Engineering false visa auditing method and system based on multi-source time series causal inference

Abstract

The application discloses an engineering false visa audit method and system based on multi-source time sequence causal inference, and relates to the field of engineering digital management.The method comprises the following steps: obtaining multi-source heterogeneous time sequence data of an engineering project and an engineering visa event to be audited, performing multi-resolution time alignment on data with different sampling frequencies, constructing a unified time sequence matrix, establishing a directed time sequence interaction graph based on lag correlation, and extracting a time sequence graph embedding representation; further performing invariant causal discovery in multiple environments corresponding to different construction states and / or data source attributes, calculating the causal influence strength between data sources, constructing a target causal network corresponding to the visa event, and optimizing the target causal network in combination with engineering design specifications, contract clause constraints and historical experience models; and calculating an audit causal score and a causal completeness index to identify the false risk of the engineering visa event. Thus, the process restoration capability, causal explainability and risk identification accuracy of the visa audit are improved.

Description

Technical Field

[0001] This invention belongs to the field of digital management technology for engineering projects, and in particular relates to a method and system for auditing false visas in engineering projects based on multi-source temporal causal inference. Background Technology

[0002] With the continuous improvement of digitalization in the construction engineering field, a large amount of data from different sources, of diverse types, and with obvious time evolution characteristics has been gradually accumulated in various stages of engineering projects, including design, construction, monitoring, schedule management, cost control, and contract performance. This data, to a certain extent, reflects the formation process, implementation status, and changes in results of engineering activities, providing richer basic information for engineering management and auditing than traditional paper documents, and making it possible to conduct more refined identification and analysis throughout the entire engineering process.

[0003] During project implementation, engineering change orders are typically used to supplement and confirm changes, adjustments to quantities, additional measures, or schedule variations that occur during construction. They are a crucial management mechanism for ensuring smooth project progress. However, in practice, change order matters often involve multiple stages, including design basis, site conditions, construction process, resource input, and settlement results, resulting in a relatively long chain of connections and complex influencing factors. For such highly procedural and interconnected matters, relying solely on scattered data for post-event verification often fails to fully reconstruct the business context in which the change order arose, and it is also difficult to make a more comprehensive judgment on whether the relevant change orders have sufficient and reasonable factual support.

[0004] Currently, auditing of engineering visas typically relies on manual verification, experience-based analysis, and confirmation of individual documents. While this approach is effective for routine projects or smaller data volumes, the increasing volume, frequent updates, and diverse sources of engineering data necessitate a growing demand for comprehensive comparison and analysis of information across different stages and sources. Furthermore, although relevant digital management platforms can achieve a certain level of data aggregation, classification management, and process tracking, the formation relationship between visa matters and related engineering facts usually requires further in-depth analysis and identification, considering temporal evolution characteristics, business connections, and engineering rules. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a method and system for auditing fraudulent engineering visas based on multi-source temporal causal inference.

[0006] The technical solution of the present invention is as follows:

[0007] A method for auditing fraudulent engineering visas based on multi-source temporal causal inference, the method comprising:

[0008] Acquire multi-source heterogeneous time-series data of engineering projects and engineering visa events to be audited. The multi-source heterogeneous time-series data comes from multiple engineering business data sources, and the multiple engineering business data sources have different sampling frequencies.

[0009] A multi-resolution dynamic time normalization algorithm is used to perform time alignment processing on the multi-source heterogeneous time series data to generate a time series matrix corresponding to a unified time axis.

[0010] Based on the time series matrix, the lag correlation between different engineering business data sources is calculated to construct a directed time series interaction graph, and the time series graph embedding representation of each engineering business data source in the directed time series interaction graph is extracted;

[0011] Based on the time series graph embedding representation, multi-source time-invariant causal discovery is performed in multiple pre-divided environments to determine the residual variation characteristics and parameter consistency characteristics of the regression model under different environments, and to calculate the causal influence strength between different engineering business data sources based on the residual variation characteristics and parameter consistency characteristics; each environment is used to indicate the corresponding construction status and / or data source attributes;

[0012] Based on the strength of the causal influence, a target causal network corresponding to the engineering visa event is constructed, and the target causal network is optimized in combination with preset engineering domain knowledge constraints to adjust the weights of the corresponding causal edges in the target causal network; the engineering domain knowledge constraints include at least one of the following: engineering design specifications, contract terms constraints, and historical experience models;

[0013] The audit causality score and causal completeness index of the engineering visa event are calculated based on the weighted target causal network, and the presence of false risk of the engineering visa event is identified based on the audit causality score and the causal completeness index.

[0014] Furthermore, the step of performing time alignment processing on the multi-source heterogeneous time series data using a multi-resolution dynamic time normalization algorithm to generate a time series matrix corresponding to a unified time axis includes:

[0015] Obtain the time series and corresponding timestamp sequences of different engineering business data sources in the multi-source heterogeneous time series data;

[0016] A weighted distance measure function with a time difference penalty term is constructed. The weighted distance measure function is used to measure the difference between sequential sample points by multiplying the norm distance of the sample feature difference vector with the exponential weight function based on time penalty. The exponential weight function uses a preset penalty hyperparameter to control the sensitivity of path alignment to time difference.

[0017] Downsampling is performed on the time series of the different engineering business data sources to obtain the corresponding coarse-grained time series, and the global cumulative distance between the coarse-grained time series is calculated based on the weighted distance metric function. The coarse-grained alignment path that minimizes the global cumulative distance is then extracted.

[0018] The coarse-grained alignment path is mapped back to the original sampling resolution, and a local search constraint boundary with a preset time radius is set with the mapped path as the center.

[0019] Within the local search constraint boundary, the weighted distance metric function is invoked to perform fine-resolution dynamic time-normalized matching iteration to obtain the corresponding fine-grained time-aligned index mapping relationship;

[0020] Using the timestamp sequence of the preset benchmark engineering business data source as a unified time axis reference, and based on the fine-grained time alignment index mapping relationship, sampling point alignment processing is performed on each of the other engineering business data sources to generate a time series matrix corresponding to the unified time axis.

[0021] Further, the step of calculating the lag correlation between different engineering business data sources based on the time series matrix to construct a directed time series interaction graph, and extracting the time series graph embedding representation of each engineering business data source in the directed time series interaction graph, includes:

[0022] Zero-mean standardization is performed on the time series of each engineering business data source in the time series matrix to eliminate dimensional differences and baseline drift between different engineering business data sources;

[0023] Set the maximum time-series lag window length and the lag decay coefficient that decreases over time;

[0024] For any two engineering business data sources after standardization, the cumulative value of the cross sequence feature product is calculated within the lag window and the absolute value is taken, based on the historical observation value of the first engineering business data source and the current observation value of the second engineering business data source. Combined with the lag attenuation coefficient, the lag influence score in the corresponding direction is calculated.

[0025] A preset significant noise truncation threshold is set, and asymmetric sparse filtering based on bidirectional comparison is performed on the lag effect score: when the positive lag effect score is strictly greater than the negative lag effect score and greater than the significant noise truncation threshold, the positive lag effect score is confirmed as an effective lag association and assigned as a directed edge weight; otherwise, the corresponding directed edge weight is reset to zero to construct an asymmetric adjacency matrix of the directed temporal interaction graph with sparse noise reduction characteristics.

[0026] The symmetric fusion matrix of the asymmetric adjacency matrix is ​​extracted, and the degree distribution of each network node is calculated based on the symmetric fusion matrix to generate a degree matrix. Then, the degree matrix and the symmetric fusion matrix are combined to construct an undirected Laplacian matrix for graph analysis.

[0027] The undirected Laplacian matrix is ​​subjected to eigenvalue decomposition to extract eigenvectors corresponding to the first few smallest non-zero eigenvalues, which are then concatenated to form a dimension-reduced feature matrix. The row vectors in the dimension-reduced feature matrix corresponding to each engineering business data source are used as the time-series graph embedding representation of each engineering business data source.

[0028] Furthermore, the step of performing multi-source time-invariant causal discovery based on the time-series graph embedding representation in pre-divided multiple environments to determine the residual variation characteristics and parameter consistency characteristics of the regression model under different environments, and calculating the causal influence strength between different engineering business data sources based on the residual variation characteristics and parameter consistency characteristics, includes:

[0029] Based on the time-series graph embedding representation of each of the engineering business data sources, the embedding vector similarity between the target engineering business data source and the other engineering business data sources is calculated, and the engineering business data sources with embedding vector similarity greater than a preset similarity threshold are extracted as a set of candidate parent variables for constructing a regression model for the target engineering business data source, so as to achieve dimensionality reduction of the causal inference search space.

