Machine learning based survey data quality supervision method and system

Abstract

The application relates to the technical field of data supervision, in particular to a surveying and mapping data quality supervision method and system based on machine learning, which comprises the following steps: real-time receiving of surveying and mapping data flow and quality attribute information representing a business process, construction of a process topology model according to the quality attribute information; quantitative compliance deviation by using an analysis engine, identification of quality key control points; subsequent focusing on the quality key control points, combination of data blood relationship tracking, automatic determination of dynamic quality attributes of the data flow through time sequence analysis, establishment of a risk prediction model, generation of a risk index, final triggering of a closed-loop business compliance decision-making process, input of the risk index as a strategy variable into a decision-making model, dynamic adjustment of the range and degree of correction, generation of n candidate versions, selection of an optimal candidate version by calculating the difference between each version and original data, and automatic replacement of the optimal candidate version. The application realizes surveying and mapping data quality supervision through machine learning.

Description

Technical Field

[0001] This invention relates to the field of data supervision technology, specifically to a method and system for supervising the quality of surveying and mapping data based on machine learning. Background Technology

[0002] Surveying and mapping data encompasses a variety of information, including geographic information, spatial location, and topography, and is widely used in fields such as urban planning, land management, transportation construction, and resource exploration. However, with the rapid increase in the amount of surveying and mapping data, data quality issues have become increasingly prominent. Fluctuations in the quality of surveying and mapping data, including issues such as missing data, large errors, redundancy, and poor consistency, seriously affect the effectiveness and reliability of surveying and mapping data in practical applications.

[0003] In the modern surveying and mapping big data ecosystem, how to transform the traditional data quality inspection process, which relies on manual labor, is passive and lagging, into an automated, proactive, and intelligent supervision system that spans the entire data lifecycle? This system needs to go beyond simple technical error correction and become a business intelligence platform that serves organizational management and supervision. It should be able to perform quantitative performance management of the data production process, support resource optimization and allocation, predict potential risks, and adaptively assess whether the data is suitable for specific purposes based on different business application scenarios, thereby elevating data quality assurance to a core strategic asset of the organization.

[0004] To address this, a method and system for monitoring the quality of surveying and mapping data based on machine learning is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for monitoring the quality of surveying and mapping data based on machine learning, thereby enabling the monitoring of surveying and mapping data quality through machine learning.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Machine learning-based methods for monitoring the quality of surveying and mapping data include:

[0008] The system receives surveying and mapping data streams in real time, synchronously acquires quality attribute information representing the execution of business processes, constructs a process topology model based on the quality attribute information, and uses an analysis engine to quantify the frequency of compliance deviations and identify key quality control points.

[0009] Based on the quality critical control points and data lineage tracing of the lifecycle of surveying and mapping data in the business process, the dynamic quality attributes of the surveying and mapping data stream are automatically determined using time series analysis methods. The dynamic quality attributes are defined by business rules and a quality risk prediction model is established. A quality risk index is generated for the surveying and mapping data stream based on the dynamic quality attributes.

[0010] Based on the quality risk index, a business compliance decision-making process is triggered. The quality risk index is used as a business strategy variable to input into the quality decision-making model to adjust the decision-making scope and degree of the surveying and mapping data, generating n decision candidate versions. The difference between each decision candidate version and the original data is calculated, and the optimal candidate version is selected for automatic replacement.

[0011] The process topology model is constructed based on the quality attribute information:

[0012] The quality attribute information representing the execution of the business process includes the executor, the equipment used, the operating environment parameters, and the processing software version.

[0013] Each processing step of the quality attribute information is represented as a node in the graph, and the flow relationship of the surveying data stream in the execution of the business process is represented as a directed edge in the graph. Based on historical logs, risk weights are assigned to specific combinations of node attributes through machine learning, and a process topology graph integrating multi-dimensional attributes and associated risks is constructed.

[0014] A process topology model is constructed based on a process topology graph training analysis engine. The analysis engine consists of a graph neural inference network, which is used to learn the nonlinear propagation and accumulation pattern of the risk weights in the process topology graph structure, quantify the frequency of compliance deviations, and identify key quality control points.

[0015] The process of identifying critical quality control points is as follows:

[0016] Using the aforementioned process topology model, a sensitivity analysis oriented towards business outcomes is performed on each node in the process topology diagram.

[0017] The sensitivity analysis is the degree of impact on the business value of the final deliverable of the entire process when the business output of the node experiences quality fluctuations that conform to historical patterns during model simulation, and the degree of impact is quantified into a business importance indicator.

[0018] The frequency of historical compliance deviations associated with each node is statistically analyzed based on historical audit logs, serving as a management indicator characterizing the stability of that node.

[0019] The business importance index and the stability index are weighted and combined to calculate a comprehensive management priority score for each node, and the node with the highest comprehensive management priority score is determined as the quality critical control point.

[0020] The specific process of automatically determining the dynamic quality attributes of the mapping data stream using time series analysis is as follows:

[0021] For the identified critical quality control points, dynamic quality attributes are extracted from the surveying and mapping data stream in a targeted manner according to the business rules using time series analysis methods.

