignoring process and application indicators such as fund efficiency, personnel contribution, and achievement transformation. The indicator system is not comprehensive enough. The weight determination mostly relies on a single analytic hierarchy process or entropy weight method, and the subjective weight is easily affected by the experience bias of experts, and the objective weight is difficult to reflect the differences in the field and the characteristics of the project phase, leading to a disconnection between the evaluation results and the actual value of the project. At the same time, the evaluation mostly uses static analysis in stages, lacks dynamic tracking of the whole cycle data of the project, cannot reflect the performance fluctuations in the progress of the project in a timely manner, and is difficult to support fine management decisions.

[0004] The short board of the risk early warning link is more prominent. The existing method mostly uses a single machine learning algorithm to build the early warning model, which cannot adapt to the complex characteristics of multiple types of risks such as technology, management, resources, and external environment, and the risk feature extraction only relies on structured data, ignoring the semantic information and potential risk signals in unstructured data, leading to narrow coverage and low precision of the early warning. The risk early warning threshold is mostly a fixed value, which does not consider the differences in the field (basic research, applied research, etc.), funding level, and implementation stage, and is difficult to match the risk bearing capacity of different projects. In addition, the data security protection mechanism is missing, and sensitive information in scientific research data is easily leaked; and there is a lack of a closed-loop process of “evaluation-early warning-feedback-optimization”, and the early warning information and improvement suggestions are difficult to effectively land, and a management closed loop cannot be formed. These problems together make it difficult for the existing method to meet the core needs of modern scientific research projects such as “accurate evaluation, timely early warning, and dynamic management and control”, and it is urgent to build a whole-cycle performance evaluation and risk early warning system that adapts to multi-source heterogeneous data. SUMMARY

[0005] The project performance evaluation and risk early warning method for multi-source heterogeneous scientific research data provided in the application solves the problems mentioned in the prior art.

[0006] In order to achieve the above-mentioned purpose, the application adopts the following technical scheme: a project performance evaluation and risk early warning method for multi-source heterogeneous scientific research data, comprising the following steps: Data acquisition and heterogeneous fusion preprocessing: collecting multi-source heterogeneous data in the whole life cycle of scientific research, using distributed crawler and API linkage collection, structured data is extracted by JDBC, semi-structured data is parsed and converted by label, and unstructured data is processed by OCR and voice transcription; in the cleaning link, the abnormal value detection and hash deduplication are used, the improved K nearest neighbor is used to fill in the numerical missing value, and the weighted mode is used to fill in the category type; finally, the data is standardized into a unified JSON-LD format by a data standardization engine, and the associated project identifier is stored in a distributed data warehouse; Multi-dimensional feature deep extraction and optimized fusion: constructing a feature system based on the preprocessed data; the feature extraction adopts a multi-algorithm collaborative strategy, the structured data is calculated for statistical features, the semi-structured data is parsed for key features, and the unstructured data is extracted for corresponding features by BERT, CNN and MFCC; the feature fusion adopts a two-level architecture of feature layer and decision layer to obtain a fused feature set; Dynamic weight performance evaluation model construction and training: based on the performance features, a three-element combined weight model of analytic hierarchy process, entropy weight method and coefficient of variation is constructed, the subjective weight is determined by consistency test, the objective weight is obtained by calculating the information entropy, and the deviation is corrected according to the dispersion degree; a dynamic weight adjustment mechanism is set to adjust the index weight according to the project stage; Multi-modal risk early warning model construction and threshold optimization: based on risk features, a hybrid modeling is constructed by deep learning and traditional algorithms; a risk transmission path model is constructed; the risk early warning threshold is dynamically determined, and an adjustment function is constructed combined with historical risk data; Full-cycle dynamic monitoring and model incremental updating: building a real-time data acquisition link, adjusting and updating the frequency according to the project stage; generating an incremental feature set, and filtering data by a sliding window; updating the parameters of the performance model by incremental learning, and adapting new data by transfer learning for the risk model; Multi-dimensional result output and closed-loop feedback optimization: the visualization engine generates results including hierarchical dashboards and three-color warning light displays; customized reports are generated according to user roles.

[0007] Further, in the multi-dimensional feature deep extraction and optimized fusion step, the extracted features are subjected to dual processing of standardization and regularization, the standardization processing adopts an improved Z-score algorithm, and the standardization formula is wherein represents the standardized feature value, represents the original feature value, denotes the mean value of the feature in the jth project field, denotes the standard deviation of the feature in the jth project field, and ε denotes a very small positive number. Through this process, the differences between different fields and different dimensional features are eliminated, and the regularization process uses L2 regularization to constrain the features.

[0008] Further, in the dynamic weight performance evaluation model construction and training step, the entropy weight method is used for secondary optimization of the ternary combined weight, and the calculation method of the optimized combined weight is wherein denotes the final combined weight of the jth performance indicator, β denotes the weight coefficient of the analytic hierarchy process, γ denotes the weight coefficient of the entropy weight method, and 1-β-γ denotes the weight coefficient of the variation coefficient method, denotes the subjective weight determined by the analytic hierarchy process, denotes the objective weight determined by the entropy weight method, denotes the correction weight determined by the variation coefficient method, which balances the influence of subjective experience, data information and feature dispersion degree, so that the weight distribution is more in line with the complexity and dynamics of scientific research projects.

[0009] Further, in the data acquisition and heterogeneous fusion preprocessing step, the multi-modal semantic alignment technology is used when structuring the unstructured data. The text type unstructured data is segmented, tagged and recognized by the BERT-WWM model, and the precision of entity extraction is optimized by combining the field dictionary. Key entities are extracted and structured; image type unstructured data is identified by YOLOv8 algorithm, and chart data is extracted by combining image semantic segmentation technology; audio data is converted into text by convolutional neural network and end-to-end speech transcription model, and key information is extracted by entity recognition; after conversion, multimedia data is associated through knowledge graph to construct project, personnel, achievement and equipment semantic network.

[0010] Further, in the multi-dimensional feature deep extraction and optimization fusion step, the feature fusion algorithm uses a stack autoencoder enhanced by attention mechanism. The encoder part includes an input layer, a hidden layer and an encoding layer. The input layer receives high-dimensional original features, the hidden layer extracts local features through convolution kernels, and the encoding layer uses an attention weight matrix to give different weights to different features. The decoder part reconstructs the features through inverse convolution operation, calculates the reconstruction error and feature correlation; and a feature effectiveness verification mechanism is introduced in the fusion process to perform variance analysis and significance test on the fused features.