[0030] For each pre-divided environment in the set of multiple pre-divided environments, for any test engineering business data source in the set of candidate parent variables, a baseline regression model that does not contain the test engineering business data source and a conditional regression model that contains the test engineering business data source are constructed respectively; wherein, the conditional regression model in each environment adopts a unified parameterized model structure.

[0031] By fitting the baseline regression model and the conditional regression model under all environments, the baseline fit residuals and conditional fit residuals under all environments are aggregated to calculate the variance of the baseline residuals and the variance of the conditional residuals across environments, and the residual change characteristics are determined based on the relative reduction ratio of the baseline residual variance and the conditional residual variance.

[0032] Extract the model parameter vectors of the conditional regression model under various environments, and determine the parameter consistency characteristics by calculating the vector space distance between the model parameter vectors under other environments and the model parameter vectors under a preset reference environment.

[0033] The residual variation characteristics are multiplied by the average satisfaction of the parameter consistency characteristics across environments after removing the reference environment, in order to calculate and quantify the causal influence strength from the test engineering business data source to the target engineering business data source.

[0034] Furthermore, the optimization of the target causal network by incorporating preset engineering domain knowledge constraints to adjust the weights of corresponding causal edges in the target causal network includes:

[0035] At least one of the engineering design specifications, contract terms and constraints and historical experience model is parsed and abstracted into entity relation triples to construct an engineering domain knowledge graph, and the engineering domain knowledge graph is transformed into a prior knowledge matrix with constraint strength labels.

[0036] For any target causal edge in the target causal network, the data source business attributes of its head node and tail node are mapped to the corresponding coordinate positions of the prior knowledge matrix in order to extract the domain logic matching degree coefficient corresponding to the target causal edge.

[0037] Based on the engineering domain knowledge constraints, the target causal edge is subjected to rule conflict detection: when the target causal edge indicates the occurrence of additional cost causal derivation, and the engineering sub-items corresponding to the head node and tail node are identified as non-additional cost items in the contract terms constraints, or when the evolution trend of the target causal edge violates the topological constraint relationship of the engineering design specification, a domain logic conflict is determined to be triggered.

[0038] When the domain logic conflict is triggered, a penalty attenuation factor greater than zero is determined based on the domain logic matching degree coefficient, and the penalty attenuation factor decreases as the domain logic matching degree coefficient increases; when the domain logic conflict is not triggered, the penalty attenuation factor is set to zero.

[0039] Based on the original causal influence strength of the target causal edge and the determined penalty attenuation factor, a nonlinear soft constraint modulation function based on prior knowledge is invoked to transform the penalty attenuation factor into a nonlinear attenuation effect on the original causal influence strength, so as to calculate the causal influence strength after weight adjustment and use it to reconstruct the target causal network.

[0040] Furthermore, the calculation of the audit causality score and causal completeness index of the engineering visa event based on the weighted target causal network includes:

[0041] The target causal network is reconstructed using the adjusted weights of the causal influence strength, and a directed graph path search algorithm is called to extract a directed connected subgraph from the reconstructed target causal network that points from the underlying engineering business data source to the engineering visa event node, and the set of causal edges contained in the directed connected subgraph is taken as the key causal edge set.

[0042] Extract the adjusted weighted causal influence strength of each causal edge in the key causal edge set, and sum it with the preset engineering business importance weight coefficient assigned to the corresponding causal edge to calculate the audit causal score used to assess the logical support strength of the visa event.

[0043] For the multiple engineering business data sources associated with the engineering visa event, each is taken as the engineering business data source to be tested. Based on the topological dependency relationship of the reconstructed target causal network, the set of effective parent nodes pointing to the engineering business data source to be tested is extracted.

[0044] The actual observed features of the effective parent node set are input into the conditional regression model corresponding to the business data source of the project under test for forward structural equation derivation to obtain the predicted state vector of the business data source of the project under test.

[0045] The norm deviation between the actual observed multidimensional vector and the predicted state vector of each engineering business data source is calculated. Then, the normalized ratio between the square of the norm deviation of each engineering business data source and the square of the norm of the actual observed multidimensional vector is calculated and averaged. The non-negative truncation result after subtracting the normalized ratio is used as the causal completeness index characterizing the consistency of global causal prediction.

[0046] Furthermore, identifying the risk of fraudulent activity in the engineering visa event based on the audit causality score and the causal completeness index includes:

[0047] A first warning threshold is set for the audit causality score, and a second warning threshold is set for the causal completeness index. A two-dimensional risk assessment coordinate system is constructed using the first warning threshold and the second warning threshold as the judgment boundary.

[0048] The engineering visa event is mapped onto the two-dimensional risk assessment coordinate system, and the risk type is determined based on the mapped location;

[0049] When the audit causality score of the engineering visa event is lower than the first warning threshold and the causal completeness index is lower than the second warning threshold, the engineering visa event is identified as an engineering visa event with a risk of fraud and marked as an inflated anomaly.

[0050] When the audit causality score of the engineering visa event is lower than the first warning threshold and the causal integrity index is greater than or equal to the second warning threshold, the engineering visa event is identified as an engineering visa event with a risk of fraud and marked as a link forgery anomaly.

[0051] When the engineering visa event does not meet the above conditions, it will be judged as a non-target abnormal event or an event requiring further review.

[0052] For the identified engineering visa events with false risks, target causal edges with adjusted weights and causal influence strength below a preset strength threshold are extracted from the key causal edge set to identify abnormal causal chain nodes, and a visual audit interactive review report containing the abnormal causal chain nodes is automatically generated.

[0053] Specifically, in the visual audit interactive review report, the fine-grained time alignment index mapping relationship established when performing time alignment processing on the multi-source heterogeneous time series data is invoked to construct a multi-dimensional spatiotemporal tracing link for the abnormal causal chain break node, so as to associate and trace back to locate the corresponding building information model component record and / or the original waveform of the field sensor under the unified time axis.

[0054] An audit system for fraudulent engineering visas based on multi-source temporal causal inference, the system comprising:

[0055] A multi-source data acquisition unit is used to acquire multi-source heterogeneous time-series data of engineering projects and engineering visa events to be audited. The multi-source heterogeneous time-series data comes from multiple engineering business data sources, and the multiple engineering business data sources have different sampling frequencies.

[0056] A multi-resolution alignment unit is used to perform time alignment processing on the multi-source heterogeneous time series data using a multi-resolution dynamic time normalization algorithm to generate a time series matrix corresponding to a unified time axis.

[0057] An interactive graph representation unit is used to calculate the lag correlation between different engineering business data sources based on the time series matrix to construct a directed time series interactive graph, and to extract the time series graph embedding representation of each engineering business data source in the directed time series interactive graph;

[0058] The invariant causal discovery unit is used to perform multi-source time-series invariant causal discovery in multiple pre-divided environments based on the time-series graph embedding representation, in order to determine the residual change characteristics and parameter consistency characteristics of the regression model in different environments, and to calculate the causal influence strength between different engineering business data sources based on the residual change characteristics and parameter consistency characteristics; each environment is used to indicate the corresponding construction status and / or data source attributes.

[0059] The causal network optimization unit is used to construct a target causal network corresponding to the engineering visa event based on the causal influence intensity, and optimize the target causal network in combination with preset engineering domain knowledge constraints to adjust the weights of the corresponding causal edges in the target causal network; the engineering domain knowledge constraints include at least one of the following: engineering design specifications, contract terms constraints, and historical experience models;

[0060] The risk scoring and identification unit is used to calculate the audit causality score and causal completeness index of the engineering visa event based on the target causal network after weight adjustment, and to identify whether there is a false risk in the engineering visa event based on the audit causality score and the causal completeness index.

[0061] An electronic device, comprising:

[0062] At least one processor;

[0063] And a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps in the method as described in any of the preceding claims.

[0064] A storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method as described in any of the preceding claims.

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] (1) Instead of simply summarizing or statically comparing multi-source business data related to engineering visa events, this approach focuses on the continuous evolution of engineering activities. It organizes heterogeneous time-series data from different sources with varying sampling frequencies using a unified time scale, and further constructs a directional interactive structure by combining the temporal relationships between different data sources. Based on this approach, data originally scattered across different stages such as design, construction, monitoring, schedule, cost, and performance can be transformed into a continuous, interconnected representation reflecting the formation process of engineering matters. This allows the business background corresponding to visa events to move beyond isolated records and be depicted within a structure that more closely reflects the actual evolutionary logic of the engineering process. This not only enhances the overall ability to reconstruct the chain of engineering facts related to visa events but also strengthens the ability to organize key preceding events, process transmission relationships, and result response relationships, providing a more process-oriented support foundation for subsequent audit judgments.