[0022] The time series analysis method adopts a multi-scale adaptive time series modeling framework, which combines graph convolutional networks with spatiotemporal co-modeling to capture the long-term and short-term dependencies of surveying and mapping data streams.

[0023] The business data includes the surveying and mapping data generation time, transmission time, processing time, equipment usage, and operators;

[0024] The dynamic quality attributes include data integrity, data accuracy, data consistency, data transmission latency, data processing latency, and surveying equipment stability.

[0025] The specific process of establishing a quality risk prediction model and generating a quality risk index for the mapping data stream based on the dynamic quality attributes is as follows:

[0026] The quality risk prediction model uses a gradient boosting tree to generate a quality risk index for the mapping data stream based on the dynamic quality attributes. The gradient boosting tree constructs a series of decision trees by progressively weighting weak learners, and adaptively adjusts the weights of the leaf nodes of each decision tree to reflect the quality risk index.

[0027] Cross-validation is used for training and validation, and reinforcement learning is used to continuously provide feedback on the deviation between the model's prediction results and the actual quality fluctuations, thereby optimizing the prediction results.

[0028] The specific process of using cross-validation for training and validation is as follows:

[0029] A hierarchical blocking cross-validation based on business dimensions is adopted, which groups historical logs and corresponding real quality labels according to the quality attribute information that characterizes the execution of business processes. The quality attribute information includes the parameters of the executor, the execution equipment, and the working environment, and divides them into data blocks. In each iteration of cross-validation, the complete data block is used as the validation set, and the remaining blocks are used as the training set.

[0030] The specific process for triggering business compliance decisions based on the quality risk index is as follows:

[0031] Based on the quality risk index, a business compliance decision-making process is triggered. The quality risk index is used as a business strategy variable and input into the quality decision-making model to adjust the decision range and degree of the mapping data, generating n decision candidate versions. The quality decision-making model is a generative adversarial network, whose generator maps the quality risk index into a modulation vector and injects it into the network layer to control the scope and degree of repair. By sampling the input random noise vector n times, n decision candidate versions are generated.

[0032] The calculation of the difference between each decision candidate version and the original data, and the selection of the optimal candidate version for automatic replacement, specifically involves:

[0033] This is accomplished collaboratively by the business value impact assessment module and the optimal decision selection module.

[0034] The business value impact assessment module calculates the weighted difference between each decision candidate version and the original data, where the weight is determined by the business importance index calculated by the process topology model for each data point, and a higher difference score will be generated when it is modified.

[0035] The optimal decision selection module executes a multi-objective optimization algorithm to calculate the comprehensive business cost for each candidate version and selects the version with the lowest cost as the optimal version. The comprehensive business cost includes: the residual quality risk index of the candidate version after correction, the weighted difference calculated by the business value impact assessment module as the data fidelity cost, and the business execution cost corresponding to the data processing operations required to generate the version.

[0036] A machine learning-based surveying and mapping data quality supervision system includes:

[0037] Data processing module: Receives surveying and mapping data streams in real time and synchronously acquires quality attribute information representing the execution of business processes;

[0038] Key identification module: Based on the quality attribute information, a process topology model is constructed. The process topology model uses an analysis engine to quantify the frequency of compliance deviations and identify key quality control points.

[0039] Quality Risk Module: Based on key quality control points and data lineage tracing of the lifecycle of surveying and mapping data in the business process, the module automatically determines the dynamic quality attributes of the surveying and mapping data stream using time series analysis methods. The dynamic quality attributes are defined by business rules, and a quality risk prediction model is established. Based on the dynamic quality attributes, a quality risk index is generated for the surveying and mapping data stream.

[0040] Quality Correction Module: Based on the quality risk index, the business compliance decision-making process is triggered. The quality risk index is used as a business strategy variable to input into the quality decision-making model to adjust the decision range and degree of the surveying and mapping data, generate n decision candidate versions, calculate the difference between each decision candidate version and the original data, and select the optimal candidate version for automatic replacement.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] 1. This invention constructs a process topology model that integrates multi-dimensional quality attributes such as personnel, equipment, and environment, and uses graph neural network inference for analysis. By combining business importance indicators and historical stability indicators, it achieves accurate and dynamic identification of key quality control points. This method changes the traditional model of average effort in quality supervision and reliance on human experience judgment. It can proactively and intelligently focus regulatory resources and attention on the business links that have the greatest impact on the value of the final result and the highest risk, thereby significantly improving regulatory efficiency and achieving precise prevention and control of major quality risks from the source.

[0043] 2. This invention employs a multi-scale adaptive temporal modeling framework to perform in-depth analysis of the data flow of key control points, and combines gradient boosting trees and reinforcement learning to construct a self-optimizing quality risk prediction model. This method breaks through the traditional quality evaluation method based on static thresholds, and can capture the dynamic evolution trend and complex patterns of data quality. In particular, the introduction of the reinforcement learning feedback mechanism enables the model to learn from the actual effects of historical predictions and continuously optimize its prediction strategy, thereby significantly improving the accuracy of the risk index and its adaptability to changing operating environments, and effectively reducing the false negative and false positive rates of quality problems.