[0011] Further, in the multi-modal risk early warning model construction and threshold optimization step, the federated learning framework is used to train the risk early warning sub-model. In view of the data privacy requirements of different scientific research units, each unit trains the sub-model locally. The server aggregates the parameters using the FedAvg algorithm and introduces a differential privacy mechanism to add noise disturbance to the parameters. A dynamic learning rate strategy is used during model training, which adjusts adaptively according to the change of the model loss function. When the loss decreases slowly, the learning rate is reduced. A multi-level threshold system is used for risk early warning.

[0012] Further, in the full-cycle dynamic monitoring and model incremental updating step, a data quality evaluation system is constructed to evaluate the quality of new data from four dimensions: accuracy, completeness, consistency, and timeliness. Accuracy is verified by comparison with authoritative data sources. Completeness calculates the proportion of missing values. Consistency detects data conflicts across data sources. Timeliness determines the time difference between data update time and current time. A data quality scoring mechanism is used. The model incremental updating uses a dual-model parallel mechanism, with the historical model and the updated model running simultaneously. The output results of the two models are compared through A / B testing. A model version management mechanism is established to record the parameters, data, and evaluation indicators of each update, supporting model version backtracking.

[0013] Further, in the multi-dimensional result output and closed-loop feedback optimization step, the visualization system uses a front-end and back-end separation architecture. The front-end is based on ECharts and Three.js to build an interactive interface. The back-end pushes real-time update data through data interfaces. Performance display is subdivided into multiple scenarios. The project progress scenario uses a Gantt chart to show the completion status of time nodes. The fund use scenario uses a Sankey diagram to present the fund flow path. The achievement distribution scenario uses a geographic heat map to display the achievement transformation area. Risk early warning adds a risk evolution prediction function. Feedback optimization introduces a reinforcement learning mechanism, taking the execution effect of improvement suggestions as a reward signal.

[0014] Further, in the multi-modal risk early warning model construction and threshold optimization step, the dynamic adjustment of the risk early warning threshold uses a combination of time series analysis and domain adaptation. Specifically, the dynamic risk early warning threshold at time t is determined as follows: where represents the dynamic risk early warning threshold at time t, represents the initial threshold, λ(s) represents the project stage influence coefficient at time s, δ(s) represents the risk trend coefficient at time s, which is determined according to the trend slope calculated from historical risk data, and t represents the time the project has been implemented.

[0015] Further, in the data acquisition and heterogeneous fusion preprocessing step, a data security protection link is added, sensitive information is processed using data desensitization technology, TLS1.3 protocol encryption is used in the data transmission process, and partition encryption strategy is used for data storage. Core data is encrypted using AES-256 algorithm, and ordinary data is encrypted using DES algorithm. The key is periodically replaced through the key management system. In addition, a data access permission control system is established, access permissions are allocated based on the RBAC model, and field-level permission control is refined. All data operation logs are recorded, which meets the scientific research data security management specification.

[0016] Compared with the existing technology, the beneficial effects of the present application are: At the data processing level, the present application innovatively adopts a multi-source heterogeneous data fusion preprocessing strategy. Through distributed acquisition and multi-modal conversion technology, structured, semi-structured and unstructured data are efficiently integrated into a unified format. Combined with improved missing value filling and deduplication cleaning methods, the data integrity and consistency are greatly improved. At the same time, a data security protection mechanism is introduced. Through desensitization encryption, permission control and log tracing, deep correlation is realized under the premise of ensuring data privacy and security, solving the problem of data dispersion and insufficient value mining in traditional methods, and laying a high-quality data foundation for subsequent evaluation and early warning.

[0017] At the performance evaluation level, the multi-dimensional index system constructed covers all factors such as output, efficiency, contribution and transformation, breaking through the limitations of traditional methods that focus on results and ignore processes. The ternary combination weight model balances the influence of subjective experience, data information and feature dispersion degree, and cooperates with the dynamic weight adjustment mechanism. The weight distribution can be adapted according to the project phase and field characteristics, so that the evaluation result is more in line with the actual value of the project. The introduction of weighted TOPSIS method and time decay factor realizes the dynamic tracking and precise quantification of performance, solving the problem of disconnection between traditional static evaluation and reality.

[0018] At the risk early warning level, the multi-modal risk early warning model uses differentiated algorithms for different risk types, combined with risk transmission path analysis, which greatly widens the coverage and precision of early warning. The dynamic adaptive threshold is determined by integrating project field, stage and risk trend, avoiding the rigid defect of fixed threshold, and realizing the grading response and precise control of risk. The application of federated learning framework improves the model generalization while ensuring data privacy, further enhancing the reliability of early warning.

[0019] In terms of management closed loop and practicality, the whole-cycle dynamic monitoring and incremental updating mechanism ensures that the evaluation and early warning can adapt to the changes in project progress in a timely manner through real-time data collection and model iteration, avoiding the lag of traditional methods. Multi-dimensional visual output and customized reports make the evaluation results and early warning information easier to understand and apply, meeting the management needs of different roles. The closed loop feedback process driven by reinforcement learning, combined with the expert correction opinions stored in the blockchain, realizes the continuous optimization of evaluation indicators, early warning thresholds and model parameters, forming a management closed loop. In addition, the data security protection system and permission control mechanism comply with the scientific research data security specifications, further improving the engineering practicality and compliance of the method. Overall, the present application completely changes the extensive mode of traditional scientific research project management, providing strong support for the optimal allocation of scientific research resources and the improvement of project quality. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 A schematic block diagram of the project performance evaluation and risk early warning method for multi-source heterogeneous scientific research data proposed by the present application is shown in the figure. Figure 2 A comparison chart of multi-source heterogeneous data processing effects is shown in the figure. Figure 3 A comparison radar chart of multi-type risk early warning accuracy is shown in the figure. Figure 4 A trend chart of closed loop optimization iteration effect is shown in the figure. Figure 5 A comparison thermal map of data security and privacy protection effect is shown in the figure. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0022] In the description of the present application, it should be understood that the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.

[0023] In addition, the terms "first", "second", etc. are used only for descriptive purposes and should not be construed as indicating or implying relative importance or an indicated number of technical features. Therefore, the features defined as "first", "second", etc. can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise specifically limited. In addition, the terms "mounting", "connecting", "connecting" should be broadly understood, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium, or it can be the communication between two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances, and the present application will be further described in detail below with reference to the accompanying drawings.

