A machine learning-based project data quality checking method and system

Abstract

The application provides an artificial intelligence and project data management cross technical field, and relates to a project data quality checking method and system based on machine learning, which comprises the following steps: S1, a large amount of historical project data is acquired to construct a data set, a data quality checking model is trained, and the trained data quality checking model is deployed as a micro service; S2, real-time project data acquired is input into the deployed data quality checking model through the micro service, and a quality checking report carrying abnormal data and abnormal root causes is obtained; S3, real-time project data is intercepted and warned based on the quality checking report; S4, a checking log including at least real-time project data, a quality checking report, checking result feedback and checking time is recorded in real time, and an incremental data set is constructed based on the checking log to iteratively optimize the data quality checking model.The application has the advantages that the expansibility, timeliness and reliability of project data quality checking are greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and project data management, and particularly discloses a project data quality checking method and system based on machine learning. BACKGROUND

[0002] In a modern project management system, project initiation as an initial stage generates a large amount of multi-dimensional project data. These data run through the entire life cycle of the project, including but not limited to project data (such as project objectives, budget estimates, and risk assessments), procurement data (such as supplier information, bidding records, and purchase orders), contract data (such as contract terms, performance status, and change records), implementation data (such as progress reports, resource allocation, and quality control indicators), etc. These project data not only support daily decision-making, but also directly affect project quality, cost, and risk control. For example, in large infrastructure or IT projects, low-quality data (such as missing key fields or logical errors) may lead to project delays, cost overruns, and even failures. Therefore, project data quality checking is a key link to ensure project success. By systematically checking the completeness, accuracy, and consistency of project data, potential problems (such as resource conflicts or compliance risks) can be identified in advance, thereby improving the overall robustness of the project.

[0003] Traditional project data quality checking methods mainly rely on static rule bases. Static rule bases are usually based on pre-defined rules (such as data format checking, field mandatory checking, or logical dependency verification). These rules are implemented through hard coding, which cannot dynamically adapt to schema evolution of project data. For example, when project requirements change, resulting in the addition of new fields (such as environmental compliance indicators) or the adjustment of logical dependencies (such as the update of association rules between budget and progress), the static rule base needs to be manually restructured, which is not only time-consuming but also prone to errors, resulting in incomplete coverage or high false positive rates. More seriously, traditional methods mostly use batch processing, i.e., centralized processing after project data is generated (such as daily or weekly batch scanning), rather than real-time interception. This delayed mechanism cannot immediately intercept quality problems (such as missing key fields or abnormal values) when data streams enter the system, causing problem data to flow into downstream systems and amplify risks. For example, in cross-department projects, missing contract data in real-time data streams may trigger incorrect payments if not checked in time, resulting in high remediation costs.

[0004] Moreover, project data is often scattered across heterogeneous systems (such as ERP, CRM, or SCM platforms), forming data silos, which makes cross-system data verification inefficient. Traditional methods rely on point-to-point integration or ETL (Extract, Transform, Load) tools for data pulling, but are slow and resource-intensive, and lack a unified view. For example, when verifying the consistency of procurement data and contract data, data needs to be extracted from multiple systems, resulting in high latency and low throughput, which cannot meet the timeliness requirements of agile project management. At the same time, existing technologies lack the ability to locate the root cause of anomalies (root cause analysis), when anomalies (such as data conflicts or logical errors) are found during verification, only surface problems (such as missing fields) can be reported, and the source (such as system interface failure or human input error) cannot be traced back, which limits the depth of problem solving, prolongs the troubleshooting cycle, and affects project continuity.

[0005] These defects are due to the inherent limitations of traditional technology: the rigid structure of the static rule base does not adapt to the dynamic business environment; post-batch verification cannot support real-time stream processing; the complexity of cross-system integration exacerbates data consistency problems; and the lack of root cause analysis is due to the shallow logic of the rule base, which cannot simulate complex dependencies.

[0006] Therefore, how to provide a project data quality verification method and system based on machine learning to improve the scalability, timeliness, and reliability of project data quality verification has become a technical problem to be solved. SUMMARY

[0007] The technical problem to be solved by the present application is to provide a project data quality verification method and system based on machine learning to improve the scalability, timeliness, and reliability of project data quality verification.

[0008] In a first aspect, the present application provides a project data quality verification method based on machine learning, comprising the following steps:

[0009] Step S1, creating a data quality verification model based on a data anomaly detection module, a root cause positioning module, and an output module, and setting a loss function of the data quality verification model;

[0010] The data anomaly detection module is constructed based on a project initiation data detection unit, a procurement data detection unit, a contract data detection unit, an implementation data detection unit, and a cross-domain feature fusion unit; the project initiation data detection unit is configured to extract project semantic anomaly features from project initiation data through a first Transformer encoder, extract project timing anomaly features from the project initiation data through a first timing decomposition subunit, and output project anomaly features containing the project semantic anomaly features and the project timing anomaly features; the procurement data detection unit is configured to extract procurement semantic anomaly features from procurement data through a second Transformer encoder, extract procurement timing anomaly features from the procurement data through a second timing decomposition subunit, and output procurement anomaly features containing the procurement semantic anomaly features and the procurement timing anomaly features; the contract data detection unit is configured to extract contract semantic anomaly features from contract item data through a third Transformer encoder, extract contract timing anomaly features from the contract data through a third timing decomposition subunit, and output contract anomaly features containing the contract semantic anomaly features and the contract timing anomaly features; the implementation data detection unit is configured to extract implementation semantic anomaly features from implementation item data through a fourth Transformer encoder, extract implementation timing anomaly features from the implementation data through a fourth timing decomposition subunit, and output implementation anomaly features containing the implementation semantic anomaly features and the implementation timing anomaly features; and the cross-domain feature fusion unit is configured to calculate cross-weights of the project anomaly features, the procurement anomaly features, the contract anomaly features, and the implementation anomaly features respectively through a gated attention mechanism, fuse the project anomaly features, the procurement anomaly features, the contract anomaly features, and the implementation anomaly features based on the cross-weights, and output a fusion anomaly feature vector.

[0011] The root cause positioning module is constructed based on a data graph construction unit and a causal reasoning unit; the data graph construction unit is configured to construct a project data graph by taking project entities in project initiation data, procurement data, contract data, and implementation data as nodes and data dependency relationships as edges through a dynamic graph neural network; and the causal reasoning unit is configured to infer the fusion anomaly feature vector through a causal graph convolution network and randomly walk on the project data graph to locate an abnormal path.

[0012] The output module is configured to output a quality review report carrying abnormal data and abnormal root causes according to the fusion anomaly feature vector and the abnormal path.

[0013] In step S2, a large amount of historical project data is acquired, the historical project data is preprocessed to construct a data set, and a sample expansion operation is performed on the data set.

[0014] Step S3, training the data quality checking model through the data set and the loss function, and deploying the trained data quality checking model as a microservice;

[0015] Step S4, obtaining real-time project data from a distributed project heterogeneous system, inputting the real-time project data into the deployed data quality checking model through the microservice, and obtaining a quality checking report carrying abnormal data and abnormal root causes;

[0016] Step S5, intercepting and warning the real-time project data based on the quality checking report;

[0017] Step S6, recording a checking log including at least the real-time project data, the quality checking report, the checking result feedback, and the checking time, constructing an incremental data set based on the checking log to iteratively optimize the data quality checking model, and encrypting and storing the checking log.

[0018] Further, in the step S1, the formula of the loss function is:

[0019] L = λ1*L det + λ2*L root ;

[0020] Wherein, L represents the loss value of the loss function; L det represents the abnormality detection loss, which adopts a split-type focal loss; L root represents the root cause positioning loss, which adopts a path similarity comparison loss; λ1 and λ2 both represent weight coefficients.

[0021] Further, the step S2 specifically comprises:

[0022] Obtaining a large amount of historical project data including historical project data, historical procurement data, historical contract data, and historical implementation data, preprocessing each of the historical project data including at least data integration, missing value processing, duplicate value processing, data standardization, and data normalization, labeling each of the preprocessed historical project data for abnormal data and abnormal root causes, and constructing a data set based on each of the labeled historical project data;

[0023] Performing sample expansion operations of noise injection, data translation, and data scaling on numerical data in the data set, performing sample expansion operations of time translation, time scaling, and time reversal on time series data, and performing sample expansion operations of synonym replacement, sentence reorganization, and random insertion on text data.

[0024] Further, the step S3 specifically comprises:

[0025] The data set is divided into a training set, a validation set and a test set in a ratio of 8:1:1 by K-fold cross-validation method, the data quality checking model is trained by the training set, and the data quality checking model is continuously optimized during the training process at least including learning rate, random dropout rate, batch size, number of hidden layers, activation function, optimizer hyperparameters, until the loss value of the loss function is less than the preset loss threshold;

[0026] The trained data quality checking model is verified by the validation set to determine whether the checking accuracy is greater than the preset accuracy threshold, if not, the verification fails, and the training set is expanded for further training; if yes, the verification passes, and:

[0027] The precision, recall, F1 score and confidence are calculated by the test set to test the data quality checking model that passes the verification, if the test fails, the training set is expanded for further training; if the test passes, the training ends, and the data quality checking model is deployed as a microservice.