[0044] 3. This invention introduces a generative adversarial network to generate multiple candidate repair versions and establishes a comprehensive business cost function that includes risk reduction, data fidelity cost, and business execution cost to select the optimal version, achieving automated and refined closed-loop processing. This method avoids the one-size-fits-all approach of traditional automatic repair, transforming the data repair problem into a multi-objective optimization problem oriented towards business value. It ensures that each automatic replacement decision is the optimal choice made after comprehensively weighing the repair effect, the impact on downstream businesses, and its own costs, ultimately forming a fully automated quality supervision system encompassing intelligent perception, accurate prediction, and optimized decision-making. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the machine learning-based surveying data quality supervision method of the present invention.

[0046] Figure 2 This is a schematic diagram of the data flow in the machine learning-based surveying data quality supervision method of the present invention;

[0047] Figure 3 This is a schematic diagram of the structure of the machine learning-based surveying data quality supervision system of the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Example 1:

[0050] Please see Figure 1 This invention provides a machine learning-based method for monitoring the quality of surveying and mapping data, the technical solution of which is as follows:

[0051] Machine learning-based methods for monitoring the quality of surveying and mapping data include:

[0052] The system receives surveying and mapping data streams in real time, synchronously acquires quality attribute information representing the execution of business processes, constructs a process topology model based on the quality attribute information, and uses an analysis engine to quantify the frequency of compliance deviations and identify key quality control points.

[0053] Based on the quality critical control points and data lineage tracing of the lifecycle of surveying and mapping data in the business process, the dynamic quality attributes of the surveying and mapping data stream are automatically determined using time series analysis methods. The dynamic quality attributes are defined by business rules and a quality risk prediction model is established. A quality risk index is generated for the surveying and mapping data stream based on the dynamic quality attributes.

[0054] Based on the quality risk index, a business compliance decision-making process is triggered. The quality risk index is used as a business strategy variable to input into the quality decision-making model to adjust the decision-making scope and degree of the surveying and mapping data, generating n decision candidate versions. The difference between each decision candidate version and the original data is calculated, and the optimal candidate version is selected for automatic replacement.

[0055] The process topology model is constructed based on the quality attribute information:

[0056] The quality attribute information representing the execution of the business process includes the executor, the equipment used, the operating environment parameters, and the processing software version.

[0057] Each processing step of the quality attribute information is represented as a node in the graph, and the flow relationship of the surveying data stream in the execution of the business process is represented as a directed edge in the graph. Based on historical logs, risk weights are assigned to specific combinations of node attributes through machine learning, and a process topology graph integrating multi-dimensional attributes and associated risks is constructed.

[0058] See Figure 2A process topology model is constructed based on a process topology graph training analysis engine. The analysis engine consists of a graph neural inference network, which is used to learn the nonlinear propagation and accumulation pattern of the risk weight in the process topology graph structure, quantify the frequency of compliance deviations, and identify key quality control points.

[0059] While receiving the surveying and mapping data stream in real time, the system simultaneously acquires and records quality attribute information related to each processing stage of the data stream. This quality attribute information is multi-dimensional and includes: personnel information (such as operator ID, technical level, and historical excellent rate), equipment information (such as total station model, GNSS receiver ID, UAV sensor parameters, and data processing workstation configuration), operating environment parameters (such as temperature, humidity, satellite visibility, and network bandwidth during field data collection), and processing software information (such as the version number of the data processing software and the plugin version of the data processing software).

[0060] Each specific processing step (e.g., data acquisition, data preprocessing, feature extraction, or manual editing) is abstractly defined as a node in a graph, and the flow path of surveying data between these steps is defined as directed edges in the graph, thus forming a flowchart reflecting the actual business logic. Next, a large amount of historical production log data is retrieved. These logs must contain the aforementioned multi-dimensional quality attribute information and the final quality results of each step (e.g., data rework records, errors in quality inspection reports, etc.). Using this historical data, a machine learning model (e.g., gradient boosting decision tree or deep neural network) is trained. This model learns the correlation between specific combinations of node attributes and the probability of quality problems occurring. The model's output is the "risk weight" for a specific attribute combination (e.g., "Employee A" + "Equipment B" + "Environment C" + "Software D"). For example, if historical data shows a significantly high rework rate for a particular combination, a higher risk weight value is assigned to that combination.

[0061] After calculating the risk weights, these weights are integrated into the corresponding nodes of the process topology diagram as a key attribute.

[0062] After this step, a static business process diagram is transformed into a dynamic process topology diagram that integrates multi-dimensional attributes and associated risks. This enhanced topology diagram not only depicts the data flow path but also reveals the potential risk level of each link in the process under specific conditions in a quantitative way. In order to understand the nonlinear propagation and accumulation pattern of risks in complex processes, this embodiment uses a graph neural network to construct the final process topology model. A large number of risk process topology diagrams generated in history are used as training samples to train the graph neural network. The training goal is to enable the network to learn the transmission law of risk weights in the entire graph structure. For example, it can identify how the risks of certain upstream nodes are amplified or suppressed by specific downstream links. After training, this network constitutes an analytical model capable of risk prediction. When a new mapping data stream is generated, a corresponding real-time risk process topology diagram is generated for it and input into the trained model for inference. The model will output the cumulative risk score or compliance deviation prediction frequency of each node. By analyzing these output values, the frequency of compliance deviations can be quantified, and those nodes with the greatest impact on overall quality and the most significant risk accumulation effect can be automatically identified as key quality control points.