[0024] Referring to Figures 1 to 5 A project performance evaluation and risk early warning method for multi-source heterogeneous scientific research data, comprising the following steps: Data collection and heterogeneous fusion preprocessing, collecting multi-source heterogeneous scientific research data covering the whole life cycle of scientific research, the multi-source heterogeneous scientific research data including structured data, semi-structured data and unstructured data, the structured data covering table data such as fund account, personnel attendance, equipment use record, the semi-structured data including project proposal, interim report and other XML / JSON format data, and the unstructured data containing academic papers, experimental images, conference audio and other text and multimedia data; distributed crawler and API interface linkage are adopted for collection, structured data is extracted through JDBC direct connection to database, semi-structured data is parsed and labeled field is converted, and unstructured data is identified by OCR technology and audio information is extracted by voice transcription technology; noise data is identified by outlier detection algorithm in the cleaning link, repeated data is removed by hash check, numerical missing values are filled by improved K nearest neighbor algorithm, filling values are calculated combined with sample similarity and feature correlation, and category missing values are filled by weighted mode filling method, and the filling category is determined according to the weight distribution of feature importance; finally, all data are converted into unified JSON-LD format through data standardization engine, and the project unique identifier is stored in distributed data warehouse to ensure data traceability; Multi-dimensional feature depth extraction and optimization fusion, based on the pre-processed data to construct the feature system, the project performance features are subdivided into scientific research output indicators, fund efficiency indicators, personnel contribution indicators and achievement transformation indicators, the scientific research output indicators include quantitative indicators such as paper citation and invention authorization, the fund efficiency indicators include efficiency indicators such as fund utilization rate and equipment purchase cost performance ratio, the personnel contribution indicators involve core member workload, title matching degree and other dimensions, and the achievement transformation indicators include application indicators such as technology transfer amount and school-enterprise cooperation times; the risk features are divided into technical risk indicators, management risk indicators, resource risk indicators and external environment risk indicators, the technical risk indicators include technical maturity, experiment success rate and other parameters, the management risk indicators include progress deviation rate, document completeness and other elements, the resource risk indicators involve fund gap rate, equipment failure frequency and other data, and the external environment risk indicators include policy change impact, industry technology iteration speed and other variables; the feature extraction adopts a multi-algorithm collaborative strategy, the structured data is extracted by statistical feature quantity calculation, the semi-structured data is extracted by analyzing the label level, and the unstructured data is extracted by using the BERT pre-training model to extract text semantic features, using CNN to extract experimental image texture features, and using MFCC algorithm to extract voice emotion features; the feature fusion adopts a two-level fusion architecture of feature layer and decision layer, the feature layer reduces the dimension of high-dimensional features by stacking autoencoders, and retains core features with cumulative variance contribution rate ≥95%, the decision layer assigns feature weights based on attention mechanism, and gives high weights to key features such as achievement transformation and technical maturity, to obtain a fusion feature set that takes into account comprehensiveness and pertinence; Dynamic weight performance evaluation model construction and training, based on the project performance features in the fusion feature set, a three-element combined weight model of analytic hierarchy process-entropy weight method-coefficient of variation method is constructed, the analytic hierarchy process constructs a hierarchical judgment matrix through expert questionnaire, and the subjective weight is determined through consistency check (CR≤0.1), the entropy weight method obtains the objective weight by calculating the feature information entropy and utility value, and the coefficient of variation method corrects the weight deviation according to the feature dispersion degree; adopt a dynamic weight adjustment mechanism, focus on fund efficiency and personnel contribution index weight in the project establishment stage, increase the process index weight such as progress deviation rate and experiment success rate in the middle stage, and strengthen the achievement transformation and scientific research output index weight in the conclusion stage; based on the combined weight, a multi-dimensional performance evaluation model is constructed, the weighted TOPSIS method is used to calculate the comprehensive performance score, the time decay factor is introduced to correct the weight of historical data, and the weight coefficient of recent data is higher than that of long-term data, so that the evaluation result is consistent with the dynamic progress of the project; train the model through historical project data (including 5000+ scientific research project samples), optimize the model parameters by five-fold cross-validation, and ensure that the model evaluation accuracy is ≥96%; The multi-modal risk early warning model is constructed and the threshold is optimized. Based on the risk features in the fusion feature set, a hybrid modeling strategy combining deep learning and traditional algorithms is adopted. The technical risk sub-model adopts LSTM network to input the time sequence of experimental data to predict the technical breakthrough probability. The management risk sub-model uses the random forest algorithm to mine the association rules between the progress, documents and management vulnerabilities. The resource risk sub-model identifies the risk of funding and equipment resource gap through the support vector machine classifier. The external environment risk sub-model analyzes the influence probability of policy, market and other variables based on the Bayesian network. A risk transmission path model is constructed to quantify the correlation strength between different risk types and identify the key nodes of risk transmission. The risk early warning threshold is determined dynamically and adaptively. The risk data and actual risk occurrence of scientific research projects in different fields in the past ten years are collected. Combined with the project field (basic research, applied research and experimental development), funding level and implementation cycle, a threshold dynamic adjustment function is established. The occurrence probability corresponding to different risk values is calculated through the cumulative distribution function to determine the initial threshold by matching the acceptable risk level of the project. The initial threshold is iteratively optimized according to the real-time risk data. Full-cycle dynamic monitoring and model incremental updating, a real-time data acquisition link is built, real-time experimental equipment operation data is obtained through a sensor network, paper publication and invention authorization information is synchronized relying on an academic database API, and fund use data is dynamically updated by connecting a financial system. The data update frequency is adjusted according to the project phase. In the early stage of project, it is updated once every two weeks, in the middle stage, it is updated once every week, and before the project is completed, it is updated three times a month. The data acquisition and preprocessing, feature extraction and fusion steps are repeated to generate an incremental feature set. A sliding window mechanism is used to select effective data, and the window size is dynamically set according to the data update frequency. The performance evaluation model adopts an incremental learning algorithm, the parameters of the base model trained by historical data are frozen, and only the feature weights and model output layer parameters corresponding to the new data are updated. The risk early warning model is adapted to the new data distribution through transfer learning, the historical model knowledge is migrated to the update process, and the cost of retraining is reduced. A model performance evaluation system is established, with accuracy, recall rate and F1 value as core indicators. When the model evaluation indicators decrease by more than 5%, full data retraining is triggered to ensure that the model adapts to the dynamic changes in data. Multi-dimensional result output and closed-loop feedback optimization, a visual engine is used to generate multi-scene display results, performance evaluation results are presented through a hierarchical dashboard, core indicators are displayed in proportion using a ring chart, trend indicators are presented through a stacked line chart to show the change rule, and comparative indicators are displayed using a radar chart to show the gap with similar projects; risk warning information is marked by a three-color warning light, with red indicating high risk, yellow indicating medium risk, and blue indicating low risk, combined with a heat map to show the risk distribution area, and a directed graph is used to visualize the risk transmission path; customized reports are generated for different user roles, progress optimization suggestions are output for project managers, resource allocation schemes are provided for management departments, and performance evaluation summaries are presented for funding agencies; the feedback mechanism adopts a double-path of manual correction and data feedback, the manual correction results are audited by experts and then included in the sample library, the data feedback monitors the implementation effect of the improvement suggestions, and the performance improvement amplitude and risk reduction degree before and after optimization are quantified, forming a closed-loop management process of evaluation, warning, feedback and optimization, and continuously improving the applicability and accuracy of the method.