[0028] Further, the step S4 is specifically:

[0029] Real-time project data including real-time project data, real-time procurement data, real-time contract data and real-time implementation data are obtained from distributed project heterogeneous systems including project subsystem, procurement subsystem, contract subsystem and implementation subsystem through a stream processing framework;

[0030] The real-time project data is input into the deployed data quality checking model through the microservice, the data quality checking model performs inference through hardware acceleration technology, and a quality checking report carrying abnormal data and abnormal root cause is obtained;

[0031] The step S5 is specifically:

[0032] The quality checking report is monitored in real time, when the value of abnormal data or abnormal root cause in the quality checking report is not empty, the project data associated with the abnormal data or abnormal root cause is intercepted, a warning notification is generated based on the quality checking report, the warning notification is encrypted into an encrypted notification through a first encryption rule, and the encrypted notification is pushed to a pre-associated management terminal in real time through a TLS protocol to perform a warning operation;

[0033] The first encryption rule is specifically:

[0034] The hash value of the early warning notice is calculated through an SM3 algorithm, the early warning notice is converted into binary data according to UTF-8 encoding, the binary data is encrypted through an SM4 algorithm to obtain first-level encrypted data, the first-level encrypted data is spliced with the hash value to obtain spliced data, the spliced data is converted into a hexadecimal string, and the numbers 1, 2, 3 and 4 in the hexadecimal string are respectively mapped to the letters B, C, D and F to obtain second-level encrypted data, and the second-level encrypted data is encrypted into an encrypted notice through an ECC algorithm;

[0035] The step S6 is specifically:

[0036] The real-time record at least includes the real-time project data, the quality check report, the check result feedback and the check log of the check time, the incremental data set is constructed based on the check log, when the data amount of the incremental data set exceeds a preset number threshold, the data quality check model is backed up, the data quality check model is trained through the incremental data set, the performance of the trained data quality check model is verified, if the verification is passed, the training is ended, if the verification is not passed, the backed up data quality check model is restored, the data quality check model is iteratively optimized, and the check log is encrypted into an encrypted log through a second encryption rule, the encrypted log is stored and distributedly backed up;

[0037] The second encryption rule is specifically:

[0038] The MAC value of the check log is calculated through an HMAC algorithm, the check log and the MAC value are encrypted through an SM9 algorithm to obtain one-layer encrypted data, the one-layer encrypted data is equally divided into a first data block and a second data block, each character of the first data block is cyclically shifted to the left by 7 bits to obtain a first obfuscated data block, each character of the second data block is cyclically shifted to the right by 4 bits to obtain a second obfuscated data block, the first obfuscated data block and the second obfuscated data block are encrypted into two-layer encrypted data through an RC6 algorithm, and the two-layer encrypted data is split in a ratio of 1:5 and the front and rear orders are exchanged to obtain the encrypted log.

[0039] In a second aspect, the present application provides a project data quality check system based on machine learning, comprising the following modules:

[0040] The data quality check model creation module is used for creating a data quality check model based on the data anomaly detection module, the root cause positioning module and the output module, and setting a loss function of the data quality check model.

[0041] The data anomaly detection module is constructed based on a project data detection unit, a procurement data detection unit, a contract data detection unit, an implementation data detection unit, and a cross-domain feature fusion unit; the project data detection unit is configured to extract project semantic anomaly features from project data through a first Transformer encoder, extract project time sequence anomaly features from the project data through a first time sequence decomposition subunit, and output project anomaly features containing the project semantic anomaly features and the project time sequence anomaly features; the procurement data detection unit is configured to extract procurement semantic anomaly features from procurement data through a second Transformer encoder, extract procurement time sequence anomaly features from the procurement data through a second time sequence decomposition subunit, and output procurement anomaly features containing the procurement semantic anomaly features and the procurement time sequence anomaly features; the contract data detection unit is configured to extract contract semantic anomaly features from contract data through a third Transformer encoder, extract contract time sequence anomaly features from the contract data through a third time sequence decomposition subunit, and output contract anomaly features containing the contract semantic anomaly features and the contract time sequence anomaly features; the implementation data detection unit is configured to extract implementation semantic anomaly features from implementation data through a fourth Transformer encoder, extract implementation time sequence anomaly features from the implementation data through a fourth time sequence decomposition subunit, and output implementation anomaly features containing the implementation semantic anomaly features and the implementation time sequence anomaly features; and the cross-domain feature fusion unit is configured to calculate cross-weights of the project anomaly features, the procurement anomaly features, the contract anomaly features, and the implementation anomaly features respectively through a gated attention mechanism, fuse the project anomaly features, the procurement anomaly features, the contract anomaly features, and the implementation anomaly features based on the cross-weights, and output a fusion anomaly feature vector.

[0042] The root cause positioning module is constructed based on a data graph construction unit and a causal reasoning unit; the data graph construction unit is configured to construct a project data graph by taking project entities in project data, procurement data, contract data, and implementation data as nodes and data dependency relationships as edges through a dynamic graph neural network; and the causal reasoning unit is configured to infer the fusion anomaly feature vector through a causal graph convolution network and randomly walk on the project data graph to locate an abnormal path.

[0043] The output module is configured to output a quality review report carrying abnormal data and abnormal root causes according to the fusion anomaly feature vector and the abnormal path.

[0044] A data set construction module is configured to obtain a large amount of historical project data, construct a data set after preprocessing each of the historical project data, and perform a sample expansion operation on the data set.

[0045] The data quality checking model training and deployment module is configured to train a data quality checking model by using the data set and a loss function, and deploy the trained data quality checking model as a microservice.

[0046] The project data quality checking module is configured to obtain real-time project data from a distributed project heterogeneous system, input the real-time project data into the deployed data quality checking model through the microservice, and obtain a quality checking report carrying abnormal data and abnormal root causes.

[0047] The project data interception and early warning module is configured to intercept and early warn the real-time project data based on the quality checking report.

[0048] The checking log management module is configured to record a checking log in real time, the checking log at least including the real-time project data, the quality checking report, checking result feedback, and checking time, build an incremental data set based on the checking log to iteratively optimize the data quality checking model, and encrypt and store the checking log.

[0049] Further, in the data quality checking model creation module, the formula of the loss function is as follows:

[0050] L = λ1*L det + λ2*L root ;

[0051] wherein L represents a loss value of the loss function, L det represents an abnormality detection loss, and a split-type focal loss is used; L root represents a root cause positioning loss, and a path similarity comparison loss is used; and λ1 and λ2 represent weight coefficients.

[0052] Further, the data set construction module is specifically configured to:

[0053] obtain a large amount of historical project data including historical project data, historical procurement data, historical contract data, and historical implementation data, pre-process the historical project data by at least including data integration, missing value processing, duplicate value processing, data standardization, and data normalization, label the pre-processed historical project data for abnormal data and abnormal root causes, and construct a data set based on the labeled historical project data.

[0054] perform sample expansion operations of noise injection, data translation, and data scaling on numerical data in the data set, perform sample expansion operations of time translation, time scaling, and time reversal on time series data, and perform sample expansion operations of synonym replacement, sentence reorganization, and random insertion on text data.

[0055] Further, the data quality checking model training and deployment module is specifically configured to:

[0056] The data set is divided into a training set, a validation set and a test set in a ratio of 8:1:1 through K-fold cross-validation method, the data quality checking model is trained through the training set, and at least hyperparameters of the data quality checking model including learning rate, random dropout rate, batch size, number of hidden layers, activation function and optimizer are continuously optimized during the training process until the loss value of the loss function is less than a preset loss threshold;

[0057] The trained data quality checking model is verified through the validation set to determine whether the checking accuracy is greater than a preset accuracy threshold, if not, the verification fails, and the training set is expanded for continuous training; if yes, the verification passes, and:

[0058] The precision, recall, F1 score and confidence are calculated through the test set to test the data quality checking model that passes the verification, if the test fails, the training set is expanded for continuous training; if the test passes, the training ends, and the data quality checking model is deployed as a microservice.

[0059] Further, the project data quality checking module is specifically configured to:

[0060] Real-time project data including real-time project data, real-time procurement data, real-time contract data and real-time implementation data is obtained from a distributed project heterogeneous system including a project subsystem, a procurement subsystem, a contract subsystem and an implementation subsystem through a stream processing framework;

[0061] The real-time project data is input into the deployed data quality checking model through the microservice, the data quality checking model performs reasoning through hardware acceleration technology to obtain a quality checking report carrying abnormal data and abnormal root causes;

[0062] The project data interception and early warning module is specifically configured to:

[0063] The quality checking report is monitored in real time, when the abnormal data or the abnormal root cause in the quality checking report has a non-empty value, the project data associated with the abnormal data or the abnormal root cause is intercepted, an early warning notification is generated based on the quality checking report, the early warning notification is encrypted into an encrypted notification through a first encryption rule, and the encrypted notification is pushed to a pre-associated management terminal in real time through a TLS protocol to perform an early warning operation;

[0064] The first encryption rule is specifically:

[0065] The hash value of the early warning notice is calculated through an SM3 algorithm, the early warning notice is converted into binary data through UTF-8 encoding, the binary data is encrypted through an SM4 algorithm to obtain first-level encrypted data, the first-level encrypted data is spliced with the hash value to obtain spliced data, the spliced data is converted into a hexadecimal string, and the numbers 1, 2, 3 and 4 in the hexadecimal string are respectively mapped to the letters B, C, D and F to obtain second-level encrypted data, and the second-level encrypted data is encrypted into an encrypted notice through an ECC algorithm.

[0066] The verification log management module is specifically used for:

[0067] The verification log at least includes the real-time project data, the quality verification report, the verification result feedback and the verification time, an incremental data set is constructed based on the verification log, the data quality verification model is backed up when the data amount of the incremental data set exceeds a preset amount threshold, the data quality verification model is trained through the incremental data set, the trained data quality verification model is subjected to performance verification, the training is ended when the verification is passed, the backup data quality verification model is restored when the verification is not passed, the data quality verification model is iteratively optimized, and the verification log is encrypted into an encrypted log through a second encryption rule, the encrypted log is stored and distributedly backed up.

[0068] The second encryption rule is specifically:

[0069] The MAC value of the verification log is calculated through an HMAC algorithm, the verification log and the MAC value are encrypted through an SM9 algorithm to obtain one layer of encrypted data, the one layer of encrypted data is equally divided into a first data block and a second data block, each character of the first data block is cyclically shifted to the left by 7 bits to obtain a first obfuscated data block, each character of the second data block is cyclically shifted to the right by 4 bits to obtain a second obfuscated data block, the first obfuscated data block and the second obfuscated data block are encrypted through an RC6 algorithm to obtain two layers of encrypted data, the two layers of encrypted data are split in a ratio of 1:5 and the front and rear orders are exchanged to obtain the encrypted log.