[0063] By transforming traditional, passive, experience-based quality inspection into a proactive, data-driven risk prediction and monitoring model, and leveraging graph neural inference networks to deeply learn the nonlinear transmission and cumulative effects of multi-dimensional factors such as personnel, equipment, and environment in complex business processes, this method can automatically and accurately identify critical quality control points. This allows for the centralized optimization of quality control resources, significantly improving monitoring efficiency and data production quality, and achieving intelligent and refined management.

[0064] For each node in the process topology diagram, a sensitivity analysis oriented towards business outcomes is performed using the constructed process topology model. This analysis involves artificially introducing a quality fluctuation that conforms to its historical patterns into a specific node (such as "data preprocessing") during model inference. For example, based on historical data statistics, an 80th quantile accuracy deviation occurs when simulating the output data of this stage. Then, under this simulation condition, model inference is continued to observe and calculate the degree of negative impact of this local fluctuation on the business value of the final deliverable of the entire process (such as the accuracy level, integrity score, or project rating of the final data). The degree of impact is quantified, and the resulting value is the business importance index of this node. This process is repeated for all nodes in the diagram to obtain the business importance of each node.

[0065] In parallel, to assess the stability of the quality of each process step, statistical analysis is performed based on historical audit logs. Historical quality audit and compliance inspection records related to surveying and mapping production are retrieved, and the actual frequency of occurrence of historical compliance deviations (such as errors, defects, rework events, etc.) associated with each node in the process topology diagram is precisely calculated. For example, if a manual revision node was executed 1,000 times in the past year, and 50 of them resulted in quality issues requiring rework, then its deviation frequency is 5%. This calculated frequency value is defined as a management indicator characterizing the stability of the node. The higher the value of this indicator, the more unstable the historical performance of the node.

[0066] The obtained business importance indicators and stability indicators are weighted and combined to calculate a comprehensive management priority score for each node in the diagram. The weighting coefficients (e.g., business importance accounts for 70%, stability accounts for 30%) can be preset according to specific management objectives to reflect the different emphases on the two types of risks: those with high impact and those with many problems. After calculating the scores of all nodes, the scores are sorted, and the nodes with the highest comprehensive management priority scores (or the top-ranked nodes) are finally determined as the quality critical control points under the current business process. These control points are those that would cause the most serious consequences if problems occur, and are the highest priority objects for quality supervision and resource investment.

[0067] By weighting the business importance of nodes (i.e., their potential impact on the final outcome) with their historical stability (i.e., the frequency of historical problems), we can accurately distinguish between critical risk points that have a significant but occasional impact and chronic hidden danger points that have a minor but frequent impact. This ensures that the key control points we ultimately identify not only reflect the probability of risk occurrence but also the severity of its consequences. As a result, we can guide quality management resources to focus on the links that truly have a decisive impact on overall quality in the most efficient way.

[0068] Using identified critical quality control points as targets, the analysis focuses on the entry, exit, and processing of mapping data stream segments within these critical nodes. The dynamic quality attributes to be extracted are defined by pre-defined business rules. For example, one business rule can define data integrity as the number of point clouds or the missing rate of key fields before and after data frame transmission; another rule can define data processing latency as the difference between the timestamps of data entering and leaving a critical control point. These rules provide clear objectives for subsequent time-series analysis.

[0069] To accurately capture dynamic quality attributes, a multi-scale adaptive temporal modeling framework is adopted. The core of this framework lies in combining graph convolutional networks with spatiotemporal collaborative modeling techniques. First, the various indicators of multi-source business data (such as surveying data generation time, processing time, equipment ID, operator ID, etc.) generated along with the data flow are regarded as nodes in a graph. The edges of the graph are constructed based on the prior associations between these indicators or the potential correlations obtained through learning. The graph convolutional network is used to propagate information on this graph structure, thereby capturing the complex spatial dependencies between different business data indicators at each time slice (e.g., the variation pattern of processing time when a specific operator uses a specific combination of equipment).

[0070] Building upon graph neural networks that capture spatial dependencies, a spatiotemporal collaborative modeling unit is introduced to learn the evolution of the entire graph state over time. This combination enables the model to understand both the mutual influence of factors at the same moment and the changing trends of these influences over time. This multi-scale framework can simultaneously analyze data from different time windows, effectively capturing both long-term dependencies (such as the slow degradation of equipment performance) and short-term dependencies (such as instantaneous transmission delays caused by network congestion) in surveying and mapping data streams. Ultimately, it outputs a series of quantified, dynamically changing index sequences. Based on the aforementioned business rules, these output sequences are automatically determined and labeled as corresponding dynamic quality attributes, such as data integrity, data accuracy, data consistency, data transmission delay, data processing delay, and surveying and mapping equipment stability, providing input for subsequent risk index generation. Through spatiotemporal collaborative modeling, a deep insight into data quality from static slicing to dynamic evolution is achieved, comprehensively understanding the immediate correlation and long-term evolution of multiple influencing factors.