[0025] In the present application, the extracted features are subjected to dual processing of standardization and regularization in the multi-dimensional feature depth extraction and optimization fusion step, the standardization processing adopts an improved Z-score algorithm, a dynamic adjustment factor is introduced to adapt to the distribution characteristics of scientific research data, and the standardization formula is wherein represents the standardized feature value, represents the original feature value, represents the mean value of the feature in the jth project field, represents the standard deviation of the feature in the jth project field, and epsilon represents a very small positive number (10^-6), which is used to avoid calculation abnormalities caused by zero standard deviation, and through this processing, the differences between different fields and different dimension features are eliminated, the regularization processing uses L2 regularization to constrain the features, further improving the stability of feature fusion, and making subsequent model training more efficient.

[0026] In the present application, in the dynamic weight performance evaluation model construction and training step, the entropy weight method is used to optimize the ternary combination weight, and the optimized combination weight calculation method is wherein represents the final combination weight of the jth performance indicator, beta represents the weight coefficient of the analytic hierarchy process (the value range is 0.3-0.4), gamma represents the weight coefficient of the entropy weight method (the value range is 0.4-0.5), and 1-beta-gamma represents the weight coefficient of the variation coefficient method, represents the subjective weight determined by the analytic hierarchy process, represents the objective weight determined by the entropy weight method, The correction weight determined by the coefficient of variation method is represented, and through the mode, the influence of subjective experience, data information and feature dispersion degree is balanced, the weight distribution is more suitable for the complexity and dynamicity of the scientific research project, and the precision of performance evaluation is improved.

[0027] In the present application, in the data acquisition and heterogeneous fusion preprocessing step, the multi-modal semantic alignment technology is adopted when the unstructured data is structured converted, the text type unstructured data is segmented, word class is marked and entity is recognized through the BERT-WWM model, the field dictionary (covering 3000+ scientific research professional terms) is combined to optimize the entity extraction precision, the key entities such as project name, research content and achievement type are extracted and structured; the image type unstructured data adopts the YOLOv8 algorithm to identify experimental equipment, chart curve and other targets, combines the image semantic segmentation technology to extract chart data, and converts into structured numerical value; the audio type data extracts speech features through the convolutional neural network, adopts the end-to-end speech transcription model to convert into text, and then extracts key information through entity recognition; after the multimedia data conversion, the knowledge graph association is carried out, the project, personnel, achievement and equipment semantic network is constructed, and the deep value mining of unstructured data is realized.

[0028] In the present application, in the multi-dimensional feature deep extraction and optimization fusion step, the feature fusion algorithm adopts the attention mechanism enhanced stacked autoencoder, the encoder part includes the input layer, the hidden layer and the encoding layer, the input layer receives the high-dimensional original feature, the hidden layer extracts local features through the convolution kernel, the encoding layer adopts the attention weight matrix to give different weights to different features, and the contribution of core features such as paper influence and invention conversion rate is highlighted; the decoder part reconstructs the features through the deconvolution operation, calculates the reconstruction error and feature correlation, and iteratively optimizes the network parameters; in the fusion process, the feature effectiveness verification mechanism is introduced, the variance analysis and significance test are carried out on the fused features, the redundant features with a contribution degree lower than 5% are removed, and finally the fusion feature set with a dimension compression of more than 60% and an information retention rate of ≥95% is output, so that the calculation complexity of subsequent model training is reduced.

[0029] In the multi-modal risk early warning model construction and threshold optimization step of the application, a federal learning framework is used to train a risk early warning sub-model, and in view of the data privacy needs of different scientific research units, each unit trains a sub-model locally and only uploads model parameters to the federal server; the server aggregates the parameters using the FedAvg algorithm, introduces a differential privacy mechanism to add noise disturbance to the parameters, and ensures data privacy and security; a dynamic learning rate strategy is used in the model training process, the initial learning rate is set to 0.01, and the learning rate is adaptively adjusted according to the change of the model loss function, the learning rate is reduced when the loss decreases slowly, and the initial value is restored when the loss rebounds; the risk early warning adopts a multi-level threshold system, the first level threshold corresponds to a risk occurrence probability of 10%-20%, triggers a risk prompt and pushes a self-check list, the second level threshold corresponds to a risk occurrence probability of 20%-40%, starts a special check of the project management department, and the third level threshold corresponds to a risk occurrence probability of 40% or more, suspends the project fund allocation and formulates a rectification scheme, and realizes the grading response and accurate control of risk.

[0030] In the full-cycle dynamic monitoring and model incremental updating step of the application, a data quality evaluation system is constructed to evaluate the quality of new data from four dimensions of accuracy, integrity, consistency and timeliness, the accuracy is verified by comparison with authoritative data sources, the integrity is calculated by the proportion of missing values, the consistency is detected by cross-data source data conflict, and the timeliness is judged by the time difference between data update time and current time; a data quality scoring mechanism is adopted, with a full score of 100, and data below 60 points need to be re-collected and pre-processed; the model incremental updating adopts a double model parallel mechanism, simultaneously running the historical model and the updating model, and comparing the output results of the two through A / B testing, when the F1 value of the updating model is more than 3% higher than that of the historical model, the updating model is automatically switched; a model version management mechanism is established to record the parameters, data and evaluation indexes of each update, and supports model version backtracking, which can quickly roll back to the optimal version when the new model appears abnormal.