[0067] (2) After forming the above-mentioned process correlation structure, it did not stop at general correlation analysis, but further introduced a multi-environment analysis mechanism for different construction states and / or data source attributes to identify invariant causal relationships across sources, and combined engineering domain knowledge to constrain and optimize the target causal network. Through this approach, the influence paths related to visa matters can not only reflect shallow statistical correlation, but also further reflect the degree of substantial stable causal support and the completeness of the support chain, thus making the audit analysis of engineering visa events closer to the real formation logic in engineering business. At the same time, the introduction of knowledge constraints such as engineering design specifications, contract terms, and historical experience models further enhances the business rationality of causal relationship screening and weight adjustment, so that the final causal network has both data-driven objectivity and correctability at the engineering rule level. As a result, the output audit causal score and causal completeness index can more effectively distinguish visa matters with sufficient factual support, weak support chains, or abnormal causal structures, improving the pertinence and credibility of false risk identification results.

[0068] This technical solution organically combines the ability to organize the temporal evolution of multi-source heterogeneous engineering data, the ability to identify stable causality under cross-environmental conditions, and the ability to make audit judgments under the constraints of engineering knowledge. This constructs an audit analysis mechanism for engineering visa events that combines process reproducibility, causal interpretability, and quantifiable results. Based on this mechanism, whether an engineering visa event has true, complete, and reasonable factual support can be identified and expressed in a way that is more in line with engineering business logic, thereby improving the accuracy of identifying false risk events and enhancing the stability and persuasiveness of audit conclusions. Attached Figure Description

[0069] Figure 1 This is a flowchart illustrating the audit method for fraudulent engineering visas based on multi-source temporal causal inference.

[0070] Figure 2 A flowchart illustrating the operation of constructing and extracting the directed time-series interaction graph corresponding to the engineering business data source through time-series graph embedding representation;

[0071] Figure 3 A flowchart illustrating the operation of optimizing the target causal network by incorporating pre-defined engineering domain knowledge constraints;

[0072] Figure 4 A schematic diagram illustrating the system operation mechanism of an audit method for fraudulent engineering visas based on multi-source temporal causal inference;

[0073] Figure 5 A schematic diagram illustrating the comparative simulation results of different methods in identifying various scenarios of fraudulent engineering visas.

[0074] Figure 6 A schematic diagram illustrating the comparative simulation results of different methods for robust identification of false visas in engineering projects under different environmental noise levels.

[0075] Figure 7 A schematic diagram of the 3D response surface for an example of hyperparameter sensitivity analysis;

[0076] Figure 8 This is a structural block diagram of an engineering fraud audit system based on multi-source temporal causal inference. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0078] It should be noted that in the specific management and verification of engineering visas, a common approach in current technologies is to collect relevant data records chronologically and then, based on the reviewer's engineering experience, determine whether there is a reasonable correspondence between the preconditions, construction process, and results. While this method can assist in the verification of some visa matters in engineering practice, it relies heavily on human experience and is typically based on limited data chains or single-type data. As engineering projects gradually shift towards digital management, the sources, frequency, and forms of data associated with visa matters have significantly increased. Simply relying on manual analysis of the chronological order of data is no longer sufficient for continuous and systematic analysis of the relationships between cross-stage and multi-source business information.

[0079] Meanwhile, to enhance the ability to mine data relationships, some time-series data analysis methods exist, identifying the influence relationships between variables by comparing the explanatory or predictive power of historical information on target results. These methods are applicable to time-series scenarios with relatively simple structures and low variable dimensionality. However, in engineering projects, various business systems such as schedule, cost, monitoring, change management, and contract fulfillment often exhibit inconsistent sampling frequencies, inconsistent time spans, and complex relationship structures. Directly applying general time-series analysis models often fails to adequately adapt to the heterogeneity and dynamism of engineering data, and is also unfavorable for making stable and clear identifications of the influence paths related to visa matters.

[0080] Furthermore, regarding digital management of engineering projects, current technologies include multi-source data integration platforms and electronic management systems for visa documents. These solutions aim to achieve unified aggregation, classification, archiving, and process tracking of building information models, geographic information data, construction monitoring information, and various visa materials. While these solutions play a positive role in information integration, process management, and data preservation, their focus is typically on the access, storage, and retrieval of underlying data, or superficial compliance verification. They often lack the ability to conduct further correlation analysis and review to determine whether there is sufficient and reasonable supporting relationship between visa matters and related engineering facts.

[0081] In addition, some studies have attempted to identify relatively stable relationship structures by leveraging differences between different environments or data distributions, thus providing a reference for relationship mining under complex data conditions. However, in the context of engineering visa audits, the relevant data not only exhibits significant multi-source heterogeneity and temporal misalignment characteristics, but is also influenced by factors such as engineering rules, contractual constraints, and business experience. Therefore, how to identify and express the relationships in the formation process of visa matters in a more business-interpretable manner under complex engineering data conditions remains a problem that requires further improvement in current related technologies.

[0082] It should be understood that the above description of the relevant technologies is intended only to help the public better understand the inventive spirit and motivation of this application, and is not intended to limit this application.

[0083] Figure 1 A flowchart illustrating an example of an engineering fraud auditing method based on multi-source temporal causal inference according to an embodiment of this application is shown.

[0084] Regarding the execution subject of the method in the embodiments of this application, it can be any controller or processor with computing or processing capabilities, such as the controller of an engineering digital management platform. By running programs or instructions stored in the storage medium, it addresses the technical scenario where the authenticity of visa matters is difficult to accurately identify during the implementation of engineering construction projects. It achieves auxiliary auditing and risk identification of visa events in the project to be audited by uniformly acquiring, aligning, modeling time-series correlations, identifying causal relationships, correcting domain knowledge constraints, and quantifying risks of multi-source heterogeneous business data generated during the project implementation.

[0085] In some examples, it may be integrated into an electronic device or terminal through software, hardware, or a combination of both, and the type of terminal or electronic device may be diverse.

[0086] like Figure 1As shown, in step S110, multi-source heterogeneous time-series data of the project and engineering visa events to be audited are obtained. The multi-source heterogeneous time-series data comes from multiple engineering business data sources, and the multiple engineering business data sources have different sampling frequencies.

[0087] In practice, engineering projects typically generate time-series data continuously across multiple business stages, including design management, construction execution, on-site monitoring, progress management, resource allocation, and cost settlement. For example, in a building construction or municipal engineering scenario, the design system continuously generates building information model component modification records, design change orders, and drawing review records; the construction management system generates daily construction logs, process completion status, and progress ledgers; the on-site monitoring system generates sensor data collected minute-by-minute or even second-by-second, such as temperature and humidity, rainfall, foundation pit displacement, and concrete curing status; the cost system generates records of material requisition, machinery operation, labor input, and capital expenditure; simultaneously, there may also be non-periodic business data such as hidden works acceptance images, drone inspection images, and on-site signature records. Due to significant differences in the collection mechanisms, recording granularity, and business meanings of the above data, an overall multi-source time-series data set is formed, characterized by heterogeneity, non-uniform sampling, and cross-system dispersion.

[0088] Simultaneously, the system acquires auditable engineering visa events. These events typically correspond to additional costs, extensions of the construction period, adjustments to the amount of work, or changes in processes resulting from design adjustments, changes in geological conditions, abnormal weather, adjustments to construction measures, or changes requested by the construction unit. For example, in a foundation pit project, the construction unit might submit a visa application stating that "continuous heavy rainfall has caused water accumulation in the foundation pit, requiring additional costs for dewatering measures." Similarly, in a road project, the construction unit might submit a visa application stating that "local foundation conditions are abnormal, requiring additional backfilling or reinforcement work." By unifying and initially linking auditable visa events with multi-source business data such as design, construction, monitoring, progress, and cost, subsequent analysis is no longer limited to the visa text itself but can establish a multi-dimensional evidentiary basis based on objective business records before and after the visa's formation.

[0089] In step S120, a multi-resolution dynamic time normalization algorithm is used to perform time alignment processing on multi-source heterogeneous time series data to generate a time series matrix corresponding to a unified time axis.

[0090] Because the sampling frequency and recording timing of different engineering business data sources often differ significantly, directly splicing, merging, or using fixed-interval interpolation can easily lead to the smoothing, mismatch, or misjudgment of critical business changes. For example, several hours of continuous short-duration heavy rainfall may only be reflected as "local waterlogging treatment on the day" in the daily construction log. If simply aligned by day, the true chronological relationship between the rainfall peak and subsequent drainage actions can be weakened. Another example is that a BIM component modification record may be saved as an event timestamp in the model system, while actual on-site construction adjustments and material requisitions occur several hours or days later. Without adaptive alignment, it is difficult to accurately determine the correlation sequence between design adjustments and subsequent cost changes. Therefore, this embodiment uses a multi-resolution dynamic time normalization algorithm to flexibly correct for misalignments in the time dimension of different data sources, rather than simply forcing all data to be mapped to an equally spaced sequence.