[0071] By applying multi-scale time series analysis only at key quality control points, we not only greatly improve analysis efficiency, but also use graph convolution and spatiotemporal collaborative modeling to deeply explore the potential correlations and dynamic evolution patterns between various business data. This makes quality monitoring no longer limited to isolated indicator threshold judgments, but can provide early warnings of potential quality decline risks based on the overall health status and long-term trends of the data flow, achieving a higher-dimensional real-time quality insight and prediction capability.

[0072] After acquiring the multidimensional dynamic quality attributes of the surveying and mapping data stream, a quality risk prediction model is established to aggregate these discrete attributes into a single and intuitive quality risk index. A gradient boosting tree is used as the core of the quality risk prediction model, which is constructed in an iterative and progressive manner. First, a preliminary, relatively simple decision tree (i.e., a weak learner) is generated based on the training data. Then, the residual or gradient between the tree's prediction and the true quality label is calculated. A new decision tree is then built. The goal of this new tree is no longer to directly predict the quality label, but to fit the residual generated in the previous step. In this way, each subsequent decision tree corrects the accumulated errors of all previous trees. This process is repeated, constructing a series of decision trees through progressively weighted learning. By adaptively adjusting the weights of the leaf nodes of each decision tree, a powerful ensemble model is finally formed. This model can effectively capture the complex nonlinear relationship between dynamic quality attributes (such as data integrity, processing latency, equipment stability, etc.) and the final quality risk, outputting a continuous numerical value, i.e., the quality risk index.

[0073] To ensure the robustness and generalization ability of the model, cross-validation is used in the initial training and validation process. The prepared historical dataset (containing multi-dimensional dynamic quality attributes and their corresponding true quality results) is divided into several disjoint subsets (i.e., "folds"). In each round of training, one subset is selected as the validation set, and the remaining subsets are used as the training set to train the gradient boosting tree model. This process is repeated until each subset has been used as a validation set at least once. By comprehensively evaluating the model's performance on all validation sets, a reliable estimate of its predictive performance can be obtained, and the model's hyperparameters can be tuned accordingly. This effectively avoids overfitting the model to specific training data and ensures that it maintains high accuracy when facing new and unseen data.

[0074] Further reinforcement learning mechanisms are introduced for continuous optimization. Under this framework, the quality risk index generated by the model can be regarded as an action. After the data stream undergoes subsequent processing or manual quality inspection, its actual quality fluctuations (e.g., ultimately being rated as qualified, requiring rework, or discarded) are revealed. The risk index predicted by the model is compared with this actual result, and the deviation between the two is used as a feedback signal (i.e., reward or punishment). For example, if the model predicts low risk but the data is ultimately judged as unqualified, the model will receive negative feedback. This feedback signal is used to adjust the internal parameters or weights of the gradient boosting tree model, with the goal of enabling the model to make decisions that maximize cumulative rewards in future predictions. Through this continuous "prediction-validation-feedback-optimization" closed loop, the model can continuously learn from new real-world cases, achieving dynamic and adaptive optimization of the prediction results.

[0075] By using a gradient boosting tree model, multidimensional and dynamic quality attributes are efficiently aggregated into a single, intuitive quality risk index, simplifying complex quality states and providing a clear basis for immediate decision-making. By introducing a continuous feedback and optimization mechanism of reinforcement learning, the limitations of traditional prediction models that become fixed and unchanging after training are completely changed. In a dynamic production environment, this model can continuously learn from the deviation between its predictions and actual results, achieving long-term, adaptive accuracy improvement and ensuring the long-term effectiveness and reliability of quality risk assessment.

[0076] To address the autocorrelation problem prevalent in surveying and mapping production data—that is, data points originating from the same business entity (such as the same personnel or equipment) often have inherent correlations, and random partitioning can lead to information leakage—a hierarchical blocking cross-validation based on business dimensions is constructed. This method groups the complete historical dataset (including logs and real quality labels) based on one or more key quality attributes representing the execution of business processes (such as the personnel, equipment, or operating environment parameters). For example, if the personnel are chosen as the grouping dimension, all data generated by operator A will be grouped into one data block, all data generated by operator B into another block, and so on, dividing the entire dataset into several independent data blocks bound to business entities.

[0077] After the block partitioning is completed, in each iteration of cross-validation, the basic unit for dividing the training and validation sets is no longer a single data point, but the entire data block. Specifically, in one iteration, one or more complete data blocks are selected as the validation set, while all other remaining blocks constitute the training set. For example, in 5-fold cross-validation with the executor as the dimension, one iteration may use all the data of operator A as the validation set, and use the data of all other operators such as B, C, D, and E to train the model. In this way, it can be ensured that the training and validation sets are completely isolated in the selected business dimension, thereby effectively testing the model's true generalization ability when facing a completely unseen business entity (such as a new operator or a new device), eliminating the inflated evaluation results caused by internal data correlation, and obtaining a more accurate evaluation of model performance that is more in line with the actual application scenario.