[0031] In the multi-dimensional result output and closed-loop feedback optimization step of the application, the visualization system adopts a front-end and back-end separation architecture, the front-end is based on ECharts and Three.js to build an interactive interface, supports chart scaling, drilling and rotation operations, and the back-end pushes real-time update data through a data interface; performance display is subdivided into multiple scenes, project progress scene uses Gantt chart to show time node completion, fund use scene uses Sankey diagram to present fund flow path, and achievement distribution scene uses geographic heat map to show achievement transformation area; risk early warning adds risk evolution prediction function, based on LSTM network to predict the risk change trend in the next 3 months, and generates risk development curve; feedback optimization introduces reinforcement learning mechanism, takes the execution effect of improvement suggestions as reward signal, optimizes subsequent evaluation index weight and early warning threshold, and expert correction opinion is stored through blockchain technology, ensuring that the feedback process is traceable and tamper-proof, forming a continuous iteration of the optimization closed loop.

[0032] In the present application, in the steps of multi-modal risk early warning model construction and threshold optimization, the dynamic adjustment of the risk early warning threshold adopts a combination of time series analysis and field adaptation, specifically through determining, wherein denotes the dynamic risk early warning threshold at time t, denotes the initial threshold, λ(s) denotes the project stage influence coefficient at time s, the value of the basic research stage is 0.8-1.0, the value of the applied research stage is 1.0-1.2, the value of the experimental development stage is 1.2-1.5, δ(s) denotes the risk trend coefficient at time s, which is determined according to the trend slope calculated from historical risk data, the value is 1.1-1.3 when the risk is rising, the value is 1.0 when the risk is stable, the value is 0.7-0.9 when the risk is falling, t denotes the time of project implementation, in this way, the threshold changes dynamically with the progress of the project and the risk trend, improving the timeliness and adaptability of the early warning.

[0033] In the present application, in the data collection and heterogeneous fusion preprocessing step, a data security protection link is added, sensitive information is processed using data desensitization technology, reversible desensitization algorithms are used for personnel identity card numbers, bank accounts, etc., to retain data statistical value while protecting privacy; during data transmission, TLS1.3 protocol encryption is used, data integrity is verified through digital signature, and transmission tampering is prevented; data storage uses partition encryption strategy, core data (such as fund data, achievement data) is encrypted using AES-256 algorithm, and ordinary data is encrypted using DES algorithm, the key is periodically replaced (period 90 days) through the key management system; in addition, a data access permission control system is established, access permissions are assigned based on the RBAC model, refined to field-level permission control, all data operation logs are recorded, log retention time ≥3 years, in line with the scientific research data security management specification; the missing value filling optimization is an adaptive filling strategy, first determine the influence degree of missing values through feature importance evaluation, if the influence degree is higher than 80%, use the improved K nearest neighbor algorithm, combined with feature correlation weighted calculation to fill the value, if the influence degree is lower than 50%, use the mean / mode filling, balance the filling precision and calculation efficiency.

[0034] The specific implementation of the system is further illustrated by two embodiments as follows: Example 1: Performance evaluation and risk early warning of new drug research and development basic research project This embodiment is aimed at a 3-year "anti-tumor small molecule new drug research and development" basic research project, which involves multi-source data (3000+ experimental images, 80+ academic papers, 12 sets of fund accounts, 40+ hours of conference recordings), and the technical risk is prominent, which needs to focus on processing unstructured data and dynamically evaluating the research progress, the specific implementation is as follows.

[0035] 1. Detailed implementation of technical scheme Data collection and heterogeneous fusion preprocessing: Scrapy distributed crawler was used to collect PubMed papers, hospital LIS system API was used to extract clinical experimental data, JDBC was used to directly connect to Oracle database to obtain fund accounts (structured), XML format interim report was parsed, ABBYY FineReader OCR was used to identify gel electrophoresis image text annotation in experimental images, and Iflytek speech transcription system was used to extract conference recording text (unstructured). Isolation forest algorithm was used to identify outliers in experimental data (such as cell activity data deviating from the mean by 3 standard deviations), hash check was used to remove duplicate uploaded paper PDFs, and the improved K nearest neighbor algorithm was used to fill in missing experimental temperature data (K=5, combined with experimental time and reagent batch similarity weighted calculation), and the weighted mode was used to fill in the missing personnel title (core member title weight 0.7). Data standardization is in JSON-LD format, and the project unique identifier "2024YF01-Drug" is stored in the Hadoop distributed warehouse, and the data security uses AES-256 encryption fund data, and the RBAC model assigns project leaders, researchers, and administrators three levels of permissions.

[0036] Multi-dimensional feature deep extraction and optimized fusion: Performance features include paper SCI impact factor (output), reagent fund proportion (efficiency), core researcher's average number of inventions (contribution), and invention transfer intention number (transformation); risk features include compound synthesis success rate (technology), experimental record completeness (management), instrument failure frequency (resources), and policy approval progress (external). Text data uses BERT-WWM model (hidden layer 768 dimensions) to extract semantic features of paper abstracts, experimental images use YOLOv8 to identify cell morphology targets, and MFCC algorithm extracts emotion features (13 dimensions) from conference recordings. Feature standardization uses the formula where is the mean value of the pharmaceutical field features (such as the mean value of the paper impact factor 4.2), is the field standard deviation (1.8), and ε=10⁻ 6 , and after standardization, L2 regularization constraint is used. Feature fusion uses a stack autoencoder enhanced by attention mechanism, with a 3x3 convolution kernel, an attention weight of 0.3 for "compound synthesis success rate", and an attention weight of 0.25 for "invention transfer intention number". The 3% contribution of "reagent procurement brand" redundant features is removed, and a 45-dimensional fusion feature set is output (information retention rate 96%).