[0091] In practical implementation, a benchmark engineering business data source that can relatively stably reflect the project's progress or the audit's main focus can be selected first. This could include main construction progress data, daily construction logs, or a unified time reference centered on the occurrence of visa events. Then, coarse-to-fine time alignment processing can be performed on the remaining data sources. Specifically, the overall trend and key change segments of each data source can be identified at a coarser time granularity. Then, within local time constraints, fine-grained sampling points can be further matched to form an alignment relationship between each data source and a unified time axis. For example, in the scenario of "abnormal rainfall—foundation pit water accumulation—drainage visa," the system can reorganize rainfall curves, displacement monitoring records, drainage equipment start / stop records, construction logs, and drainage fee applications according to the same timeline. This allows records with inconsistent frequencies and different triggering methods to be jointly analyzed under a unified time slice, effectively reducing the interference of sampling differences and time mismatches on subsequent delayed correlation analysis. This provides a consistent time reference basis for subsequently constructing the temporal action chain between multiple business sources.

[0092] In step S130, the lag correlation between different engineering business data sources is calculated based on the time series matrix to construct a directed time series interaction graph, and the time series graph embedding representation of each engineering business data source in the directed time series interaction graph is extracted.

[0093] Under a unified timeline, the system further analyzes whether there are any sequential transmission relationships between different engineering business data sources. In engineering scenarios, many business changes are not immediate but have a significant lag. For example, continuous rainfall does not directly manifest as increased expenses at the same moment, but usually first manifests as water accumulation in the foundation pit, the deployment of pumping equipment, construction suspension, or adjustments to local procedures, and only then is it reflected in cost ledgers, machinery shifts, or visa applications. Similarly, after a design change record is generated in the BIM system, it may require adjustments to material plans, on-site briefings, and construction implementation before finally being reflected as adjustments to the workload or changes in the schedule. Based on this, the steps in this embodiment no longer only analyze static correlations at the same moment, but focus on analyzing whether historical changes in one engineering business data source will affect another engineering business data source in the subsequent time.

[0094] In practical implementation, the system can examine the correlation strength between historical and subsequent observations of any two engineering business data sources within a preset lag window, and determine whether a directed connection is established between them and the corresponding connection weights. A directed temporal interaction graph can be formed from multiple engineering business data sources and their directed connections, where nodes represent different business data sources, directed edges represent the direction of potential temporal influence, and edge weights represent the strength of the influence. Furthermore, since the initial interaction graph is typically high-dimensional, sparse, locally strongly coupled, and has a complex overall structure, the system can also extract temporal graph embedding representations of each node from the graph, allowing each engineering business data source to retain its own temporal evolution characteristics and its structural position in the overall interaction network in a lower-dimensional space. For example, in the scenario of "design change—component adjustment—material addition—cost change," the adjacency, hierarchical relationship, and coupling strength of related data sources in the graph structure can all be compressed and reflected in the corresponding embedding representation, transforming the complex multi-source temporal interaction relationship into a structural feature input suitable for subsequent causal inference processing.

[0095] In step S140, multi-source time-invariant causal discovery is performed in multiple pre-divided environments based on time-series graph embedding representation to determine the residual variation characteristics and parameter consistency characteristics of the regression model under different environments, and to calculate the causal influence strength between different engineering business data sources based on the residual variation characteristics and parameter consistency characteristics; each environment is used to indicate the corresponding construction status and / or data source attributes.

[0096] In engineering implementation, superficial correlation does not equate to true causation. For example, the simultaneous occurrence of rising material costs and delayed construction progress at a certain stage does not necessarily indicate a direct causal relationship between the two; it could also be due to concentrated rainfall, limited work areas, equipment scheduling delays, or other third-party factors. Therefore, this embodiment further introduces a multi-environment invariant causality discovery mechanism based on time-series interaction analysis to filter out pseudo-correlation relationships that only occasionally appear in local scenarios and retain more stable supporting relationships. Here, "environment" can be categorized not only based on construction stage, process status, work area, weather conditions, and other construction status information, but also based on the data source attributes inherent in different engineering business data sources. Specifically, data source attributes can include the business category to which the data source belongs, sampling method, update rhythm, formation level, record carrier type, and the business semantic attributes reflected. For example, sensor monitoring data, construction management records, cost settlement records, and design model change records can be regarded as business sets with different data source attributes; high-frequency automatic sampling data, low-frequency manual data, and event-triggered record data can also be regarded as data sets with different attribute categories.

[0097] It should be noted that incorporating data source attributes into the environmental classification is because data sources with different attributes often correspond to different information formation mechanisms and noise patterns in engineering scenarios. For example, on-site sensor data typically directly reflects changes in physical state and has strong temporal continuity, but is easily affected by factors such as equipment accuracy and environmental disturbances. Construction logs, progress reports, or signature records are more inclined towards manual management, with a slower update pace, and may be affected by reporting habits, management granularity, or approval delays. BIM component change records and design modification records, on the other hand, reflect more of the logical business adjustment process. For the same visa event, if a certain business relationship is prominent only under one type of data source attribute, but lacks stability in other attribute categories, this relationship is more likely caused by differences in recording mechanisms, sampling methods, or data formation biases, rather than constituting a true causal driving relationship. Conversely, if a candidate relationship remains relatively consistent under different construction states and can still be stably established under different data source attribute conditions, it better demonstrates the robustness of this relationship across scenarios and sources, making it more suitable as a key supporting relationship in visa audits.

[0098] In practical implementation, the system can classify samples into environments based on both the construction status dimension and the data source attribute dimension, or combine the two to form composite environment units. For example, the foundation construction stage, the main structure stage, and the decoration and installation stage can be considered as different construction status environments; simultaneously, sample sets dominated by physical monitoring data, management record data, and design change data can be considered as different data source attribute environments. Subsequently, based on time series graph embedding representation, the system selects candidate input variables around the target engineering business data source and establishes corresponding local regression relationships in each environment. By comparing the changes in the model fitting residuals before and after the introduction of candidate variables, it can be determined whether the explanatory power of the model for the target engineering business data source has been substantially improved; furthermore, by combining whether the model parameters remain relatively consistent in different environments, it can be determined whether the relationship has cross-environment stability. Finally, the system uses the degree of residual improvement and parameter consistency as the criteria for determining causality and calculates the causal influence strength between different engineering business data sources. Therefore, candidate relationships can be constrained from both the perspectives of changes in the construction process and differences in data formation mechanisms, thereby reducing misjudgments caused by fluctuations in a single working condition or deviations in a single data type, and improving the stability and reliability of causal identification results in engineering visa audits.

[0099] In step S150, a target causal network corresponding to the engineering visa event is constructed based on the causal influence strength. This target causal network is then optimized using pre-defined engineering domain knowledge constraints to adjust the weights of the corresponding causal edges. The engineering domain knowledge constraints include at least one of the following: engineering design specifications, contractual clause constraints, and historical experience models.

[0100] After obtaining the causal influence strength between different engineering business data sources, the system further constructs a target causal network around the engineering visa events to be audited. This network describes the supporting relationship structure that starts from the underlying engineering business data source, is gradually transmitted through intermediate business states, and ultimately affects the visa events. In other words, this embodiment does not only focus on the direct relationship between a single business factor and visa matters, but also unifies the transmission relationships at multiple levels such as design, construction, monitoring, schedule, resources, and costs into the same network. For example, in the scenario of "design modifications leading to adjustments in local components, which further cause material additions and cost changes," the target causal network can include design change records, BIM component status changes, material requisition records, cost fluctuations, and visa events in the same supporting chain; as another example, in the scenario of "abnormal rainfall causing water accumulation in the foundation pit, which further causes drainage input and schedule impact," the target causal network can associate meteorological data, on-site monitoring data, equipment input records, construction progress anomalies, and visa matters.

[0101] On the other hand, engineering visa audits are not purely data correlation issues; they are also constrained by engineering specifications, contract boundaries, and rules of experience. Therefore, this embodiment further introduces engineering domain knowledge constraints to modify and optimize the causal edges in the target causal network. Engineering domain knowledge constraints can originate from at least one of engineering design specifications, contractual clause constraints, and historical experience models. These constraints reflect which business evolution relationships are logically permissible in engineering, which cost changes should not result in additional requests within contract boundaries, and which combinations of business states are considered high-risk or unreasonable in historical projects. For example, if data analysis shows a strong correlation between the cost of a certain measure and a construction adjustment, but the contract stipulates that such measures are already included in a fixed-price lump sum, the system can identify a business rule conflict and weaken the weight of the corresponding causal edge. Similarly, if an engineering measure is inconsistent with the structural or procedural constraints specified in the design specifications, the system can also modify the corresponding connection. This allows the target causal network to retain its data-driven discovery capabilities while further conforming to the business rules and review logic in engineering audit scenarios, reducing the misleading impact of unreasonable causal chains on subsequent risk identification.