[0078] By adopting a hierarchical blocking cross-validation method based on business dimensions, the problem of internal correlation that is common in surveying and mapping production data is fundamentally solved. This method eliminates the problem of inflated model evaluation results caused by information leakage in standard random validation. The method forces the model to learn the general rules across different business entities (such as different operators or equipment) rather than memorizing the individual characteristics of specific entities. Therefore, its validation results can more realistically reflect the model's generalization ability when facing new business scenarios (such as new employees joining the company), providing a more reliable and trustworthy performance guarantee for the actual deployment of the model.

[0079] Once the quality risk index of the surveying and mapping data stream is generated, if its value exceeds a preset threshold, it will automatically trigger the business compliance decision-making process and enter the data correction stage. The core of this decision-making process is a generative adversarial network that serves as a quality decision-making model. In this stage, the previously calculated quality risk index is not directly used as the input to the generative network, but is used as a key business strategy variable. Specifically, the risk index is first converted into a modulation vector through a small mapping network, which is used to dynamically and finely adjust the scope and degree of subsequent data correction.

[0080] The modulation vector is injected into multiple layers of the generator network of the generative adversarial network. It acts like a continuously adjustable knob. The higher the risk index, the stronger the adjustment effect of the modulation vector. This will guide the generator network to apply a wider range or deeper degree of adjustment when correcting the original data. In order to generate multiple correction schemes with diversity, n different random noise vectors are sampled at the input of the generator. Each sampling, under the uniform constraint of the modulation vector, will drive the generator to produce a decision candidate version with slightly different content but all in line with the repair target under the current risk level. This results in n selectable data versions that have been corrected in different ways.

[0081] After obtaining n decision candidate versions, the optimal version needs to be selected. The core is to calculate the weighted difference between each decision candidate version and the original data to measure the loss of data fidelity. The weights here directly reuse the business importance index calculated for each data point by the previous process topology model. This means that modifications to key data points with higher business value will generate a much higher difference score than modifications to ordinary data points, quantifying the potential impact of each modification on core business value and prioritizing the protection of important data.

[0082] The final solution is determined through an optimal decision-making selection process. This process calculates a comprehensive business cost for each candidate version. This cost is a multi-dimensional evaluation result obtained through a multi-objective optimization algorithm, specifically including three core components: 1) Residual quality risk: The data of the candidate version is re-input into the risk prediction model to obtain its corrected residual quality risk index, representing the effectiveness of the repair; 2) Data fidelity cost: The weighted difference calculated in the previous step is directly used, representing the degree of deviation of the modification from the original information; 3) Business execution cost: The computational resources and time consumption of the data processing operations required to generate the candidate version are assessed. Taking into account these three costs, the total cost of each candidate version is calculated. The candidate version with the lowest cost is ultimately selected as the optimal decision and used to automatically replace the original high-risk data fragments, completing the entire closed-loop supervision and disposal process. By unifying the three mutually constraining core dimensions of repair effectiveness, data fidelity, and business execution cost into the comprehensive business cost consideration, the one-sided decision-making that may be caused by single-objective optimization is completely avoided.

[0083] Generative adversarial networks can dynamically generate multiple appropriate and controllable remediation candidate versions based on risk indices, enabling flexible decision-making. Meanwhile, the optimal version selection mechanism adopts multi-objective optimization, seeking the best balance between remediation effectiveness, data fidelity (especially for high business value data), and execution cost. This ensures that the final automatic replacement decision is not only effective but also the globally optimal solution under comprehensive consideration, greatly improving the automation level of data quality supervision and the intelligence level of decision-making.

[0084] This application constructs a closed-loop intelligent quality supervision system covering the entire process from pre-event warning and in-event monitoring to post-event automatic correction. First, through macro-modeling of the business process, it can proactively identify the key control points most prone to quality problems, achieving source prediction of risks. Second, focusing on these key points, it conducts in-depth analysis of real-time data streams, quantifying the complex and ever-changing quality status into an intuitive risk index, achieving accurate and dynamic real-time monitoring. Based on this risk index, it automatically triggers and executes optimal correction decisions, completing a seamless connection from problem discovery to problem resolution. This changes the traditional quality inspection model that relies on manual and delayed processing, greatly improving the automation level, processing efficiency, and reliability of the final results of surveying and mapping data production.

[0085] Example 2:

[0086] Unmanned aerial photogrammetry (UAV) is one of the mainstream technologies in the current surveying and mapping field. Its operation process is long and there are many influencing factors (such as flight attitude, environmental changes, software algorithms, etc.), which leads to high uncertainty in data quality. This embodiment provides a surveying and mapping data quality supervision system based on machine learning to supervise the quality of surveying and mapping data.