[0070] The application has the advantages that:

[0071] 1. A data quality verification model is created through a data anomaly detection module, a root cause localization module, and an output module, and a loss function is set for the data quality verification model. Then, a large amount of historical project data is acquired, preprocessed, and used to construct a dataset, which is then augmented. Next, the data quality verification model is trained using the dataset and the loss function, and deployed as a microservice. Then, real-time project data is obtained from a distributed heterogeneous project system, and input into the deployed data quality verification model via the microservice to obtain a quality verification report containing abnormal data and root causes. Based on the quality verification report, real-time project data is intercepted and alerted. A verification log, including at least the real-time project data, the quality verification report, verification result feedback, and verification time, is recorded in real time. An incremental dataset is constructed based on the verification log to iteratively optimize the data quality verification model, and the verification log is encrypted and stored. In other words, project data quality verification is performed using a pre-trained data quality verification model instead of a traditional static rule base. Combined with iterative optimization of the data quality verification model, it can… The system dynamically adapts to changes in project data patterns; it obtains real-time project data from distributed heterogeneous project systems, inputs it into the data quality verification model for inference, and outputs a quality verification report. Based on the quality verification report, it performs real-time interception and early warning, realizing streaming processing rather than post-event batch verification, ensuring that problematic data is intercepted in a timely manner and preventing it from flowing into downstream systems. It obtains real-time project data from heterogeneous project systems through a streaming processing framework, calculates and merges the cross-weights of different data sources through the cross-domain feature fusion unit (gated attention mechanism) of the data anomaly detection module, providing a unified view; the data graph construction unit (dynamic graph neural network) of the root cause localization module integrates cross-system entities and dependencies, reducing the resource consumption and latency of traditional ETL or point-to-point integration, and improving verification efficiency and throughput; through the causal graph convolutional network of the root cause localization module, it randomly walks on the project data graph to locate abnormal paths, outputs a quality verification report carrying the root cause of the anomaly, and performs early warning and optimization based on the root cause of the anomaly, surpassing traditional surface anomaly detection, enabling the tracing of the source of the problem, and ultimately greatly improving the scalability, timeliness, and reliability of project data quality verification.

[0072] 2. Multiple dedicated detection units (detection units for project initiation data, procurement data, contract data, and implementation data) process different types of project data respectively, and integrate semantic and temporal features (for example, the Transformer encoder extracts semantic anomaly features, and the temporal decomposition subunit extracts temporal anomaly features), covering the entire life cycle of project data, avoiding the limitations of a single data source, and improving the completeness of anomaly detection; multimodal feature extraction (semantic + temporal) captures the structure and dynamic changes of project data, improving detection accuracy, especially in handling complex project scenarios, it can identify potential hidden anomalies.

[0073] 3. The cross-domain feature fusion unit uses a gated attention mechanism to calculate and weight the fusion of abnormal features from different data types (such as project abnormal features, procurement abnormal features, etc.), which can adaptively adjust the cross-weight, highlight the correlation between important features, solve the data island problem, enhance the sensitivity of abnormal detection (for example, identify complex abnormalities caused by cross-department data dependence), and reduce the false positive rate, effectively improving the detection efficiency compared with traditional simple splicing or average fusion methods.

[0074] 4. The root cause positioning module provides end-to-end root cause tracking (abnormal path output) through the data graph construction unit (using a dynamic graph neural network to construct project entity nodes and data dependency edges) and the causal reasoning unit (using a causal graph convolution network to infer and random walk to locate abnormal paths), enhancing the explainability of the output (users can understand the source of the abnormality); combined with causal reasoning, it can locate deep-rooted causes (such as contract data problems leading to implementation abnormalities), rather than just surface abnormalities, which is beneficial for subsequent decision-making and problem eradication, and improves the trustworthiness of the system.

[0075] 5. The integration of abnormal detection and root cause positioning tasks is achieved through a custom loss function (L = λ1*L det + λ2*L root ), where the focal loss handles data imbalance (such as minority class abnormal samples), improving detection robustness; the path similarity loss optimizes the accuracy of root cause positioning; the weight coefficient can be adjusted to balance task priorities, speed up model convergence, and improve overall model performance.

[0076] 6. Through full-process automation covering data collection, model training, real-time deployment (microservices) to application response, reducing manual intervention; through sample expansion to effectively improve model generalization ability; microservice architecture facilitates deployment and expansion, supports high-concurrency real-time processing, ensures low-latency response (milliseconds), and improves operational efficiency.

[0077] 7、By integrating multi-modal feature extraction (using Transformer encoder and temporal decomposition unit to capture semantic and temporal anomaly features of project, procurement, contract and implementation data), gated attention mechanism driven cross-domain feature fusion (adaptive weighted fusion of multi-source anomaly features), and root cause positioning based on dynamic graph neural network and causal reasoning (constructing project data graph and locating abnormal path through random walk), high-precision anomaly detection and explainable root cause analysis of project life cycle data are realized; the end-to-end automated process (from historical data training, micro-service real-time deployment to distributed system access) combined with multi-task loss function optimization (weighted balance of focal loss and path similarity contrast loss) significantly improves detection efficiency and model robustness; at the same time, through real-time interception and early warning, incremental data iterative optimization and encrypted log storage, a closed-loop quality control system is formed, which has millisecond-level response capability, continuous adaptive evolution and industrial-level security compliance, and fully solves the core pain points of incomplete coverage, high false positive rate, difficult root cause tracing and delayed response in project data quality review.

[0078] 8、For different data types (numeric, time series, text), customized sample expansion operations are used: noise injection, data translation and scaling for numeric data (such as procurement amount), which increases data diversity and prevents machine learning model overfitting; time translation, scaling and inversion for time series data (such as project implementation timeline), which simulates real-world time changes and enhances the robustness of time series models (such as predicting project delays); synonym replacement, sentence restructuring and random insertion for text data (such as contract terms), which enriches language expression and improves natural language processing model understanding; this type-sensitive expansion method has significant innovation, avoiding the limitations of "one-size-fits-all" expansion, effectively solving small sample data problems and improving model generalization performance in real-world scenarios.

[0079] 9、Through systematic data preprocessing (including integration, missing value processing, standardization, etc.), the quality and consistency of historical project data are significantly improved, combined with precise labeling of abnormal data and root causes, a highly reliable data set is constructed; further, customized sample expansion strategies (such as noise injection, time translation, synonym replacement, etc.) are designed for numeric, time series and text data, effectively enhancing data diversity, solving the model overfitting problem in small sample scenarios, and significantly improving the generalization ability and prediction accuracy of machine learning models, while the automated process reduces the cost of manual intervention, providing efficient and scalable data support for project risk assessment, decision optimization and other advanced applications.

[0080] 10、By adopting K-fold cross-validation combined with an 8:1:1 data division strategy, the generalization ability and robustness of the model are significantly improved, effectively avoiding overfitting. In the training process, key hyperparameters such as learning rate and random dropout rate are dynamically optimized, and the training is terminated by a loss function threshold, greatly improving the parameter tuning efficiency and resource utilization. At the same time, a dual protection mechanism of verification and testing is designed - in the verification stage, the model is screened by accuracy threshold, and in the testing stage, multiple-dimensional indicators such as precision, recall, F1 score and confidence are comprehensively evaluated to ensure the reliability of model decision-making. The closed-loop feedback mechanism of dynamically expanding the training set can adaptively optimize the data deficiency problem, and finally realize high scalability and low operation and maintenance cost through micro-service deployment, forming an efficient, stable and iterative data quality verification model development paradigm.

[0081] 11、Through the real-time integration of multi-source project data of distributed heterogeneous systems (including project initiation, procurement, contract and implementation subsystems) by the stream processing framework, and the use of hardware-accelerated micro-service model for efficient data quality verification, the real-time performance and cross-system compatibility of data processing are significantly improved, breaking the data silos and realizing global monitoring. Relying on hardware acceleration technology to optimize model reasoning efficiency, the computing delay and resource consumption are greatly reduced. Synchronous output of abnormal data and root cause analysis enhances the accuracy of problem diagnosis and decision support capability. Finally, through the design of an automated and scalable architecture, operation and maintenance costs are reduced and project risks are prevented, forming an efficient, accurate and economical integrated project data quality verification solution.

[0082] 12、By real-time monitoring of quality verification reports, automatically intercepting related data and generating warnings when anomalies are triggered, using a multi-level encryption mechanism (combining SM3 hash verification, SM4 symmetric encryption, character mapping optimization and ECC asymmetric encryption) to ensure data security and integrity, and combining with the TLS protocol to push to the management terminal in real time, the efficiency of abnormal response and the level of processing automation are significantly improved. At the same time, through the combination and transmission encryption of national encryption algorithms, end-to-end security protection is achieved, effectively solving the problems of response lag, high dependence on manual intervention and high risk of data leakage in traditional monitoring systems, providing a high-reliability, low-latency closed-loop solution for data quality management in high-sensitivity fields.

[0083] 13、By building an incremental log containing complete verification elements to realize closed-loop iterative optimization of the data quality model, innovatively using a threshold trigger mechanism to ensure the stability and computational efficiency of model updates, and designing a four-layer nested encryption system - combining the national SM9 algorithm to verify data integrity, asymmetric character bit shift confusion to build a proactive defense layer, RC6 algorithm for deep encryption, and 1:5 ratio data reordering to form the final ciphertext, on the basis of ensuring the continuous evolution of data quality management, the problem of single-point encryption vulnerability in sensitive log protection in traditional solutions is overcome, and the real-time project data quality monitoring capability and attack resistance are significantly improved.