[0102] In step S160, the audit causal score and causal completeness index of the engineering visa event are calculated based on the target causal network after weight adjustment, and the presence of false risks in the engineering visa event is identified based on the audit causal score and causal completeness index.

[0103] Here, the system performs a two-dimensional quantitative evaluation of audit engineering visa events based on the target causal network optimized by domain knowledge constraints. On the one hand, the system extracts the key supporting paths leading to the visa event and combines the influence strength of each causal edge on the path with the importance of the engineering business to form an audit causal score that characterizes the strength of the preceding support. On the other hand, based on the dependencies depicted by the current target causal network, the system performs forward inference on the status of multiple engineering business data sources related to the visa event and compares the inference results with the actual observation results to form a causal completeness index that characterizes the overall logical consistency. The former focuses more on "whether the visa has sufficient objective cause support," while the latter focuses more on "whether the data chain related to the visa is generally self-consistent."

[0104] For example, in the scenario of "additional drainage fee visa," if the system extracts a relatively complete key supporting path from rainfall, water accumulation monitoring, equipment investment, construction suspension to visa application, the audit causality score is likely to be high. If the observation results from various relevant data sources are also relatively consistent with the results of the target causal network deduction, the causal completeness index is also high, indicating that the visa is more likely to be a normal visa with an actual engineering background. Conversely, if a visa document shows a large increase in engineering volume, but the preceding business chain lacks corresponding design adjustments, changes in site conditions, or evidence of resource investment, its audit causality score will be low. If there are further significant inconsistencies between site monitoring, progress records, material ledgers, and the visa document, the causal completeness index will also be low, indicating a high risk of fraud in the visa matter. Thus, the system can transform visa auditing from a traditional method relying on manual experience and fragmented data verification into an auxiliary auditing method based on a joint assessment of causal support chains and global data consistency, thereby improving the objectivity, stability, and interpretability of engineering visa risk identification.

[0105] Regarding the implementation details of generating the time series matrix through time alignment processing in step S120, in some examples of embodiments of this application, firstly, the time series of different engineering business data sources and their corresponding timestamp sequences in multi-source heterogeneous time series data are obtained.

[0106] It should be noted that due to differences in the operation of data acquisition equipment, management platforms, and business systems at engineering sites, different engineering business data sources not only have different sampling frequencies, but also typically inconsistent start and end times and timeline density. For example, on-site temperature and humidity, displacement, or stress sensor data can be represented as high-frequency, equidistant sequences at the minute or second level, while construction progress records, manual logs, and approval records may be represented as low-frequency, non-equidistant sequences updated daily or filled out irregularly; building information model component change records and concealed works acceptance images may also present event-triggered, discrete recording methods.

[0107] To facilitate subsequent unified time alignment processing, any two engineering business data sources in the multi-source heterogeneous time series data can be represented as corresponding time series and timestamp sequences, respectively. For example, the time series of the first engineering business data source can be determined. and its corresponding timestamp sequence and the time series of the second engineering business data source and its corresponding timestamp sequence ;in, and These represent the total number of sampling points for the corresponding time series, and typically... Using the above representation method, data from different sources and with different sampling densities can be uniformly transformed into time-series objects that can participate in subsequent distance calculations and path searches.

[0108] Then, a weighted distance measure function with a time difference penalty term is constructed. The weighted distance measure function is used to measure the difference between sequence sample points by multiplying the norm distance of the sample feature difference vector with the exponential weight function based on time penalty. The exponential weight function uses a preset penalty hyperparameter to control the sensitivity of path alignment to time difference.

[0109] Here, the sample feature difference vector has been pre-processed to be locally dimensionless. In some examples, arbitrary sample points... and Weighted distance between The calculation formula is:

[0110] ;

[0111] In the formula, and These are the sampling point indices for the corresponding time series. and These are the sample feature values ​​under the corresponding timestamps. To ensure the subtractability between different physical quantities, the sample feature values ​​have undergone local dimensionless normalization before participating in the distance calculation. The squared L2 norm of the feature difference vector is represented by... The square of the time difference between the sample point pairs. For the time-penalized exponential weighting function, This is a penalty hyperparameter used to control the sensitivity of path alignment to time difference.

[0112] The above formula considers both numerical and temporal similarity. The first part characterizes the difference between two sample points in the feature space, while the second part imposes a higher cost on sample point pairs with larger time spans. When the time offset between two sample points is large, the exponential term will rapidly increase with the square of the time difference, thereby increasing the cost of their participation in optimal path matching. This avoids mispairing engineering events that are far apart in time due to local numerical similarity, and makes the obtained temporal matching relationship more consistent with the actual occurrence sequence and transmission law of engineering activities.

[0113] Subsequently, downsampling processing is performed on the time series of different engineering business data sources to obtain the corresponding coarse-grained time series, and the global cumulative distance between the coarse-grained time series is calculated based on the weighted distance metric function. The coarse-grained alignment path that minimizes the global cumulative distance is then extracted.

[0114] In practical implementation, if we directly target the lengths respectively and Performing global matching on the raw time series often leads to high computational complexity and storage overhead. This is especially true in scenarios with long project cycles and high sensor data sampling frequencies, where direct global dynamic matching in the raw space significantly increases processing costs. Therefore, it is advisable to first perform smoothing and downsampling on the raw time series to obtain a shorter, coarse-grained time series that still retains the macroscopic trend of change. and Based on this, the system can construct a cumulative distance matrix. The optimal path in the coarse-grained space is calculated step by step using dynamic programming.

[0115] For example, arbitrary coordinates The cumulative distance at a point can be calculated using the following recursive formula: ;in, and These are coarse-grained time series. and The sampling point index, This represents the local weighted distance between corresponding coarse-grained sampling points. This recursive relationship indicates that the optimal cumulative distance for the current state is determined by the local distance at the current position and the minimum historical cost among adjacent valid preceding states. Furthermore, the initial boundary values ​​of the cumulative distance matrix are preset according to dynamic programming boundary conditions (e.g., the starting position is set to zero, and the remaining boundaries are set to infinity). After traversal, the coarse-grained alignment path that minimizes the global cumulative distance can be extracted by backtracking from the end point of the cumulative distance matrix to the starting point. Although it is not yet precise to the level of the original sampling points, it can already reflect the overall correspondence between different engineering business data sources on a macroscopic time scale.

[0116] Subsequently, the coarse-grained aligned path is mapped back to the original sampling resolution, and a local search constraint boundary with a preset time radius is set with the mapped path as the center.

[0117] In practical implementation, the coarse-grained alignment path can be reverse-expanded or interpolated based on the original timestamp relationship to form an approximate time correspondence main line in the original sampling space. Then, a local search constraint boundary is constructed centered on this main line, combined with a preset time radius. The function of this local search constraint boundary is to compress the original high-dimensional search space into a finite band-shaped region surrounding the coarse-grained path, thereby concentrating subsequent fine-grained matching within a local area where true correspondences are more likely to form. For example, when the coarse-grained path indicates that changes in construction progress at a certain stage mainly correspond to on-site monitoring data for a specific time period, then only a finer correspondence of sampling points needs to be searched around that time period in the original sampling space, without needing to perform an indiscriminate traversal of the entire global space. Thus, the macroscopic time trend reflected by the coarse-grained path is preserved while reducing the redundant computational load of the fine-grained search.

[0118] Then, within the local search constraint boundary, the weighted distance metric function is called to perform fine-resolution dynamic time-normalized matching iterations to obtain the corresponding fine-grained time-aligned index mapping relationship.

[0119] Within the local search constraint boundary, the system again invokes the aforementioned weighted distance metric function and uses the dynamic programming path search method to perform finer time matching on the original sampling points. Since the search range is now limited by the coarse-grained path, computational resources can be more concentrated on local detail alignment, thereby improving the ability to identify short-term changes and local emergencies. In engineering scenarios, this fine-grained matching can more accurately capture correspondences such as between short-term heavy rainfall and on-site water accumulation records, between instantaneous equipment downtime and stress fluctuation records, or between local construction adjustments and abnormal monitoring values. After this iterative process, the system can output a fine-grained time alignment index mapping relationship accurate to the original sampling point level. This mapping relationship can be in the form of a one-to-one, one-to-many, or many-to-one correspondence, used to directly characterize the time binding method of different engineering business data sources at the original sampling level.

[0120] Furthermore, using the timestamp sequence of the preset benchmark engineering business data source as a unified time axis reference, and based on the fine-grained time alignment index mapping relationship, sampling point alignment processing is performed on each of the other engineering business data sources to generate a time series matrix corresponding to the unified time axis.