[0087] See Figure 3 When a drone mapping operation is initiated, the data processing module receives the aerial survey data stream transmitted from the drone in real time. This includes the original aerial photographs (image data) and high-frequency POS data (position and attitude) from the flight control log. It also retrieves and records quality attribute information related to the operation, such as: the pilot ID of the pilot who performed the flight (e.g., Zhang San), the model of the drone and sensor used (e.g., DJI M300 RTK + P1 camera), the environmental parameters of the operation area at that time (e.g., wind speed 5 m / s, good lighting conditions), and the version number of the ground station software used for data processing (e.g., DJI Terra v3.9).

[0088] After receiving the above information, the key identification module immediately constructs a process topology model for this task based on historical data from similar projects. This model includes nodes such as flight path planning, data acquisition, POS calculation, aerial triangulation, and dense matching. The analysis engine discovers through analysis of historical logs that under the condition of "wind speed of 5 m / s", the failure rate or reprocessing rate of the aerial triangulation node has increased significantly in the past. Combined with sensitivity analysis, it is confirmed that if a problem occurs in this node, it will directly lead to the entire project's accuracy failing to meet the standard. Therefore, the system automatically identifies and marks "aerial triangulation" as the key quality control point for this task.

[0089] The quality risk module utilizes a time-series analysis model to specifically monitor the data flow input to this stage and determine its dynamic quality attributes in real time. For example, it detected a slight drift in the accuracy of the POS data after processing (dynamic quality attribute), and the processing time for image tie point matching (dynamic quality attribute) was 15% longer than normal. Based on the combination of these dynamic attributes, the quality risk prediction model comprehensively judged that there was a high risk of model failure and generated a quality risk index as high as 0.82 for the data flow being processed.

[0090] A risk index as high as 0.82 immediately triggered the quality correction module. The quality decision model used this index as a strong strategy variable to control the degree of repair, generating three candidate versions: Version A removed images with the lowest tie point matching and recalculated; Version B kept all images but automatically optimized camera intrinsic parameters and increased the number of iterations; Version C introduced historical images from nearby areas as auxiliary constraints for joint adjustment. Subsequently, the system evaluated the three versions: Version A had the lowest residual risk but high data fidelity cost; Version B had a balanced cost across all aspects; Version C had the best effect but the highest business execution cost (retrieving historical data and increasing calculations); Finally, the multi-objective optimization algorithm calculated that Version B had the lowest overall business cost, so the system selected the aerial triangulation result of Version B to automatically replace the original high-risk solution result.

[0091] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine learning-based method for monitoring the quality of surveying and mapping data, characterized in that, include: The system receives real-time surveying data streams and synchronously acquires quality attribute information characterizing the execution of business processes. Based on this quality attribute information, a process topology model is constructed. The quality attribute information includes executors, equipment, operating environment parameters, and processing software versions. The process topology model utilizes an analysis engine to quantify the frequency of compliance deviations and identify key quality control points. The analysis engine is composed of a graph neural network inference network. The process of identifying key quality control points specifically involves: Each processing step of the quality attribute information is represented as a node in the graph, and the flow relationship of the surveying data stream in the execution of the business process is represented as a directed edge in the graph. Based on historical logs, risk weights are assigned to node attributes through machine learning, and a process topology graph integrating multi-dimensional attributes and associated risks is constructed. Using the aforementioned process topology model, a sensitivity analysis oriented towards business outcomes is performed on each node in the process topology diagram. The sensitivity analysis is the degree of impact on the business value of the final deliverable of the entire process when the business output of the node experiences quality fluctuations that conform to historical patterns during model simulation, and the degree of impact is quantified into a business importance indicator. The frequency of historical compliance deviations associated with each node is statistically analyzed based on historical audit logs, serving as a management indicator characterizing the stability of that node. The business importance index and the stability management index are weighted and combined to calculate a comprehensive management priority score for each node, and the node with the highest comprehensive management priority score is determined as the quality critical control point. Based on key quality control points and data lineage tracing, the lifecycle of surveying and mapping data streams in the business process is traced. A time-series analysis method is used to automatically determine the dynamic quality attributes of the surveying and mapping data streams. This method employs a multi-scale adaptive time-series modeling framework, combining graph convolutional networks and spatiotemporal co-modeling, simultaneously analyzing data from different time windows to capture the long-term and short-term dependencies of the surveying and mapping data streams. The dynamic quality attributes are defined by business rules and include data integrity, data accuracy, data consistency, data transmission latency, data processing latency, and surveying equipment stability. A quality risk prediction model is established to generate a quality risk index for the surveying and mapping data streams based on these dynamic quality attributes. Based on the quality risk index, a business compliance decision-making process is triggered. The quality risk index is used as a business strategy variable and input into the quality decision-making model to adjust the decision scope and degree of the surveying and mapping data stream. The quality decision-making model is a generative adversarial network. Its generator maps the quality risk index into a modulation vector and injects it into the network layer to control the scope and degree of repair. By sampling the input random noise vector multiple times, multiple decision candidate versions are generated. The difference between each decision candidate version and the surveying and mapping data stream is calculated, and the optimal candidate version is selected for automatic replacement.