[0084] 14、Through constructing machine learning model of fusion multi-source data anomaly detection (project, procurement, contract, implementation) and root cause positioning, combining with Transformer encoder to extract semantic features and time decomposition to capture dynamic anomalies, using gated attention mechanism to realize cross-domain feature fusion, and through dynamic graph neural network to build data graph for causal reasoning, the comprehensiveness of anomaly detection and the accuracy of root cause positioning are significantly improved; The end-to-end automated process (from data preprocessing, sample expansion, model training to microservice deployment) supports real-time stream processing and hardware acceleration, efficiently outputs quality check report carrying abnormal data and root cause, realizes active interception and multi-layer encryption early warning (such as SM4+ECC), at the same time, through incremental learning to continuously optimize the model, and using national secret algorithm (SM3 / SM9 / RC6) and confusion technology to strengthen log security, with the core advantages of high robustness, strong security, scalability and low operation and maintenance cost. BRIEF DESCRIPTION OF DRAWINGS

[0085] The application will be further described below with reference to the accompanying drawings in conjunction with the embodiments.

[0086] Fig. 1 is a flow chart of a project data quality checking method based on machine learning.

[0087] Fig. 2 is a structural schematic diagram of a project data quality checking system based on machine learning. DETAILED DESCRIPTION

[0088] The technical solutions in the embodiments of the present application have the following general idea: the project data quality checking is performed by using a pre-trained data quality checking model to replace the traditional static rule base, and the iteration optimization of the data quality checking model can dynamically adapt to the mode change of the project data; real-time project data is obtained from a distributed project heterogeneous system, input into the data quality checking model for reasoning, and a quality checking report is output, real-time interception and early warning are performed based on the quality checking report, streaming processing is realized instead of post-batch checking, problem data is ensured to be intercepted in time, and the problem data is prevented from flowing into a downstream system; real-time project data is obtained from the project heterogeneous system through a streaming processing framework, cross-weight of different data sources is calculated and fused through a cross-domain feature fusion unit (a gating attention mechanism) of a data anomaly detection module, and a unified view is provided; a data graph construction unit (a dynamic graph neural network) of a root cause positioning module integrates cross-system entities and dependency relationships, reduces resource consumption and delay of the traditional ETL or point-to-point integration, and improves checking efficiency and throughput; a causal graph convolution network of the root cause positioning module randomly walks on a project data graph, locates an abnormal path, and outputs a quality checking report carrying an abnormal root cause, early warning and optimization are performed based on the abnormal root cause, and the traditional surface anomaly detection is surpassed, the problem source can be traced back, and then the expansibility, timeliness and reliability of the project data quality checking are improved.

[0089] Please refer to Figs. 1-2 The preferred embodiment of the project data quality checking method based on machine learning provided by the present application comprises the following steps:

[0090] In step S1, a data quality checking model is created based on a data anomaly detection module, a root cause positioning module and an output module, and a loss function of the data quality checking model is set;

[0091] The data anomaly detection module is constructed based on a project initiation data detection unit, a procurement data detection unit, a contract data detection unit, an implementation data detection unit, and a cross-domain feature fusion unit. The project initiation data detection unit is configured to extract project semantic anomaly features from project initiation data through a first Transformer encoder, extract project temporal anomaly features from the project initiation data through a first temporal decomposition subunit, and output project anomaly features including the project semantic anomaly features and the project temporal anomaly features. The procurement data detection unit is configured to extract procurement semantic anomaly features from procurement data through a second Transformer encoder, extract procurement temporal anomaly features from the procurement data through a second temporal decomposition subunit, and output procurement anomaly features including the procurement semantic anomaly features and the procurement temporal anomaly features. The contract data detection unit is configured to extract contract semantic anomaly features from contract item data through a third Transformer encoder, extract contract temporal anomaly features from the contract data through a third temporal decomposition subunit, and output contract anomaly features including the contract semantic anomaly features and the contract temporal anomaly features. The implementation data detection unit is configured to extract implementation semantic anomaly features from implementation item data through a fourth Transformer encoder, extract implementation temporal anomaly features from the implementation data through a fourth temporal decomposition subunit, and output implementation anomaly features including the implementation semantic anomaly features and the implementation temporal anomaly features. The cross-domain feature fusion unit is configured to calculate cross-weights of the project anomaly features, the procurement anomaly features, the contract anomaly features, and the implementation anomaly features through a gated attention mechanism, fuse the project anomaly features, the procurement anomaly features, the contract anomaly features, and the implementation anomaly features based on the cross-weights, and output a fusion anomaly feature vector.

[0092] The multiple dedicated detection units (project initiation data, procurement data, contract data, and implementation data detection units) process different types of project data respectively, and fuse semantic and temporal features (for example, the Transformer encoder extracts semantic anomaly features, and the temporal decomposition subunit extracts temporal anomaly features), which covers the project full life cycle data, avoids the limitation of a single data source, and improves the integrity of anomaly detection. The multi-modal feature extraction (semantic + temporal) captures the structure and dynamic changes of the project data, improves the detection accuracy, and especially can identify potential hidden anomalies when processing complex project scenarios.

[0093] The cross-domain feature fusion unit adopts a gated attention mechanism to calculate and weight the fusion of abnormal features from different data types (such as project initiation abnormal features, procurement abnormal features, etc.), can adaptively adjust the cross-weight, highlight the correlation between important features, solve the data island problem, enhance the sensitivity of abnormal detection (for example, identify complex abnormalities caused by cross-department data dependence), and reduce the false positive rate, effectively improving the detection efficiency compared with traditional simple splicing or average fusion methods.

[0094] The root cause positioning module is constructed based on a data graph construction unit and a causal reasoning unit; the data graph construction unit is used to construct a project data graph by a dynamic graph neural network, taking project entities in project data, procurement data, contract data and implementation data as nodes and data dependency relationships as edges; the causal reasoning unit is used to infer the fusion abnormal feature vector by a causal graph convolution network, and randomly walk on the project data graph to locate an abnormal path;

[0095] The root cause positioning module provides end-to-end root cause tracking (abnormal path output) by the data graph construction unit (constructing project entity nodes and data dependency edges using a dynamic graph neural network) and the causal reasoning unit (inference and random walk positioning of abnormal paths by a causal graph convolution network), and enhances the explainability of the output (users can understand the source of the abnormality); combined with causal reasoning, it can locate deep root causes (such as contract data problems leading to implementation abnormalities), rather than just surface abnormalities, which is beneficial to subsequent decision-making and problem eradication, and improves the trustworthiness of the system.

[0096] The output module is used to output a quality check report carrying abnormal data and abnormal root causes according to the fusion abnormal feature vector and the abnormal path;

[0097] Step S2, a large amount of historical project data is obtained, and after pre-processing each of the historical project data, a data set is constructed, and a sample expansion operation is performed on the data set;

[0098] Step S3, the data quality check model is trained by the data set and a loss function, and the trained data quality check model is deployed as a microservice;

[0099] Step S4, real-time project data is obtained from a distributed project heterogeneous system, the real-time project data is input into the deployed data quality check model through the microservice, and a quality check report carrying abnormal data and abnormal root causes is obtained;

[0100] Step S5, based on the quality check report, the real-time project data is intercepted and warned;

[0101] Step S6, real-time record at least includes the real-time project data, quality inspection report, inspection result feedback and inspection time, and the inspection log is encrypted and stored based on the inspection log.

[0102] Through covering data collection, model training, real-time deployment (microservice) to application response, full-process automation, reducing manual intervention; through sample expansion to effectively improve the model generalization ability; microservice architecture facilitates deployment and expansion, supports high-concurrency real-time processing, ensures low-latency response (millisecond level), and improves operation efficiency.

[0103] Through the integration of multi-modal feature extraction (using Transformer encoder and time decomposition unit to capture semantic and time sequence anomaly features of project, procurement, contract and implementation data), gated attention mechanism driven cross-domain feature fusion (adaptive weighted fusion of multi-source anomaly features), and root cause positioning based on dynamic graph neural network and causal reasoning (constructing project data graph and locating abnormal path through random walk), high-precision anomaly detection and interpretable root cause analysis of project full life cycle data are realized; The end-to-end automated process (from historical data training, microservice real-time deployment to distributed system access) combined with multi-task loss function optimization (weighted balance of focal loss and path similarity comparison loss) significantly improves detection efficiency and model robustness; At the same time, through real-time interception and early warning, incremental data iterative optimization and encrypted log storage, a closed-loop quality control system is formed, which has millisecond-level response capability, continuous adaptive evolution and industrial-level security compliance, and fully solves the core pain points of incomplete coverage, high false alarm rate, root cause difficult to trace, response delay and other problems in project data quality inspection.

[0104] In the step S1, the formula of the loss function is:

[0105] L=λ1*L det +λ2*L root ;

[0106] Wherein, L represents the loss value of the loss function; L det represents the anomaly detection loss, which adopts the focal loss of the division type; L root represents the root cause positioning loss, which adopts the path similarity comparison loss; λ1, λ2 both represent the weight coefficient.

[0107] Through self-defined loss function (L=λ1*L det +λ2*L root) integrates anomaly detection and root cause localization tasks, focuses on loss processing data imbalance (such as minority class anomaly samples), improves detection robustness; path similarity loss optimizes root cause localization accuracy; weight coefficient can be adjusted to balance task priority, speed up model convergence, and improve overall model performance.

[0108] The step S2 is specifically:

[0109] A large amount of historical project data including historical project data, historical procurement data, historical contract data and historical implementation data is obtained, each of the historical project data is preprocessed including at least data integration, missing value processing, duplicate value processing, data standardization and data normalization, each of the preprocessed historical project data is labeled for abnormal data and abnormal root cause, and a data set is constructed based on each of the labeled historical project data.

[0110] The sample expansion operation of noise injection, data translation and data scaling is performed on the numerical data in the data set, the sample expansion operation of time translation, time scaling and time reversal is performed on the time series data, and the sample expansion operation of synonym replacement, sentence reorganization and random insertion is performed on the text data.