[0121] In order to form a standardized input data structure that can be directly used for subsequent analysis, the system may select an engineering business data source with strong temporal continuity, relatively complete coverage, or close correlation with the visa event to be audited as the unified timeline reference. For example, in some implementation modes, a comprehensive construction log or a master progress account can be selected as the benchmark of the unified timeline. For other engineering business data sources, the system reorganizes the observation values according to the fine-grained time alignment index mapping relationship. When multiple original sampling points correspond to the same reference time step, aggregation by means of mean value, extreme value or other suitable methods can be performed on the corresponding features; when there is no effectively matched sampling point for a certain reference time step, reasonable complementation or interpolation can be performed based on adjacent matching results. After the above processing, the original engineering business data sources with different sampling frequencies and different time expression modes can be uniformly organized into a time series matrix where rows correspond to unified time steps and columns correspond to different business features.

[0122] Thereby, under the premise of not destroying the sequence constraint relationship of engineering events, unified reorganization is performed on the multi-source heterogeneous data with inconsistent original time expression, so that different engineering business data can be compared and subjected to linkage analysis under the same time reference. The time series matrix obtained in this way can well retain the main time series features of the original business changes, and reduce the time mismatch problem caused by sampling differences, recording delays or different event triggering modes, thereby improving the accuracy and availability of the unified expression of multi-source engineering data.

[0123] Figure 2 Shows an exemplary operation flow chart of constructing and extracting the time series graph embedding representation of the directed time-series interaction graph corresponding to engineering business data sources in the method according to the embodiment of the present application.

[0124] As shown in Figure 2 , in step S210, zero-mean standardization processing is performed on the time series of each engineering business data source in the time series matrix to eliminate dimensional differences and baseline drift between different engineering business data sources.

[0125] In engineering business systems, the physical dimensions and numerical ranges corresponding to different data sources often differ significantly. For example, the readings of on-site temperature sensors may range from -10 degrees Celsius to 40 degrees Celsius, while the cost of machinery shifts, material expenditures, or engineering quantity statistics may reach tens of thousands or even hundreds of thousands of yuan. If these absolute values ​​are directly used for time series correlation calculations, sequences with larger numerical ranges will have a dominant influence on the calculation results, making it difficult to effectively identify data with smaller numerical ranges but obvious fluctuations. Therefore, this embodiment performs zero-mean standardization on each column of the time series matrix, i.e., the time series corresponding to each engineering business data source. For example, for the original observation sequence of a certain engineering business data source, a standardized sequence with a mean of zero and a variance of one can be obtained by subtracting the historical mean of the sequence and dividing by its standard deviation.

[0126] This processing maps different engineering business data sources to a dimensionless standard scale. Subsequent calculations no longer focus on the original numerical values, but rather on the fluctuation range, trends, and relative fluctuations of different data sources over time. This reduces dimensional interference between different business dimensions and mitigates the baseline impact caused by long-term stable offsets, enabling subsequent time-series correlation analysis across business data sources to be conducted on a unified scale.

[0127] In step S220, the maximum time-series lag window length and the lag decay coefficient that decreases over time are set. For any two engineering business data sources after standardization, the cumulative value of the cross sequence feature product is calculated within the lag window around the historical observation value of the first engineering business data source and the current observation value of the second engineering business data source, and the absolute value is taken. Combined with the lag decay coefficient, the lag influence score in the corresponding direction is calculated.

[0128] In the actual business transmission process of engineering projects, business changes are usually not immediate but have a certain degree of delayed propagation. For example, a drawing change order may gradually cause changes in material requisition records, machinery scheduling, and cost fluctuations over several days after its issuance; continuous rainfall may also gradually be reflected in on-site drainage records, progress anomaly records, or visa application records over subsequent time steps. Meanwhile, this impact usually weakens over time. Based on these characteristics, this embodiment sets a maximum time lag window length for different engineering business data sources under a unified time axis. and the corresponding hysteresis decay coefficient: ;in, This represents the discrete lag step count. For any two standardized engineering business data sources... and From engineering business data sources Pointing to engineering business data source Lagged effect score It can be calculated using the following formula: ;in, This represents the total length of the aligned time series in the time series matrix. For time indexing, and These represent the standardized engineering business data sources. and engineering business data sources The observed characteristic value at the corresponding time, This indicates the absolute value operation.

[0129] In the above formula, the inner summation term is used to statistically analyze the engineering business data source at a certain lag step. Historical changes and engineering business data sources The overall coupling between current changes is considered, and irrelevant random noise will cancel each other out during the accumulation process. The outer layer uses an attenuation coefficient to weight the influence of different lag steps, giving higher weight to more recent historical states and lower weight to more distant historical states. This effectively reflects the gradual attenuation of engineering events over time. Furthermore, the absolute value of the accumulated result of the cross-sequence product is taken, so regardless of whether the two changes in the same or opposite direction, as long as their temporal linkage is strong, it can be reflected in the lag influence score.

[0130] In step S230, a preset significant noise truncation threshold is set, and asymmetric sparse filtering based on bidirectional comparison is performed on the lag effect score: when the positive lag effect score is strictly greater than the negative lag effect score and greater than the significant noise truncation threshold, the positive lag effect score is confirmed as an effective lag association and assigned as a directed edge weight; otherwise, the corresponding directed edge weight is reset to zero to construct an asymmetric adjacency matrix of a directed temporal interaction graph with sparse noise reduction characteristics.

[0131] In actual engineering data, if edges are connected solely based on the aforementioned lag effect scores, a relatively dense relation matrix is ​​typically obtained. This matrix contains both genuine time-series driven relationships and numerous weak associations arising from shared background, seasonal variations, synchronized management operations, or accidental co-occurrence. Therefore, this embodiment further sets a significant noise truncation threshold. And compare the scores of the lag effects in both positive and negative directions. For example, when the following conditions are met: ;

[0132] Will Confirmed as a data source from engineering business. Pointing to engineering business data source Effective lag associations are assigned as directed edge weights. Otherwise, the directed edge weight will be... Set to zero. After this processing, the asymmetric adjacency matrix corresponding to the directed temporal interaction graph can be constructed. .

[0133] Through the above filtering process, noisy relationships are eliminated from both the intensity and direction of influence. Only when the lag effect score in a certain direction is significantly stronger than that in the opposite direction and exceeds a preset threshold is the directional relationship retained. This reduces redundant connections caused by bidirectional co-occurring changes, weak synchronous fluctuations, or background noise, making the resulting directed time-series interaction graph more prominent in terms of the dominant time-series driving relationships between different engineering business data sources.

[0134] In step S240, the symmetric fusion matrix of the asymmetric adjacency matrix is ​​extracted, and the degree distribution of each network node is calculated based on the symmetric fusion matrix to generate the degree matrix. Then, the degree matrix and the symmetric fusion matrix are combined to construct an undirected Laplacian matrix for graph analysis.

[0135] It should be noted that since the asymmetric adjacency matrix directly corresponds to the directed graph structure, and in subsequent spectral analysis, directly performing eigenvalue decomposition on the asymmetric matrix often introduces complex eigenvalues ​​or numerical instability problems, this embodiment first performs symmetric fusion processing on the asymmetric adjacency matrix. For example, the asymmetric adjacency matrix can be extracted. Symmetric fusion matrix: ;

[0136] This matrix is ​​used to comprehensively represent the overall connection strength of the bidirectional relationship between any two engineering business data sources. Subsequently, the system calculates the degree distribution of each network node based on this symmetric fusion matrix and generates a degree matrix accordingly. Furthermore, the undirected Laplace matrix can be constructed in the following form: ;

[0137] Through the above processing, the adjacency matrix, which originally focused on local edge weight relationships, is further transformed into a Laplacian matrix that can characterize the overall network structure. This matrix contains the connection density, clustering trend, and overall topological distribution information between nodes, thus providing a stable mathematical foundation for subsequent spectral decomposition and extraction of low-dimensional node representations.

[0138] In step S250, an eigenvalue decomposition operation is performed on the undirected Laplacian matrix to extract the eigenvectors corresponding to the first few smallest non-zero eigenvalues ​​and concatenate them to form a dimension-reduced feature matrix. The row vectors in the dimension-reduced feature matrix corresponding to each engineering business data source are used as the time series graph embedding representation of each engineering business data source.

[0139] Graph networks typically have high dimensionality and sparse structure in their original representation space. Directly using adjacency relationships for subsequent analysis can easily lead to high computational burden. Therefore, this embodiment focuses on the undirected Laplacian matrix. Perform feature decomposition to extract features corresponding to the previous features. The eigenvectors with the smallest non-zero eigenvalues ​​are extracted, and the extracted eigenvectors are concatenated column-wise to form a dimensionality-reduced feature matrix. Among them, The preset time series graph embedding dimension is then used. Subsequently, the row vectors in this dimensionality-reduced feature matrix corresponding to each engineering business data source are used as the time series graph embedding representation of each engineering business data source.