[0037] Dynamic weight performance evaluation model construction: AHP invites 5 pharmaceutical experts to construct a judgment matrix, and passes the consistency test (CR=0.08≤0.1) to get the subjective weight (output index 0.3); the entropy weight method calculates the information entropy to get the objective weight (performance index 0.35); the coefficient of variation method is corrected (transformation index 0.25). The combined weight formula Take β = 0.35, γ = 0.45, 1-β-γ = 0.2, calculate the output 0.28, performance 0.32, contribution 0.18, transformation 0.22. The performance weight is raised to 0.38 in the project stage (1st year), and the technical risk weight is raised to 0.3 in the middle stage (2nd year), and the weighted TOPSIS method is used to calculate the comprehensive performance score, the time decay factor is multiplied by 0.8 for the 1st year data and 0.95 for the 2nd year.

[0038] Multi-modal risk early warning model construction: technical risk sub-model uses LSTM (input dimension 45, hidden layer 128), input compound synthesis time series data (updated every quarter); management risk uses random forest (tree number 100); resource risk uses support vector machine (kernel function RBF). The federal learning framework accesses 3 cooperative units, uploads parameters after 3 rounds of local training, and adds differential privacy noise (ε = 1.5) when FedAvg aggregation. Dynamic threshold formula Initial threshold = 0.3, basic research stage λ(s) = 0.9, 2nd year risk increase δ(s) = 1.2, t = 2 years integral , trigger secondary warning.

[0039] Dynamic monitoring and closed-loop feedback: experimental data is collected every two weeks, and paper and fund data is updated every month, with a sliding window size of 3 months. Incremental learning freezes the first 3 layers of LSTM parameters, only updates the output layer, and triggers full retraining when the F1 value of model performance evaluation decreases by 3%. Visualization uses ECharts to generate scientific research output radar chart and risk trend line chart, pushes the "optimized synthesis route" suggestion to the project leader, and the expert correction result is stored in the blockchain, after feedback, the transformation index weight is raised to 0.25.

[0040] 2. Effect verification data Table 1: Comparison of data processing and technical risk early warning effect Indicator Traditional method Inventive method Unstructured data conversion rate 65% 98% Performance evaluation and actual fit degree 72% 94% Technical risk early warning accuracy 68% 92% Data security vulnerability number 4 0 Early warning response time 48 hours 2 hours The data in Table 1 verifies the advantages of the application in unstructured data processing and risk early warning. The traditional method has a non-structured data conversion rate of only 65% due to the backwardness of OCR and speech transcription technology, and the evaluation fit degree is low due to the lack of fusion of field characteristics; the application realizes 98% conversion through BERT-WWM and YOLOv8, balances subjective and objective factors through the combination weight model, and the fit degree is improved to 94%. The technical risk early warning adopts LSTM time series prediction and dynamic threshold, with an accuracy of 92%, which is 24% higher than the traditional method, and the data security protection realizes zero vulnerability, with the early warning response time shortened to 2 hours. This reflects the synergistic effect of multi-modal data processing, dynamic weight and risk model, and solves the pain points of complex data in basic research projects and lagging risk early warning.

[0041] Embodiment 2: Performance evaluation and risk early warning of intelligent manufacturing application research project This embodiment is aimed at a 2-year "industrial robot precision optimization" application research project. The project is intensive in structured data (20 sets of fund account, 5000+ device usage records), focuses on achievement transformation and fund efficiency, and highlights resource and management risks. The specific implementation is as follows.

[0042] 1. Detailed implementation of technical scheme Data collection and heterogeneous fusion preprocessing: Collect fund payment data (structured) through enterprise ERP system API, obtain running parameters (semi-structured JSON) from device Internet of Things sensors (collect once every 10 minutes), collect related invention literature through KnowNet crawler, and shoot robot operation video (unstructured) through industrial camera. Clean up the fund abnormal expenditure identified by DBSCAN algorithm (clustering radius 0.5), remove duplicate device records by hash check, fill in the missing device maintenance time by improved K nearest neighbor (K=3, combined with device model and running time), and fill in the missing supplier category by weighted mode. Standardize the data to JSON-LD, store the associated identifier "2024YF02-Robot" in ClickHouse warehouse, and encrypt the core data with AES-256 and the general data with DES. The key is rotated every 90 days through Huawei Cloud KMS.

[0043] Multi-dimensional feature deep extraction and optimized fusion: Performance features include invention authorization number (output), device utilization rate (efficiency), engineer work hour contribution rate (contribution), and technology contract amount (transformation); risk features include algorithm iteration failure rate (technology), progress deviation rate (management), fund gap rate (resource), and raw material price increase amplitude (external). Calculate statistical features from structured data, extract technical keyword features from invention text using BERT-WWM, and extract robot trajectory features from operation video using CNN. In the feature standardization formula, The mean value of the intelligent manufacturing field is 65% (device utilization rate), = 12%, ε = 10⁻6 The feature fusion uses a stacked autoencoder, with attention weights allocating 0.3 to "technology contract amount" and 0.28 to "funding gap rate". The "engineer education" feature with a contribution of 2% is excluded, and a 42-dimensional fusion feature set is output (information retention rate 97%).

[0044] Dynamic weight performance evaluation model construction: AHP invites 4 intelligent manufacturing experts to judge the matrix CR=0.07≤0.1, (Conversion index 0.32); Entropy weight method (Efficiency index 0.3); Variance coefficient method (Resource index 0.2). The combined weight formula takes β=0.3, γ=0.45, and 1-β-γ=0.25, resulting in output 0.2, efficiency 0.31, contribution 0.19, and conversion 0.3. The efficiency weight in the project stage is 0.35, and the conversion weight in the conclusion stage is 0.35. The weighted TOPSIS method calculates the score, and the time decay factor is multiplied by 0.85 in the first year.

[0045] Multi-modal risk early warning model construction: management risk uses random forest (tree number 120), resource risk uses support vector machine, and external risk uses Bayesian network. Federal learning accesses 2 enterprises and 1 university, local training for 4 rounds, parameter aggregation adds differential privacy noise (ε=1.2). Dynamic threshold formula =0.35, application research stage λ(s)=1.1, funding gap expansion δ(s)=1.3 in the first 1.5 years, t=1.5 years , triggering a three-level early warning.

[0046] Dynamic monitoring and closed-loop feedback: Collect equipment data every week, update funding data every two weeks, and use a sliding window of 45 days. Incremental learning uses dual models in parallel, and the model F1 value is higher than the historical model by 4% when switching. Version management records 12 parameter updates. Visualization uses Three.js to build a 3D model of device operation, Sankey diagram to show funding flow, and "adjust funding allocation pace" suggestion to management department. Reinforcement learning takes the growth of technology contract amount by 10% as the reward signal, and optimizes the conversion index weight to 0.32.