[0111] Customized sample expansion operations are adopted for different data types (numerical, time series, text): noise injection, data translation and scaling of numerical data (such as procurement amount), which increases data diversity and prevents machine learning model overfitting; time translation, scaling and reversal of time series data (such as project implementation timeline), which simulates real-world time changes and enhances the robustness of time series models (such as predicting project delays); synonym replacement, sentence reorganization and random insertion of text data (such as contract terms), which enriches language expression and improves natural language processing model understanding; this type-sensitive expansion method has significant innovation, avoiding the limitations of "one-size-fits-all" expansion, effectively solving the small sample data problem and improving the generalization performance of the model in real scenarios.

[0112] Through systematic data preprocessing (including integration, missing value processing, standardization, etc.), the quality and consistency of historical project data are significantly improved, combined with precise labeling of abnormal data and root causes, a highly reliable data set is constructed; further, customized sample expansion strategies (such as noise injection, time translation, synonym replacement, etc.) are designed for numerical, time series and text data, effectively enhancing data diversity, solving the model overfitting problem in small sample scenarios, greatly improving the generalization ability and prediction accuracy of machine learning models, and the automated process reduces the cost of manual intervention, providing efficient and scalable data support for project risk assessment, decision optimization and other advanced applications.

[0113] The step S3 is specifically:

[0114] The data set is divided into a training set, a verification set and a test set in a ratio of 8:1:1 by K-fold cross-validation method, the data quality checking model is trained by the training set, and the data quality checking model is continuously optimized during the training process at least including learning rate, random dropout rate, batch size, number of hidden layers, activation function, and hyperparameters of the optimizer, until the loss value of the loss function is less than a preset loss threshold;

[0115] The trained data quality checking model is verified by the verification set to determine whether the checking accuracy is greater than a preset accuracy threshold, if not, the verification fails, and the training set is expanded to continue training; if yes, the verification passes, and:

[0116] The precision, recall, F1 score and confidence are calculated by the test set to test the data quality checking model that passes the verification, if the test fails, the training set is expanded to continue training; if the test passes, the training ends, and the data quality checking model is deployed as a microservice.

[0117] By adopting K-fold cross-validation combined with an 8:1:1 data division strategy, the generalization ability and robustness of the model are significantly improved, overfitting is effectively avoided, key hyperparameters such as learning rate and random dropout rate are dynamically optimized during the training process, and the training is terminated by the loss function threshold, which greatly improves the parameter tuning efficiency and resource utilization. At the same time, a double protection mechanism of verification and testing is designed, the accuracy threshold is used to filter the model in the verification stage, and multiple dimensional indexes such as precision, recall, F1 score and confidence are comprehensively evaluated in the testing stage to ensure the reliability of the model decision. The closed-loop feedback mechanism of dynamically expanding the training set can adaptively optimize the data deficiency problem, and finally realizes high scalability and low operation and maintenance cost through microservice deployment, forming an efficient, stable and iterative data quality checking model development paradigm.

[0118] The step S4 is specifically:

[0119] Real-time project data including real-time project data, real-time procurement data, real-time contract data and real-time implementation data are obtained from a distributed project heterogeneous system including a project subsystem, a procurement subsystem, a contract subsystem and an implementation subsystem by a stream processing framework; the stream processing framework can be Apache Flink or Apache SparkStreaming, and the working principle is to distribute data streams to multiple computing nodes for parallel processing, supporting real-time data processing and analysis;

[0120] The microservice inputs real-time project data into a deployed data quality checking model, which performs inference through hardware acceleration technology to obtain a quality checking report carrying abnormal data and abnormal root causes;

[0121] The flow processing framework integrates multi-source project data of distributed heterogeneous systems (including project initiation, procurement, contract, and implementation subsystems) in real time, and uses a hardware-accelerated microservice model for efficient data quality checking, significantly improving real-time data processing and cross-system compatibility, breaking down data silos and enabling global monitoring. Relying on hardware acceleration technology to optimize model inference efficiency, significantly reducing computational latency and resource consumption; synchronously outputting abnormal data and root cause analysis, enhancing the accuracy of problem diagnosis and decision support capabilities; ultimately, through an automated and scalable architecture design, reducing operational costs and preventing project risks, forming an efficient, accurate, and cost-effective integrated project data quality checking solution.

[0122] The step S5 specifically includes:

[0123] Real-time monitoring of the quality checking report, when the value of the abnormal data or abnormal root cause in the quality checking report is not empty, intercepting the project data associated with the abnormal data or abnormal root cause, generating an early warning notification based on the quality checking report, encrypting the early warning notification into an encrypted notification through a first encryption rule, and pushing the encrypted notification to a pre-associated management terminal in real time through a TLS protocol to perform an early warning operation;

[0124] The first encryption rule specifically includes:

[0125] Calculating the hash value of the early warning notification through the SM3 algorithm, converting the early warning notification into binary data according to UTF-8 encoding, encrypting the binary data through the SM4 algorithm to obtain first-level encrypted data, concatenating the first-level encrypted data with the hash value to obtain concatenated data, converting the concatenated data to a hexadecimal string, mapping the numbers 1, 2, 3, and 4 in the hexadecimal string to the letters B, C, D, and F respectively to obtain second-level encrypted data, and encrypting the second-level encrypted data into an encrypted notification through the ECC algorithm;

[0126] By monitoring the quality verification report in real time, the associated data is automatically intercepted and a warning is generated when an anomaly is triggered. A multi-level encryption mechanism (combining SM3 hash verification, SM4 symmetric encryption, character mapping optimization, and ECC asymmetric encryption) is used to ensure data security and integrity. The data is pushed to the management terminal in real time using the TLS protocol, significantly improving the efficiency of abnormal response and the level of automation. The combination of national encryption algorithms and transmission encryption ensures end-to-end security, effectively solving the problems of response lag, high dependence on manual intervention, and high risk of data leakage in traditional monitoring systems. This provides a high-reliability, low-latency closed-loop solution for data quality management in sensitive areas.

[0127] The step S6 is specifically:

[0128] The real-time record includes at least the real-time project data, the quality verification report, the verification result feedback, and the verification log of the verification time. Based on the verification log, an incremental data set is constructed. When the data volume of the incremental data set exceeds a preset quantity threshold, the data quality verification model is backed up. The data quality verification model is trained through the incremental data set. The performance of the trained data quality verification model is verified. If the verification is passed, the training is ended. If the verification is not passed, the backed-up data quality verification model is restored to iteratively optimize the data quality verification model. The verification log is encrypted into an encrypted log through a second encryption rule. The encrypted log is stored and distributedly backed up.

[0129] The second encryption rule is specifically:

[0130] The MAC value of the verification log is calculated through the HMAC algorithm. The verification log and the MAC value are encrypted through the SM9 algorithm to obtain one layer of encrypted data. The one layer of encrypted data is equally divided into a first data block and a second data block. Each character of the first data block is cyclically shifted to the left by 7 bits to obtain a first obfuscated data block. Each character of the second data block is cyclically shifted to the right by 4 bits to obtain a second obfuscated data block. The first obfuscated data block and the second obfuscated data block are encrypted into two layers of encrypted data through the RC6 algorithm. The two layers of encrypted data are split in a ratio of 1:5 and the order is exchanged to obtain the encrypted log.

[0131] The closed-loop iterative optimization of the data quality model is realized by constructing the incremental log containing the complete verification elements, the threshold trigger mechanism is innovatively used to ensure the stability and calculation efficiency of the model updating, and the four-layer nested encryption system is designed, that is, the national SM9 algorithm is used to verify the data integrity, the asymmetric character bit shift confusion is used to build the active defense layer, the RC6 algorithm is used to implement deep encryption, and the 1:5 ratio data sequencing is used to form the final ciphertext, so that the single-point encryption vulnerability problem existing in the sensitive log protection of the traditional scheme is solved on the basis of ensuring the continuous evolution of the data quality management, and the real-time project data quality monitoring capability and attack resistance are significantly improved.

[0132] By constructing a machine learning model combining multi-source data anomaly detection (project initiation, procurement, contract, implementation) and root cause positioning, combining the Transformer encoder to extract semantic features and time series decomposition to capture dynamic anomalies, using the gated attention mechanism to realize cross-domain feature fusion, and using the dynamic graph neural network to build a data graph for causal reasoning, the comprehensiveness of anomaly detection and the accuracy of root cause positioning are significantly improved; Its end-to-end automated process (from data preprocessing, sample expansion, model training to microservice deployment) supports real-time stream processing and hardware acceleration, efficiently outputs quality check reports carrying abnormal data and root causes, realizes active interception and multi-layer encryption warning (such as SM4+ECC), continuously optimizes the model through incremental learning, and uses national encryption algorithms (SM3 / SM9 / RC6) and confusion techniques to strengthen log security, with the core advantages of high robustness, strong security, scalability and low operation and maintenance cost.

[0133] The preferred embodiment of the project data quality check system based on machine learning includes the following modules:

[0134] The data quality check model creation module is used to create a data quality check model based on the data anomaly detection module, the root cause positioning module and the output module, and set the loss function of the data quality check model;

[0135] The data anomaly detection module is constructed based on a project initiation data detection unit, a procurement data detection unit, a contract data detection unit, an implementation data detection unit, and a cross-domain feature fusion unit; the project initiation data detection unit is configured to extract project semantic anomaly features from project initiation data through a first Transformer encoder, extract project timing anomaly features from the project initiation data through a first timing decomposition subunit, and output project anomaly features containing the project semantic anomaly features and the project timing anomaly features; the procurement data detection unit is configured to extract procurement semantic anomaly features from procurement data through a second Transformer encoder, extract procurement timing anomaly features from the procurement data through a second timing decomposition subunit, and output procurement anomaly features containing the procurement semantic anomaly features and the procurement timing anomaly features; the contract data detection unit is configured to extract contract semantic anomaly features from contract item data through a third Transformer encoder, extract contract timing anomaly features from the contract data through a third timing decomposition subunit, and output contract anomaly features containing the contract semantic anomaly features and the contract timing anomaly features; the implementation data detection unit is configured to extract implementation semantic anomaly features from implementation item data through a fourth Transformer encoder, extract implementation timing anomaly features from the implementation data through a fourth timing decomposition subunit, and output implementation anomaly features containing the implementation semantic anomaly features and the implementation timing anomaly features; and the cross-domain feature fusion unit is configured to calculate cross-weights of the project anomaly features, the procurement anomaly features, the contract anomaly features, and the implementation anomaly features respectively through a gated attention mechanism, fuse the project anomaly features, the procurement anomaly features, the contract anomaly features, and the implementation anomaly features based on the cross-weights, and output a fusion anomaly feature vector.