[0140] After this processing, each engineering business data source no longer exists as a single time series, but is represented as a low-dimensional dense vector. This vector retains the connectivity characteristics of the engineering business data source within the overall network, and also reflects its relative position and clustering attributes within the entire engineering business interaction structure. Thus, the complex, high-dimensional, and sparse temporal interaction network is compressed into a more compact and computationally efficient low-dimensional representation, while preserving as much of the original network's main structural information as possible. This improves the accuracy and stability of the unified representation of cross-business data source relationships.

[0141] In this embodiment, by sequentially performing standardization, lag effect score calculation, bidirectional comparison filtering, graph structure transformation, and spectral decomposition dimensionality reduction on multi-source engineering business data, the relatively scattered local linkages between the original time series can be organized into a computable and comparable low-dimensional graph embedding representation. This processing result can better reflect the relative relationships, connection strengths, and structural distribution of different engineering business data sources in the overall time-series interaction network, while reducing the complexity and noise interference caused by the direct participation of the original high-dimensional graph structure in the calculation.

[0142] Regarding the implementation details of feature extraction and calculation of causal influence strength in step S140, in some examples of embodiments of this application, firstly, based on the time series graph embedding representation of each engineering business data source, the embedding vector similarity between the target engineering business data source and the other engineering business data sources is calculated, and the engineering business data sources with embedding vector similarity greater than a preset similarity threshold are extracted as a set of candidate parent variables for constructing a regression model for the target engineering business data source, so as to achieve dimensionality reduction of the causal inference search space.

[0143] In practice, engineering construction projects often involve a large number of engineering business data sources, including high-frequency monitoring data as well as business records such as construction management, progress reporting, material consumption, expense expenditure, and design changes. Performing a full causal search across all engineering business data sources would significantly increase the number of combinations and computational overhead. Since the aforementioned time-series graph embedding representation has compressed the structural positional relationships of each engineering business data source in the overall time-series interaction network into low-dimensional vectors, the similarity between the low-dimensional vectors corresponding to different engineering business data sources can be used to pre-exclude marginal variables with weak relationships to the target engineering business data source. For example, cosine similarity can be used to calculate the embedding vector similarity between the target engineering business data source and the embedding vectors corresponding to the other engineering business data sources. Engineering business data sources with embedding vector similarities greater than a preset similarity threshold are selected as data sources with strong correlation to the target engineering business data source in the embedding space. These selected data sources are then extracted as a set of candidate parent variables for constructing a regression model for the target engineering business data source.

[0144] For example, when using a specific visa expense as the target engineering business data source, the system can prioritize retaining engineering business data sources that are closer to it in the embedding space, such as rainfall monitoring records, drainage equipment start-up and shutdown records, construction log anomalies, and material requisition changes, while excluding marginal business variables that are clearly less relevant to the visa item. Through this process, the scope of regression modeling can be narrowed while maintaining highly relevant candidate relationships, thereby reducing the variable dimensionality and computational burden in multi-environment analysis and mitigating the interference of a large number of irrelevant variables on the statistical results.

[0145] Then, for each pre-divided environment in the set of multiple pre-divided environments, for any data source of the business to be tested in the candidate parent variable set, a baseline regression model without the data source of the business to be tested and a conditional regression model with the data source of the business to be tested are constructed respectively; wherein, the conditional regression model in each environment adopts a unified parameterized model structure.

[0146] In this embodiment, the pre-divided environment can be formed by dividing the construction state, dividing it by data source attributes, or dividing it by a combination of construction state and data source attributes. For example, the foundation construction stage, the main structure construction stage, and the decoration and installation stage can be regarded as different construction state environments. Alternatively, sample sets mainly composed of physical monitoring data, sample sets mainly composed of construction management record data, and sample sets mainly composed of design change and model record data can be regarded as different data source attribute environments. By observing the interaction between candidate variables and target variables in different environments, it can be determined whether a certain relationship is only valid in individual scenarios or remains relatively stable under multiple environmental conditions.

[0147] For example, for a set of environments consisting of multiple pre-divided environments. Each pre-partitioned environment For the set of candidate parent variables Any business data source of the project under test Construct separate data sources that do not contain the business data of the project under test. The baseline regression model and the data source containing the engineering business data to be tested Conditional regression model: ; ;

[0148] In the formula, Data source for target engineering business At any moment The observed values, For the candidate parent variable set at time The input feature vector, This indicates removing the data source of the engineering business to be tested from the input feature vector. The subsequent feature vectors, and respectively environment Baseline local regression function and conditional local regression function, and These represent the baseline fit residuals and conditional fit residuals, respectively; the conditional local regression functions under different environments employ a unified parameterized model structure. Using a unified parameterized model structure helps ensure that the model parameter vectors obtained under different environments maintain consistency in dimensionality and meaning, thus facilitating subsequent cross-environment parameter comparisons.

[0149] Then, by fitting all environments The baseline regression and conditional regression models are used to pool the baseline fit residuals and conditional fit residuals across all environments, in order to calculate the variance of the baseline residuals across environments. and conditional residual variance The residual variation characteristics were determined based on the relative reduction ratio of baseline residual variance to conditional residual variance.

[0150] In practical implementation, the system can aggregate the baseline fitting residuals and conditional fitting residuals obtained under various environments in environmental order to form a complete residual sequence across environments, and calculate the variances of the baseline residuals and conditional residuals accordingly. The relative reduction ratio between the two can be expressed as: This ratio is used to characterize the data source of the project's business. After being introduced, the target project business data source The degree of reduction resulting from the uncertainty of forecasting. If this relative reduction rate is high, it indicates that the data source of the project under test is... The addition of this feature can significantly improve the model's accuracy in targeting engineering business data sources. The explanatory or predictive power of the data source; conversely, if the ratio is low, it indicates that the data source of the project under test is... The information gain provided is limited, and it has limited impact on the target engineering business data source. Its contribution was relatively weak.

[0151] Subsequently, the conditional regression model was extracted for various environments. The model parameter vector Furthermore, the parameter consistency characteristics are determined by calculating the vector space distance between the model parameter vectors in other environments and the model parameter vectors in a preset reference environment.

[0152] In this embodiment, to determine whether a candidate relationship possesses cross-environment stability, the system further compares the parameter vectors of the conditional regression model under different environments. For example, a representative reference environment can be selected. For example, environments with high data quality, minimal disturbance, or relatively stable construction conditions, and calculating model parameter vectors under other environments. With reference environment parameter vector Euclidean distance between them: .

[0153] If this distance is small in multiple environments, it indicates that the data source of the project under test is business data. Data source for target engineering business The interaction between parameters shows relatively small changes across different environments, and its parameters exhibit relatively consistent behavior. However, if the parameter distance fluctuates significantly across different environments, it indicates that the relationship is easily affected by environmental changes and has weak stability. Through the above processing, we can distinguish whether a relationship holds true only by chance in a local environment, or whether it remains relatively consistent across multiple environments.

[0154] Furthermore, the residual variation characteristics are multiplied by the average satisfaction of the cross-environment parameter consistency characteristics after removing the reference environment, in order to calculate and quantify the causal influence strength from the data source of the engineering business under test to the data source of the target engineering business.

[0155] For example, based on the residual variation characteristics and the Euclidean distance between model parameter vectors in different environments, the data source of the engineering project under test is calculated using the following formula. To the target engineering business data source causal influence strength : In the formula, The total number of pre-divided environments. This is a pre-defined reference environment used as a comparison benchmark. For the conditional regression model in the reference environment The model parameter vector below, This represents the L2 norm operator for calculating the Euclidean distance between vectors. To prevent smooth, extremely small positive numbers with a denominator of zero; This is a parameter consistency indicator function, when the environment... Compared with the reference environment The Euclidean distance between the model parameter vectors is strictly less than the preset consistency tolerance. When the parameter consistency indicator function is active, its value is 1; otherwise, it is 0.

[0156] In the above formula, the first term reflects the data source of the business of the project under test. Data source for target engineering business The first term represents the predicted gain, and the second term reflects the average proportion of the interaction relationship that meets stability requirements across multiple environments. Multiplying the two terms reveals the result only when the data source of the engineering project under test is available. It can significantly reduce the data source of the target engineering business A higher causal influence strength is obtained only when the predicted residual variance is both accurate and maintains high parameter consistency across different environments. Therefore, a unified quantitative framework can simultaneously reflect both relationship strength and environmental stability.