2. The method for monitoring the quality of surveying and mapping data based on machine learning according to claim 1, characterized in that, The process topology model constructed based on the quality attribute information is as follows: A process topology model is constructed based on a process topology graph training analysis engine. The analysis engine consists of a graph neural inference network, which is used to learn the nonlinear propagation and accumulation pattern of the risk weights in the process topology graph structure, quantify the frequency of compliance deviations, and identify key quality control points.

3. The method for monitoring the quality of surveying and mapping data based on machine learning according to claim 1, characterized in that, The specific process of establishing a quality risk prediction model and generating a quality risk index for the mapping data stream based on the dynamic quality attributes is as follows: The quality risk prediction model uses a gradient boosting tree to generate a quality risk index for the mapping data stream based on the dynamic quality attributes. The gradient boosting tree constructs a series of decision trees by progressively weighting weak learners, and adaptively adjusts the weights of the leaf nodes of each decision tree to reflect the quality risk index. Cross-validation is used for training and validation, and reinforcement learning is used to continuously provide feedback on the deviation between the model's prediction results and the actual quality fluctuations, thereby optimizing the prediction results.

4. The machine learning-based surveying data quality supervision method according to claim 3, characterized in that, The specific process of using cross-validation for training and validation is as follows: A hierarchical blocking cross-validation based on business dimensions is adopted, which groups historical logs and their corresponding real quality labels into data blocks according to the quality attribute information that represents the execution of business processes; and in each iteration of cross-validation, the complete data block is used as the validation set, and the remaining blocks are used as the training set.

5. The method for monitoring the quality of surveying and mapping data based on machine learning according to claim 1, characterized in that, The calculation of the difference between each decision candidate version and the mapping data stream, and the selection of the optimal candidate version for automatic replacement, specifically involves: This is accomplished collaboratively by the business value impact assessment module and the optimal decision selection module. The business value impact assessment module calculates the weighted difference between each decision candidate version and the original data, where the weight is determined by the business importance index calculated by the process topology model for each data point, and a higher difference score will be generated when it is modified. The optimal decision selection module executes a multi-objective optimization algorithm to calculate the comprehensive business cost for each candidate version and selects the version with the lowest cost as the optimal version. The comprehensive business cost includes: the residual quality risk index of the candidate version after correction, the weighted difference calculated by the business value impact assessment module as the data fidelity cost, and the business execution cost corresponding to the data processing operations required to generate the version.

6. A machine learning-based surveying and mapping data quality supervision system, characterized in that, include: Data processing module: Receives surveying and mapping data streams in real time and synchronously acquires quality attribute information representing the execution of business processes; Key Identification Module: Based on the quality attribute information, a process topology model is constructed. This quality attribute information includes executors, equipment, operating environment parameters, and processing software version. The process topology model utilizes an analysis engine to quantify the frequency of compliance deviations and identify key quality control points. The analysis engine is composed of a graph neural network inference system. The process of identifying key quality control points specifically involves: Each processing step of the quality attribute information is represented as a node in the graph, and the flow relationship of the surveying data stream in the execution of the business process is represented as a directed edge in the graph. Based on historical logs, risk weights are assigned to node attributes through machine learning, and a process topology graph integrating multi-dimensional attributes and associated risks is constructed. Using the aforementioned process topology model, a sensitivity analysis oriented towards business outcomes is performed on each node in the process topology diagram. The sensitivity analysis is the degree of impact on the business value of the final deliverable of the entire process when the business output of the node experiences quality fluctuations that conform to historical patterns during model simulation, and the degree of impact is quantified into a business importance indicator. The frequency of historical compliance deviations associated with each node is statistically analyzed based on historical audit logs, serving as a management indicator characterizing the stability of that node. The business importance index and the stability management index are weighted and combined to calculate a comprehensive management priority score for each node, and the node with the highest comprehensive management priority score is determined as the quality critical control point. Quality Risk Module: Based on key quality control points and data lineage tracing, the module traces the lifecycle of surveying and mapping data streams within the business process. It automatically determines the dynamic quality attributes of the data streams using time-series analysis. This analysis employs a multi-scale adaptive time-series modeling framework, combining graph convolutional networks with spatiotemporal collaborative modeling. It simultaneously analyzes data from different time windows to capture the long-term and short-term dependencies of the surveying and mapping data streams. The dynamic quality attributes are defined by business rules and include data integrity, data accuracy, data consistency, data transmission latency, data processing latency, and surveying equipment stability. A quality risk prediction model is established, and a quality risk index is generated for the surveying and mapping data streams based on these dynamic quality attributes. Quality Correction Module: Based on the quality risk index, a business compliance decision-making process is triggered. The quality risk index is used as a business strategy variable and input into the quality decision-making model to adjust the decision range and degree of the mapping data stream. The quality decision-making model is a generative adversarial network. Its generator maps the quality risk index into a modulation vector and injects it into the network layer to control the scope and degree of repair. By sampling the input random noise vector multiple times, multiple decision candidate versions are generated. The difference between each decision candidate version and the mapping data stream is calculated, and the optimal candidate version is selected for automatic replacement.

Images