[0136] The multiple dedicated detection units (project initiation data, procurement data, contract data, and implementation data detection units) respectively process different types of project data, and fuse semantic and timing features (for example, a Transformer encoder extracts semantic anomaly features, and a timing decomposition subunit extracts timing anomaly features), which covers project lifecycle data, avoids single data source limitations, and improves the integrity of anomaly detection; multi-modal feature extraction (semantic + timing) captures the structure and dynamic changes of project data, improves detection accuracy, and can identify potential hidden anomalies, especially in complex project scenarios.

[0137] The cross-domain feature fusion unit adopts a gated attention mechanism to calculate and weight the fusion of abnormal features from different data types (such as project initiation abnormal features, procurement abnormal features, etc.), can adaptively adjust the cross-weight, highlight the correlation between important features, solve the data island problem, enhance the sensitivity of abnormal detection (for example, identify composite anomalies caused by cross-department data dependence), and reduce the false positive rate, effectively improving the detection efficiency compared with traditional simple splicing or average fusion methods.

[0138] The root cause positioning module is constructed based on a data graph construction unit and a causal reasoning unit; the data graph construction unit is used to construct a project data graph by a dynamic graph neural network, taking project entities in project data, procurement data, contract data and implementation data as nodes and data dependence relationships as edges; the causal reasoning unit is used to infer the fusion abnormal feature vector by a causal graph convolution network, and randomly walk on the project data graph to locate an abnormal path;

[0139] The root cause positioning module provides end-to-end root cause tracking (abnormal path output) through the data graph construction unit (constructing project entity nodes and data dependence edges using a dynamic graph neural network) and the causal reasoning unit (inference and random walk to locate an abnormal path through a causal graph convolution network), enhancing the explainability of the output (users can understand the source of the abnormality); combined with causal reasoning, it can locate deep root causes (such as contract data problems leading to implementation abnormalities), rather than just surface abnormalities, which is beneficial to subsequent decision-making and problem eradication, and improves the trustworthiness of the system.

[0140] The output module is used to output a quality check report carrying abnormal data and abnormal root causes according to the fusion abnormal feature vector and the abnormal path;

[0141] The data set construction module is used to obtain a large amount of historical project data, pre-process each of the historical project data to construct a data set, and perform a sample expansion operation on the data set;

[0142] The data quality check model training and deployment module is used to train a data quality check model through the data set and a loss function, and deploy the trained data quality check model as a microservice;

[0143] The project data quality check module is used to obtain real-time project data from a distributed project heterogeneous system, input the real-time project data into the deployed data quality check model through the microservice, and obtain a quality check report carrying abnormal data and abnormal root causes;

[0144] The project data interception and early warning module is used to intercept and warn real-time project data based on the quality check report;

[0145] The verification log management module is used for recording a verification log in real time, the verification log at least including the real-time project data, the quality verification report, the verification result feedback and the verification time, constructing an incremental data set based on the verification log, iteratively optimizing the data quality verification model, and encrypting and storing the verification log.

[0146] By covering data collection, model training, real-time deployment (microservice) to application response, full-process automation is reduced manual intervention; by sample expansion effectively improving the model generalization ability; microservice architecture facilitates deployment and expansion, supports high-concurrency real-time processing, ensures low-latency response (millisecond level), and improves operation efficiency.

[0147] By integrating multi-modal feature extraction (using Transformer encoder and time decomposition unit to capture semantic and time sequence anomaly features of project, procurement, contract and implementation data), gated attention mechanism driven cross-domain feature fusion (adaptive weighted fusion of multi-source anomaly features), and root cause positioning based on dynamic graph neural network and causal reasoning (constructing project data graph and locating abnormal path through random walk), high-precision anomaly detection and interpretable root cause analysis of project full life cycle data are realized; the end-to-end automated process (from historical data training, microservice real-time deployment to distributed system access) combined with multi-task loss function optimization (weighted balance of focal loss and path similarity comparison loss) significantly improves detection efficiency and model robustness; at the same time, through real-time interception and early warning, incremental data iterative optimization and encrypted log storage, a closed-loop quality control system is formed, which has millisecond-level response capability, continuous adaptive evolution and industrial-level safety compliance, and fully solves the core pain points of incomplete coverage, high false positive rate, root cause difficult to trace and response delay in project data quality verification.

[0148] In the data quality verification model creation module, the formula of the loss function is:

[0149] L=λ1*L det +λ2*L root ;

[0150] Wherein, L represents the loss value of the loss function; L det represents the anomaly detection loss, which adopts the focal loss of different types; L root represents the root cause positioning loss, which adopts the path similarity comparison loss; λ1 and λ2 both represent the weight coefficient.

[0151] By defining the loss function (L=λ1*L det +λ2*L root) integrate anomaly detection and root cause localization tasks, where the focus loss handles data imbalance (such as minority class anomaly samples), improving detection robustness; the path similarity loss optimizes root cause localization accuracy; the weight coefficient is adjustable to balance task priorities, speed up model convergence, and improve overall model performance.

[0152] The data set construction module is specifically used for:

[0153] A large amount of historical project data including historical project data, historical procurement data, historical contract data and historical implementation data is obtained, and each of the historical project data is preprocessed including at least data integration, missing value processing, duplicate value processing, data standardization and data normalization. Each of the historical project data after preprocessing is labeled with abnormal data and abnormal root cause, and a data set is constructed based on each of the labeled historical project data.

[0154] The sample expansion operation of noise injection, data translation and data scaling is performed on the numerical data in the data set, the sample expansion operation of time translation, time scaling and time reversal is performed on the time series data, and the sample expansion operation of synonym replacement, sentence reorganization and random insertion is performed on the text data.

[0155] Customized sample expansion operations are adopted for different data types (numerical, time series, text): noise injection, data translation and scaling of numerical data (such as procurement amount), which increases data diversity and prevents machine learning model overfitting; time translation, scaling and reversal of time series data (such as project implementation timeline), which simulates real-world time changes and enhances the robustness of time series models (such as predicting project delays); synonym replacement, sentence reorganization and random insertion of text data (such as contract terms), which enriches language expression and improves natural language processing model understanding ability; this type-sensitive expansion method has significant innovation, avoiding the limitations of "one-size-fits-all" expansion, effectively solving the small sample data problem and improving the generalization performance of the model in real scenarios.

[0156] Through systematic data preprocessing (including integration, missing value processing, standardization, etc.), the quality and consistency of historical project data are significantly improved, and combined with precise labeling of abnormal data and root causes, a highly reliable data set is constructed; further, customized sample expansion strategies (such as noise injection, time translation, synonym replacement, etc.) are designed for numerical, time series and text data, effectively enhancing data diversity, solving the problem of model overfitting in small sample scenarios, greatly improving the generalization ability and prediction accuracy of machine learning models, and the automated process reduces the cost of manual intervention, providing efficient and scalable data support for project risk assessment, decision optimization and other advanced applications.

[0157] The data quality checking model training and deployment module is specifically configured to:

[0158] The data set is divided into a training set, a validation set and a test set in a ratio of 8:1:1 through K-fold cross-validation, the data quality checking model is trained through the training set, and at least hyperparameters of the data quality checking model including a learning rate, a random dropout rate, a batch size, a number of hidden layers, an activation function and an optimizer are continuously optimized during the training process until a loss value of the loss function is less than a preset loss threshold;

[0159] The trained data quality checking model is verified through the validation set to determine whether a checking accuracy is greater than a preset accuracy threshold, if not, the verification fails, and the training set is expanded to continue training, if yes, the verification passes, and:

[0160] The precision, recall, F1 score and confidence are calculated through the test set to test the data quality checking model that passes the verification, if the test fails, the training set is expanded to continue training, if the test passes, the training ends, and the data quality checking model is deployed as a microservice.

[0161] By adopting K-fold cross-validation combined with an 8:1:1 data division strategy, the generalization ability and robustness of the model are significantly improved, overfitting is effectively avoided, key hyperparameters such as the learning rate and the random dropout rate are dynamically optimized during the training process, the training termination is controlled through the loss function threshold, the parameter tuning efficiency and resource utilization are greatly improved, a double protection mechanism of verification and testing is designed, the model is screened through the accuracy threshold in the verification stage, and multiple-dimensional indexes such as the precision, recall, F1 score and confidence are comprehensively evaluated in the testing stage to ensure the reliability of the model decision, the closed-loop feedback mechanism of dynamically expanding the training set can adaptively optimize the data deficiency problem, and finally the high scalability and low operation and maintenance cost are realized through the microservice deployment, forming an efficient, stable and iterative data quality checking model development paradigm.