[0047] 2. Effectiveness verification data Table 2: Comparison of performance evaluation and closed-loop management effect Indicator Traditional method Inventive method Weight allocation rationality 70% 95% Achievement transformation evaluation deviation 18% 4% Resource risk early warning recall rate 71% 93% Closed-loop optimization response cycle 30 days 7 days Expenditure use efficiency improvement 5% 22% The data in Table 2 highlights the advantages of the application in performance evaluation and closed-loop management. The single weight of the traditional method results in a rationality of only 70%, and does not dynamically track the transformation of achievements, with a deviation of 18%; the three-element combined weight of the application combined with stage adjustment, the rationality is improved to 95%, and the weighted TOPSIS method accurately quantifies the transformation value, with a deviation of 4%. The resource risk early warning adopts support vector machine and dynamic threshold, with a recall rate of 93%, which is 22% higher than the traditional method. The closed-loop feedback is stored through reinforcement learning and blockchain, and the response cycle is shortened to 7 days, and the efficiency of fund use is improved by 22%. This verifies the effectiveness of the dynamic weight, federal learning early warning and closed-loop optimization module, and solves the problems of inaccurate application research project achievement transformation evaluation and low resource management efficiency.

[0048] Referring to Figure 2 : The figure directly verifies the core value of the data preprocessing module of the application. The traditional method lacks multi-modal conversion technology, and the conversion rate of unstructured data is generally less than 60%, especially for images and voice, which is only 30%-45%; the application combines technologies such as OCR recognition, BERT semantic extraction, and CNN image analysis to improve the conversion rate of all types of data to more than 95%, of which the text type reaches 98%. Although the processing time of unstructured data is slightly increased, the data integrity is greatly improved, solving the pain point of the traditional method that it is difficult to mine the value of unstructured data. The data reflects the synergistic effect of distributed collection, multi-modal conversion and standardized processing, and lays a foundation of high-quality data for subsequent evaluation and early warning.

[0049] Referring to Figure 3 : The figure reveals the synergistic effect of the multi-modal early warning model and the dynamic threshold. The traditional method has an external environment risk identification accuracy of only 55%, and no risk transmission identification ability; the basic model of the application identifies risks by type (LSTM technology risk, random forest management risk), with an accuracy of 80%-88%, and after superimposing the dynamic threshold formula (fusion stage and trend coefficient), the accuracy of technology and resource risks reaches 92%-94%, and the newly added risk transmission identification accuracy is 90%. This is because the dynamic threshold solves the problem of poor adaptability of the fixed threshold, and the type-based modeling adapts to the complex characteristics of different risks. The data verifies the comprehensiveness and accuracy of the early warning model, covering the blind area of the traditional method.

[0050] Referring to Figure 4The figure shows the continuous optimization value of the closed-loop feedback mechanism. The initial model evaluation accuracy is only 82%, the response period is 48 days, and the compliance is 80%; after 1 feedback (expert correction weight), the accuracy is improved to 88%; 2 feedback (optimize early warning threshold), recall rate reaches 89%; 3 feedback (reinforcement learning adjusts feature weight), accuracy 94%, response period is reduced to 7 days; 4 feedback (perfect security encryption), compliance reaches 100%. This is because the closed-loop process converts evaluation results, early warning effects and expert opinions into optimization signals, and updates the model through incremental learning. Data confirms the effectiveness of the evaluation, early warning, feedback and optimization closed loop, and solves the pain points of model solidification and difficulty in iteration of traditional methods.

[0051] Reference Figure 5 The figure directly reflects the advantages of the data security protection module. The traditional method scores less than 70 points in desensitization accuracy, permission control and other dimensions, and there is a risk of privacy leakage; the present application uses reversible desensitization algorithm (accuracy 98%), TLS1.3 transmission encryption (99%), AES-256 storage encryption (97%), RBAC field-level permission control (96%) and blockchain operation traceability (99%), and the full-dimension score reaches more than 96%. This is because the security protection runs through the whole link from data collection to transmission to storage, and meets the needs of scientific research data "privacy protection and statistical value". Data verifies the effectiveness of the security module, solves the problem of "heavy evaluation and light security" of traditional methods, and meets the scientific research data management specifications.

[0052] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for project performance evaluation and risk early warning for multi-source heterogeneous scientific research data, characterized in that, Includes the following steps: Data Acquisition and Heterogeneous Fusion Preprocessing: Multi-source heterogeneous data from the entire scientific research lifecycle are collected using distributed crawlers and API-linked acquisition. Structured data is extracted using JDBC, semi-structured data is parsed and tagged, and unstructured data is processed using OCR and speech-to-text. In the cleaning process, outlier detection and hash deduplication are used. Numerical missing values ​​are filled with improved K-nearest neighbors, and categorical missing values ​​are filled with weighted mode. Finally, the data is converted into a unified JSON-LD format by a data standardization engine, associated with project identifiers, and stored in a distributed data warehouse. Multidimensional feature deep extraction and optimization fusion: constructing a feature system based on preprocessed data; Feature extraction employs a multi-algorithm collaborative strategy: statistical features are calculated for structured data, key features are extracted by parsing labels for semi-structured data, and corresponding features are extracted for unstructured data using BERT, CNN, and MFCC. feature The fusion process employs a two-tier architecture consisting of a feature layer and a decision layer to obtain a fused feature set. Construction and training of dynamic weighted performance evaluation model: Based on performance characteristics, a three-element combined weighted model of analytic hierarchy process, entropy weight method and coefficient of variation method is constructed. Subjective weights are determined by consistency test, objective weights are obtained by calculating information entropy, and deviations are corrected according to the degree of dispersion. A dynamic weight adjustment mechanism is set to adjust the indicator weights according to the project stage. Multimodal risk warning model construction and threshold optimization: Based on risk characteristics, a hybrid model of deep learning and traditional algorithms is used; a risk transmission path model is constructed; the risk warning threshold is dynamically determined, and an adjustment function is built by combining historical risk data; Full-cycle dynamic monitoring and incremental model updates: Establish a real-time data acquisition link and adjust the update frequency according to project stages; generate incremental feature sets and filter data using a sliding window; update parameters of the performance model using incremental learning and adapt the risk model to new data using transfer learning; Multi-dimensional result output and closed-loop feedback optimization: The visualization engine generates results including a layered dashboard and a three-color warning light display; and generates customized reports according to user roles.