[0162] The project data quality checking module is specifically configured to:

[0163] Real-time project data including real-time project data, real-time procurement data, real-time contract data and real-time implementation data is acquired from a distributed project heterogeneous system including a project subsystem, a procurement subsystem, a contract subsystem and an implementation subsystem through a stream processing framework, the stream processing framework can be Apache Flink or Apache SparkStreaming, and the working principle is to distribute data streams to multiple computing nodes for parallel processing to support real-time data processing and analysis;

[0164] The microservice inputs real-time project data into a deployed data quality checking model, which performs inference through hardware acceleration technology to obtain a quality checking report carrying abnormal data and abnormal root causes;

[0165] The multi-source project data of distributed heterogeneous systems (including project initiation, procurement, contract and implementation subsystems) is integrated in real time through a stream processing framework, and hardware-accelerated microservice models are used for efficient data quality checking, significantly improving the real-time performance and cross-system compatibility of data processing, breaking down data silos and achieving global monitoring; relying on hardware acceleration technology to optimize model inference efficiency, significantly reducing computational delay and resource consumption; synchronously outputting abnormal data and root cause analysis, enhancing the accuracy of problem diagnosis and decision support capability; ultimately, through an automated and scalable architecture design, reducing operation and maintenance costs and preventing project risks, forming an efficient, accurate and economical integrated project data quality checking solution.

[0166] The project data interception and early warning module is specifically configured to:

[0167] The quality checking report is monitored in real time, and when the value of the abnormal data or abnormal root cause in the quality checking report is not empty, the project data associated with the abnormal data or abnormal root cause is intercepted, a warning notification is generated based on the quality checking report, the warning notification is encrypted into an encrypted notification through a first encryption rule, and the encrypted notification is pushed to a pre-associated management terminal in real time through a TLS protocol to perform a warning operation;

[0168] The first encryption rule is specifically configured to:

[0169] The hash value of the warning notification is calculated through an SM3 algorithm, the warning notification is converted into binary data according to UTF-8 encoding, the binary data is encrypted through an SM4 algorithm to obtain first-level encrypted data, the first-level encrypted data is spliced with the hash value to obtain spliced data, the spliced data is converted into a hexadecimal string, and the numbers 1, 2, 3 and 4 in the hexadecimal string are respectively mapped to the letters B, C, D and F to obtain second-level encrypted data, and the second-level encrypted data is encrypted into an encrypted notification through an ECC algorithm;

[0170] By monitoring the quality check report in real time, the associated data is automatically intercepted and a warning is generated when an anomaly is triggered. A multi-level encryption mechanism (combining SM3 hash verification, SM4 symmetric encryption, character mapping optimization, and ECC asymmetric encryption) is used to ensure data security and integrity. The data is pushed to the management terminal in real time using the TLS protocol, significantly improving the efficiency of abnormal response and the level of automation. The combination of national encryption algorithms and transmission encryption ensures end-to-end security, effectively solving the problems of response lag, high dependence on manual intervention, and high risk of data leakage in traditional monitoring systems. This provides a high-reliability, low-latency closed-loop solution for data quality management in sensitive areas.

[0171] The verification log management module is specifically used for:

[0172] The verification log includes at least the real-time project data, quality check report, verification result feedback, and verification time. Based on the verification log, an incremental data set is constructed. When the data volume of the incremental data set exceeds a preset quantity threshold, the data quality check model is backed up. The data quality check model is trained using the incremental data set. The trained data quality check model is performance-verified. If the verification is passed, the training is ended. If the verification is not passed, the backed-up data quality check model is restored for iterative optimization. The verification log is encrypted into an encrypted log using a second encryption rule. The encrypted log is stored and distributed for backup.

[0173] The second encryption rule is specifically:

[0174] The MAC value of the verification log is calculated using the HMAC algorithm. The verification log and MAC value are encrypted using the SM9 algorithm to obtain one layer of encrypted data. The one layer of encrypted data is equally divided into a first data block and a second data block. Each character in the first data block is cyclically shifted 7 bits to the left to obtain a first obfuscated data block. Each character in the second data block is cyclically shifted 4 bits to the right to obtain a second obfuscated data block. The first obfuscated data block and the second obfuscated data block are encrypted into two layers of encrypted data using the RC6 algorithm. The two layers of encrypted data are split in a 1:5 ratio and the order is exchanged to obtain the encrypted log.

[0175] The closed-loop iterative optimization of the data quality model is realized by constructing the incremental log containing complete verification elements, the threshold trigger mechanism is innovatively used to ensure the stability and computational efficiency of model updating, and a four-layer nested encryption system is designed, which integrates the SM9 algorithm to verify data integrity, the asymmetric character bit shift confusion to build an active defense layer, the RC6 algorithm to implement deep encryption, and a 1:5 ratio data sequencing to form the final ciphertext. On the basis of ensuring the continuous evolution of data quality management, the single-point encryption vulnerability problem existing in the traditional scheme in sensitive log protection is solved, and the real-time project data quality monitoring capability and attack resistance are significantly improved.

[0176] By constructing a machine learning model that integrates multi-source data anomaly detection (project initiation, procurement, contract, implementation) and root cause positioning, combining the Transformer encoder to extract semantic features and time series decomposition to capture dynamic anomalies, using the gated attention mechanism to realize cross-domain feature fusion, and constructing a data graph through a dynamic graph neural network for causal reasoning, the comprehensiveness of anomaly detection and the accuracy of root cause positioning are significantly improved; its end-to-end automated process (from data preprocessing, sample expansion, model training to microservice deployment) supports real-time stream processing and hardware acceleration, efficiently outputs quality check reports carrying abnormal data and root causes, realizes active interception and multi-layer encryption warning (such as SM4+ECC), continuously optimizes the model through incremental learning, and uses the SM algorithm (SM3 / SM9 / RC6) and confusion technology to strengthen log security, with the core advantages of high robustness, strong security, scalability and low operation and maintenance cost.

[0177] Although the specific embodiments of the present application are described above, those skilled in the art should understand that the specific examples described are only illustrative, and are not intended to limit the scope of the present application, and equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present application should be covered within the scope of the claims of the present application.

Claims

1. A machine learning based project data quality check method, characterized in that: The method comprises the following steps: Step S1, creating a data quality checking model based on a data anomaly detection module, a root cause positioning module and an output module, and setting a loss function of the data quality checking model; Step S2, obtaining a large amount of historical project data, preprocessing each of the historical project data to construct a data set, and performing a sample expansion operation on the data set; Step S3, training the data quality checking model through the data set and the loss function, and deploying the trained data quality checking model as a microservice; Step S4, obtaining real-time project data including real-time project data, real-time procurement data, real-time contract data and real-time implementation data from a distributed project heterogeneous system including a project establishment subsystem, a procurement subsystem, a contract subsystem and an implementation subsystem through a stream processing framework; inputting the real-time project data into the deployed data quality checking model through the microservice, and obtaining a quality checking report carrying abnormal data and abnormal root causes through the data quality checking model by using a hardware acceleration technology; Step S5, real-time monitoring the quality checking report, when the value of the abnormal data or the abnormal root cause in the quality checking report is not empty, intercepting the project data associated with the abnormal data or the abnormal root cause, generating an early warning notification based on the quality checking report, encrypting the early warning notification into an encrypted notification through a first encryption rule, and pushing the encrypted notification to a pre-associated management terminal in real time through a TLS protocol to perform an early warning operation; The first encryption rule is specifically: calculating the hash value of the early warning notification through an SM3 algorithm, converting the early warning notification into binary data according to UTF-8 encoding, encrypting the binary data through an SM4 algorithm to obtain first-level encrypted data, splicing the first-level encrypted data and the hash value to obtain spliced data, converting the spliced data into a hexadecimal string, mapping the numbers 1, 2, 3 and 4 in the hexadecimal string to the letters B, C, D and F respectively to obtain second-level encrypted data, and encrypting the second-level encrypted data into an encrypted notification through an ECC algorithm; Step S6, real-time recording of a checking log including at least the real-time project data, the quality checking report, the checking result feedback and the checking time, constructing an incremental data set based on the checking log, when the data amount of the incremental data set exceeds a preset quantity threshold, backing up the data quality checking model, training the data quality checking model through the incremental data set, verifying the performance of the trained data quality checking model, verifying through, and ending the training, verifying not through, and restoring the backed up data quality checking model to iteratively optimize the data quality checking model, and encrypting the checking log into an encrypted log through a second encryption rule, storing and distributing the encrypted log; The second encryption rule is specifically: The MAC value of the check log is calculated by an HMAC algorithm, the check log and the MAC value are encrypted by an SM9 algorithm to obtain one layer of encrypted data, the one layer of encrypted data is equally divided into a first data block and a second data block, each character of the first data block is cyclically shifted to the left by 7 bits to obtain a first obfuscated data block, each character of the second data block is cyclically shifted to the right by 4 bits to obtain a second obfuscated data block, the first obfuscated data block and the second obfuscated data block are encrypted into two layers of encrypted data by an RC6 algorithm, and the two layers of encrypted data are split in a 1:5 ratio and the front and rear orders are exchanged to obtain an encrypted log.

2. The machine learning based project data quality check method as claimed in claim 1, wherein: In the step S1, the data anomaly detection module is constructed based on a project initiation data detection unit, a procurement data detection unit, a contract data detection unit, an implementation data detection unit, and a cross-domain feature fusion unit; the project initiation data detection unit is configured to extract project semantic anomaly features from project initiation data through a first Transformer encoder, extract project timing anomaly features from the project initiation data through a first timing decomposition subunit, and output project anomaly features including the project semantic anomaly features and the project timing anomaly features; The procurement data detection unit is configured to extract procurement semantic anomaly features from procurement data through a second Transformer encoder, extract procurement timing anomaly features from the procurement data through a second timing decomposition subunit, and output procurement anomaly features including the procurement semantic anomaly features and the procurement timing anomaly features; The contract data detection unit is configured to extract contract semantic anomaly features from contract data through a third Transformer encoder, extract contract timing anomaly features from the contract data through a third timing decomposition subunit, and output contract anomaly features including the contract semantic anomaly features and the contract timing anomaly features; The implementation data detection unit is configured to extract implementation semantic anomaly features from implementation data through a fourth Transformer encoder, extract implementation timing anomaly features from the implementation data through a fourth timing decomposition subunit, and output implementation anomaly features including the implementation semantic anomaly features and the implementation timing anomaly features; and the cross-domain feature fusion unit is configured to calculate cross-weights of the project anomaly features, the procurement anomaly features, the contract anomaly features, and the implementation anomaly features respectively through a gated attention mechanism, fuse the project anomaly features, the procurement anomaly features, the contract anomaly features, and the implementation anomaly features based on the cross-weights, and output a fusion anomaly feature vector; The root cause positioning module is constructed based on a data graph construction unit and a causal reasoning unit; the data graph construction unit is configured to construct a project data graph by taking project entities in project initiation data, procurement data, contract data, and implementation data as nodes and data dependency relationships as edges through a dynamic graph neural network; and the causal reasoning unit is configured to reason the fusion anomaly feature vector through a causal graph convolution network and randomly walk on the project data graph to locate an abnormal path. The output module is configured to output a quality check report carrying abnormal data and an abnormal root cause according to the fused abnormal feature vector and the abnormal path. The formula of the loss function is: L = λ1*L det + λ2*L root ; Wherein, L represents the loss value of the loss function; L det represents the anomaly detection loss, adopts the focal loss of the split type; L root represents the root cause positioning loss, adopts the path similarity comparison loss; λ1, λ2 both represent the weight coefficient.

3. The machine learning based project data quality check method as claimed in claim 1, wherein: The step S2 specifically includes: A large amount of historical project data including historical project initiation data, historical procurement data, historical contract data and historical implementation data is acquired, each of the historical project data is preprocessed including at least data integration, missing value processing, duplicate value processing, data standardization and data normalization, each of the preprocessed historical project data is labeled for abnormal data and an abnormal root cause, and a data set is constructed based on each of the labeled historical project data. The numerical data in the data set is subjected to sample expansion operations of noise injection, data translation and data scaling, the time series data is subjected to sample expansion operations of time translation, time scaling and time reversal, and the text data is subjected to sample expansion operations of synonym replacement, sentence reorganization and random insertion.

4. The machine learning based project data quality check method as claimed in claim 1, wherein: The step S3 specifically includes: The data set is divided into a training set, a validation set and a test set in a ratio of 8:1:1 through K-fold cross-validation, the data quality check model is trained through the training set, and the hyperparameters of the data quality check model including learning rate, random dropout rate, batch size, number of hidden layers, activation function and optimizer are continuously optimized during the training process until the loss value of the loss function is less than a preset loss threshold; The trained data quality check model is verified through the validation set to determine whether the check accuracy is greater than a preset accuracy threshold, if not, the verification fails, and the training set is expanded for continuous training; if yes, the verification passes, and the precision, recall, F1 score and confidence are calculated through the test set to test the data quality check model that passes the verification, and if the test fails, the training set is expanded for continuous training; If the test passes, the training is ended, and the data quality check model is deployed as a microservice.

5. A machine learning based project data quality verification system, characterized in that: The following modules are included: A data quality check model creation module is configured to create a data quality check model based on a data anomaly detection module, a root cause positioning module and an output module, and set a loss function of the data quality check model; A data set construction module is configured to acquire a large amount of historical project data, construct a data set after preprocessing each of the historical project data, and perform sample expansion operations on the data set; A data quality check model training and deployment module is configured to train a data quality check model through the data set and the loss function, and deploy the trained data quality check model as a microservice; A project data quality check module is configured to acquire real-time project data including real-time project initiation data, real-time procurement data, real-time contract data and real-time implementation data from a project heterogeneous system including a project initiation subsystem, a procurement subsystem, a contract subsystem and an implementation subsystem through a stream processing framework. The microservice inputs real-time project data into a deployed data quality checking model, which performs inference through a hardware acceleration technology to obtain a quality checking report carrying abnormal data and abnormal root causes; A project data interception early warning module is configured to monitor the quality checking report in real time, intercept project data associated with abnormal data or abnormal root causes in the quality checking report when the abnormal data or abnormal root causes have non-empty values, generate an early warning notification based on the quality checking report, encrypt the early warning notification into an encrypted notification through a first encryption rule, and push the encrypted notification to a pre-associated management terminal in real time through a TLS protocol to perform an early warning operation. The first encryption rule is specifically as follows: The hash value of the early warning notification is calculated through an SM3 algorithm, the early warning notification is converted into binary data according to UTF-8 encoding, the binary data is encrypted through an SM4 algorithm to obtain first-level encrypted data, the first-level encrypted data is spliced with the hash value to obtain spliced data, the spliced data is converted into a hexadecimal string, and the numbers 1, 2, 3 and 4 in the hexadecimal string are respectively mapped to the letters B, C, D and F to obtain second-level encrypted data, and the second-level encrypted data is encrypted into an encrypted notification through an ECC algorithm; A checking log management module is configured to record a checking log in real time, the checking log at least including the real-time project data, the quality checking report, checking result feedback and checking time, construct an incremental data set based on the checking log, back up the data quality checking model when the data volume of the incremental data set exceeds a preset quantity threshold, train the data quality checking model through the incremental data set, verify the performance of the trained data quality checking model, end the training when the verification is passed, restore the backed-up data quality checking model when the verification is not passed, and iteratively optimize the data quality checking model, and encrypt the checking log into an encrypted log through a second encryption rule, store and distribute the encrypted log. The second encryption rule is specifically as follows: The MAC value of the checking log is calculated through an HMAC algorithm, the checking log and the MAC value are encrypted through an SM9 algorithm to obtain one-layer encrypted data, the one-layer encrypted data is equally divided into a first data block and a second data block, each character of the first data block is cyclically shifted to the left by 7 bits to obtain a first obfuscated data block, each character of the second data block is cyclically shifted to the right by 4 bits to obtain a second obfuscated data block, the first obfuscated data block and the second obfuscated data block are encrypted through an RC6 algorithm to obtain two-layer encrypted data, and the two-layer encrypted data is split at a ratio of 1:5 and the front and rear orders are exchanged to obtain an encrypted log.

6. The machine learning based project data quality verification system as claimed in claim 5, wherein: In the data quality checking model creation module, the data anomaly detection module is constructed based on a project data detection unit, a procurement data detection unit, a contract data detection unit, an implementation data detection unit, and a cross-domain feature fusion unit; the project data detection unit is configured to extract project semantic anomaly features from project data through a first Transformer encoder, extract project time sequence anomaly features from the project data through a first time sequence decomposition subunit, and output project anomaly features containing the project semantic anomaly features and the project time sequence anomaly features; the procurement data detection unit is configured to extract procurement semantic anomaly features from procurement data through a second Transformer encoder, extract procurement time sequence anomaly features from the procurement data through a second time sequence decomposition subunit, and output procurement anomaly features containing the procurement semantic anomaly features and the procurement time sequence anomaly features; the contract data detection unit is configured to extract contract semantic anomaly features from contract data through a third Transformer encoder, extract contract time sequence anomaly features from the contract data through a third time sequence decomposition subunit, and output contract anomaly features containing the contract semantic anomaly features and the contract time sequence anomaly features; the implementation data detection unit is configured to extract implementation semantic anomaly features from implementation data through a fourth Transformer encoder, extract implementation time sequence anomaly features from the implementation data through a fourth time sequence decomposition subunit, and output implementation anomaly features containing the implementation semantic anomaly features and the implementation time sequence anomaly features; the cross-domain feature fusion unit is configured to calculate cross weights of the project anomaly features, the procurement anomaly features, the contract anomaly features, and the implementation anomaly features respectively through a gated attention mechanism, fuse the project anomaly features, the procurement anomaly features, the contract anomaly features, and the implementation anomaly features based on the cross weights, and output a fusion anomaly feature vector; The root cause positioning module is constructed based on a data graph construction unit and a causal reasoning unit; the data graph construction unit is configured to construct a project data graph by taking project entities in project data, procurement data, contract data, and implementation data as nodes and data dependency relationships as edges through a dynamic graph neural network; and the causal reasoning unit is configured to infer the fusion anomaly feature vector through a causal graph convolution network and randomly walk on the project data graph to locate an abnormal path; The output module is configured to output a quality checking report carrying abnormal data and abnormal root causes according to the fusion anomaly feature vector and the abnormal path; The formula of the loss function is: L = λ1*L det + λ2*L root ; Wherein, L represents the loss value of the loss function; L det represents the anomaly detection loss, adopts the focal loss of the split type; L root represents the root cause positioning loss, adopts the path similarity comparison loss; λ1, λ2 both represent the weight coefficient.

7. The machine learning based project data quality verification system as claimed in claim 5, wherein: The data set construction module is specifically configured to: obtain a large amount of historical project data including historical project data, historical procurement data, historical contract data, and historical implementation data, pre-process each of the historical project data through at least data integration, missing value processing, duplicate value processing, data standardization, and data normalization, label each of the pre-processed historical project data for abnormal data and abnormal root causes, and construct a data set based on each of the labeled historical project data; The sample expansion operations of noise injection, data translation and data scaling are performed on the numerical data in the data set, the sample expansion operations of time translation, time scaling and time reversal are performed on the time series data, and the sample expansion operations of synonym replacement, sentence reorganization and random insertion are performed on the text data.

8. The machine learning based project data quality verification system as claimed in claim 5, wherein: The data quality checking model training and deployment module is specifically configured to: divide the data set into a training set, a validation set and a test set in a ratio of 8:1:1 through K-fold cross-validation, train the data quality checking model through the training set, and continuously optimize the hyperparameters of the data quality checking model including learning rate, random dropout rate, batch size, number of hidden layers, activation function and optimizer in the training process until the loss value of the loss function is less than a preset loss threshold; verify the trained data quality checking model through the validation set, determine whether the checking accuracy is greater than a preset accuracy threshold, if not, the verification fails, the training set is expanded and the training is continued, if yes, the verification passes, and the precision, recall, F1 score and confidence are calculated through the test set to test the data quality checking model that passes the verification, if the test fails, the training set is expanded and the training is continued; if the test passes, the training ends and the data quality checking model is deployed as a microservice.

Images