2. The method for project performance evaluation and risk warning for multi-source heterogeneous scientific research data as described in claim 1, characterized in that, This also includes a dual process of standardization and regularization in the multi-dimensional feature depth extraction and optimization fusion step. The standardization process uses an improved Z-score algorithm, and the standardization formula is as follows: in Represents the standardized eigenvalues. Represents the original feature values. This represents the mean of the feature in the j-th item domain. ε represents the standard deviation of the feature in the j-th item domain, and ε represents a very small positive number. This process eliminates the differences between features in different domains and with different dimensions. The regularization process uses L2 regularization to constrain the features.

3. The method for project performance evaluation and risk early warning for multi-source heterogeneous scientific research data as described in claim 1, characterized in that, This also includes a secondary optimization of the ternary combination weights using the entropy weight method in the construction and training steps of the dynamic weighted performance evaluation model. The optimized combination weights are calculated as follows: in Let represent the final combined weight of the j-th performance indicator, β represent the weight coefficient of the analytic hierarchy process (AHP), γ represent the weight coefficient of the entropy weight method, and 1-β-γ represent the weight coefficient of the coefficient of variation method. This represents the subjective weights determined by the analytic hierarchy process (AHP). This represents the objective weight determined by the entropy weight method. This represents the corrected weights determined by the coefficient of variation method. This method balances the influence of subjective experience, data information, and the degree of feature dispersion, making the weight allocation more in line with the complexity and dynamism of scientific research projects.

4. The method for project performance evaluation and risk warning for multi-source heterogeneous scientific research data as described in claim 1, characterized in that, In the data acquisition and heterogeneous fusion preprocessing steps, multimodal semantic alignment technology is used to transform unstructured data into structured data. Text-type unstructured data is segmented, labeled with parts of speech, and identified using the BERT-WWM model. The entity extraction accuracy is optimized by combining a domain dictionary, and key entities are extracted and structured. Image-type unstructured data is identified using the YOLOv8 algorithm, and image semantic segmentation technology is used to extract chart data. Audio data is extracted for speech features through convolutional neural networks, converted into text using an end-to-end speech-to-text model, and then key information is extracted through entity recognition. After multimedia data conversion, it is associated through a knowledge graph to construct a semantic network of projects, personnel, results, and equipment.

5. The method for project performance evaluation and risk early warning for multi-source heterogeneous scientific research data according to claim 1, characterized in that, In the multidimensional feature deep extraction and optimization fusion step, the feature fusion algorithm adopts a stacked autoencoder enhanced by attention mechanism. The encoder part includes an input layer, a hidden layer and an encoding layer. The input layer receives high-dimensional original features, the hidden layer extracts local features through convolution kernels, and the encoding layer uses an attention weight matrix to assign differentiated weights to different features. The decoder reconstructs features through deconvolution operations and calculates the correlation between reconstruction error and features. A feature validity verification mechanism is introduced during the fusion process, and variance analysis and significance tests are performed on the fused features.

6. The method for project performance evaluation and risk warning for multi-source heterogeneous scientific research data according to claim 1, characterized in that, In the multimodal risk warning model construction and threshold optimization steps, a federated learning framework is used to train the risk warning sub-model. To meet the data privacy requirements of different research units, each unit trains the sub-model locally. The server uses the FedAvg algorithm to aggregate parameters and introduces a differential privacy mechanism to add noise perturbation to the parameters. During model training, a dynamic learning rate strategy is adopted, which is adaptively adjusted according to changes in the model's loss function. The learning rate is reduced when the loss decreases slowly. The risk warning system employs a multi-level threshold system.

7. The method for project performance evaluation and risk warning for multi-source heterogeneous scientific research data as described in claim 1, characterized in that, In the full-cycle dynamic monitoring and incremental model update process, a data quality assessment system is constructed to evaluate the quality of newly added data from four dimensions: accuracy, completeness, consistency, and timeliness. Accuracy is verified by comparison with authoritative data sources; completeness is calculated by the proportion of missing values; consistency is checked for data conflicts across data sources; and timeliness is determined by the difference between the data update time and the current time. A data quality scoring mechanism is adopted. The incremental model update adopts a dual-model parallel mechanism, running the historical model and the updated model simultaneously, and comparing the output results of the two through A / B testing. A model version management mechanism is established to record the parameters, data, and evaluation indicators of each update, and to support model version backtracking.

8. The method for project performance evaluation and risk early warning for multi-source heterogeneous scientific research data according to claim 1, characterized in that, In the multi-dimensional result output and closed-loop feedback optimization steps, the visualization system adopts a front-end and back-end separation architecture. The front-end is based on ECharts and Three.js to build an interactive interface, and the back-end pushes updated data in real time through data interfaces. The performance display is divided into multiple scenarios. The project progress scenario uses Gantt charts to show the completion status of time nodes, the funding usage scenario uses Sankey charts to show the funding flow path, and the results distribution scenario uses geographic heat maps to show the results transformation areas. The risk warning adds a risk evolution prediction function. The feedback optimization introduces a reinforcement learning mechanism, using the implementation effect of improvement suggestions as a reward signal.

9. The method for project performance evaluation and risk warning for multi-source heterogeneous scientific research data according to claim 1, characterized in that, In the construction and threshold optimization steps of the multimodal risk warning model, the dynamic adjustment of the risk warning threshold adopts a combination of time series analysis and domain adaptation, specifically through... Confirmed, among which This represents the dynamic risk warning threshold at time t. λ(s) represents the initial threshold, λ(s) represents the project stage impact coefficient at time s, δ(s) represents the risk trend coefficient at time s, which is determined by the trend slope calculated based on historical risk data, and t represents the time the project has been implemented.

10. The method for project performance evaluation and risk early warning for multi-source heterogeneous scientific research data according to claim 1, characterized in that, In the data acquisition and heterogeneous fusion preprocessing steps, a data security protection step has been added, and data desensitization technology is used to process sensitive information. Data transmission is encrypted using the TLS 1.3 protocol; data storage employs a partitioned encryption strategy, with core data encrypted using the AES-256 algorithm and ordinary data encrypted using the DES algorithm, and keys are rotated periodically through a key management system; in addition, a data access control system is established, allocating access permissions based on the RBAC model, refining access control down to the field level, and recording all data operation logs, in compliance with scientific research data security management standards.

Images