[0157] In this embodiment, by screening the candidate parent variable set, constructing baseline regression and conditional regression models under multiple environments, statistically analyzing cross-environment residual variation characteristics, and comparing cross-environment parameter consistency characteristics, the original temporal interaction relationship can be further transformed into a causal influence strength with environmental stability constraints. This processing result can better distinguish between stable interactions that remain consistent under multiple engineering conditions and accidental associations that only appear in local environments, thereby improving the reliability of multi-source engineering business relationship identification.

[0158] Figure 3 A flowchart illustrating an example of optimizing a target causal network in accordance with the method of an embodiment of this application, incorporating pre-set engineering domain knowledge constraints.

[0159] like Figure 3 As shown, in step S310, at least one of the engineering design specifications, contract terms and historical experience models is parsed and abstracted into entity relation triples to construct an engineering domain knowledge graph, and the engineering domain knowledge graph is transformed into a prior knowledge matrix with constraint strength labels.

[0160] In practice, project implementation and visa approval are typically constrained by multiple rules, including technical requirements in design specifications regarding structural layout, construction conditions, material usage, and process dependencies; commercial agreements in contract terms regarding cost attribution, lump-sum scope, and adjustable boundaries; and empirical review rules developed from past projects. These rules often exist in the form of natural language text, tabular agreements, or empirical descriptions, making them difficult to directly participate in numerical calculations. Therefore, in this embodiment, at least one of the aforementioned engineering domain knowledge can be extracted in a structured manner. For example, rule parsing, keyword extraction, entity recognition, syntactic analysis, or manually pre-set templates can be used to transform relevant knowledge into entity relation triples consisting of a head entity, relation type, and tail entity. For instance, business constraints such as "dewatering measures fees belong to the lump-sum package," "specific geological conditions correspond to specific foundation treatment requirements," and "adjustments to a certain type of process may cause changes in specific resource inputs" can be transformed into structured representations.

[0161] After obtaining the entity relation triples, a knowledge graph for the engineering domain can be further constructed, and relations in the graph can be assigned different strength labels based on relation type, rule source, and constraint level. Subsequently, the system transforms the engineering domain knowledge graph into a prior knowledge matrix. In this matrix, the elements not only represent whether constraints exist between different business entities, but also characterize the strength of those constraints. Generally, constraints derived from mandatory design specifications or explicit contractual agreements can be set to a higher strength, while constraints derived from historical experience models or general rules of experience can be set to a relatively lower strength. Through this process, engineering expertise, originally existing in textual or experiential form, can be transformed into addressable and computable structured prior information, providing a computational basis for subsequent business consistency verification of target causal edges.

[0162] In step S320, for any target causal edge in the target causal network, the data source business attributes of its head node and tail node are mapped to the corresponding coordinate positions of the prior knowledge matrix in order to extract the domain logic matching degree coefficient corresponding to the target causal edge.

[0163] In this embodiment, the target causal network formed by the preceding data-driven analysis typically contains multiple target causal edges pointing from the first engineering business data source to the second engineering business data source. For any one of these target causal edges, the data source business attributes corresponding to its head node and tail node can be obtained. For example, the head node corresponds to business attributes such as abnormal rainfall monitoring, groundwater level changes, increased material consumption, and schedule deviations, while the tail node corresponds to additional drainage costs, basic treatment visas, project extensions, or other visa-related statuses. The system maps the data source business attributes of the aforementioned head and tail nodes to a priori knowledge matrix. The corresponding coordinates are used to extract the domain logic matching coefficient corresponding to the target causal edge. .

[0164] Here, the domain logic matching coefficient This coefficient is used to characterize the degree of matching between the target causal edge identified by data-driven approaches and the engineering business knowledge system. In practice, this coefficient is usually normalized to ensure that its value satisfies... If a target causal edge is consistent with common business logic in design specifications, contractual constraints, or historical experience rules, its corresponding matching coefficient can be high; if a target causal edge lacks support from common engineering sense or is rarely found in existing knowledge, its matching coefficient can be low. By introducing this matching coefficient, subsequent verification of target causal edges can no longer rely solely on statistical data results, but can simultaneously refer to business rationality information in the engineering field.

[0165] In step S330, rule conflict detection is performed on the target causal edge based on engineering domain knowledge constraints: when the target causal edge indicates the occurrence of additional cost causal derivation, and the engineering sub-items corresponding to the head node and tail node are identified as non-additional cost items in the contract terms constraints, or when the evolution trend of the target causal edge violates the topological constraint relationship of the engineering design specifications, a domain logic conflict is determined to be triggered.

[0166] In practical implementation, rule conflict detection is mainly used to identify whether there are obvious unreasonable situations at the business level regarding target causal edges. The first type of detectable conflict is contract boundary conflict; for example, a target causal edge indicates that a change in a preceding business will cause additional expenses, but after checking the contract terms, it is found that the corresponding engineering sub-item has been agreed to be a lump-sum item or a non-additional expense item. This indicates that although the target causal edge has a certain correlation in the data, its business result does not conform to the contract agreement, and should be judged as having a domain logic conflict. The second type of detectable conflict is design specification or process logic conflict; for example, if the business evolution direction reflected by the target causal edge violates the topological constraints defined by the engineering design specifications, or clearly does not conform to the usual dependency sequence of engineering activities, then it is also judged as having a domain logic conflict.

[0167] For example, in an engineering scenario, if the system identifies a strong correlation between "local measure adjustment" and "addition of a fixed-price sub-item cost" from the data, but the contract explicitly states that the sub-item cost cannot be separately calculated through visas, then this target causal edge should be identified as a contract boundary conflict. Furthermore, if a target causal edge shows that a change in the state of an upstream component is directly driven in reverse by a downstream decoration process, and this relationship is inconsistent with the structural dependency sequence reflected in the design specifications, then it should be identified as a design topology conflict. Therefore, rule-level business reviews can be performed on target causal edges in the target causal network to identify connections inconsistent with engineering professional logic.

[0168] In step S340, when a domain logic conflict is triggered, a penalty attenuation factor greater than zero is determined based on the domain logic matching degree coefficient, and the penalty attenuation factor decreases as the domain logic matching degree coefficient increases; when no domain logic conflict is triggered, the penalty attenuation factor is set to zero.

[0169] In order to enable the rule conflict detection results to further influence the adjustment of the target causal edge weights, this embodiment introduces a conflict indicator variable in a preferred example. The value is 1 when a domain logic conflict is triggered, and 0 when no domain logic conflict is triggered. This conflict indicator variable and the aforementioned domain logic matching coefficient are used as the basis for this calculation. The penalty decay factor for the corresponding target causal edge can be calculated. For example, it can be in the following form: ;in, These are preset smoothing control parameters, and .

[0170] As can be seen from the above expression, when no domain logic conflict is triggered, due to Then the penalty decay factor The value is directly zero, without any additional intervention in the strength of the original causal influence; when a domain logic conflict is triggered, due to... The penalty decay factor is determined by the domain logic matching coefficient. Decision. If a causal edge of a target triggers a conflict, but its matching degree within the domain knowledge system is still relatively high, then... The smaller the value of a conflicting edge, the smaller the corresponding penalty attenuation factor; conversely, if a target causal edge triggers a conflict and has a low matching degree, the penalty attenuation factor is relatively large. This approach allows the impact of rule conflicts to act on the edge weight correction process as a continuous quantity, rather than completely eliminating all conflicting edges.

[0171] In step S350, based on the original causal influence strength of the target causal edge and the determined penalty attenuation factor, a nonlinear soft constraint modulation function based on prior knowledge is invoked to transform the penalty attenuation factor into a nonlinear attenuation effect on the original causal influence strength, so as to calculate the causal influence strength after weight adjustment and use it to reconstruct the target causal network.

[0172] For example, based on the business data source of the project under test To the target engineering business data source Original causal influence strength and the determined penalty decay factor The adjusted causal influence strength is calculated using the following modulation model. : In the formula, For nonlinear soft-constraint modulation functions based on prior knowledge, These are global control hyperparameters used to adjust the severity of penalties, and ; Causal influence strength after weight adjustment Used to replace the corresponding original causal influence strength to reconstruct the target causal network.

[0173] In the above formula, when Sometimes, At this point, the strength of the causal influence after weight adjustment remains consistent with the strength of the original causal influence; when At this time, the exponential decay term will take values ​​between zero and one, thus exerting a continuous decay effect on the strength of the original causal influence. Furthermore, as... As the number of parameters increases, the exponential decay intensifies, further weakening the weight of the corresponding target causal edge. The settings enable the system to make overall adjustments based on the degree of influence of application scenarios on domain knowledge constraints. Through the above processing, the original target causal network obtained from data-driven analysis can be modified into an optimized network that takes into account both data relationships and engineering business constraints. In this network, the weights of target causal edges that clearly conflict with engineering domain knowledge are reduced, while the weights of target causal edges that are consistent with engineering rules retain more of their original influence strength.

[0174] In this embodiment, by transforming at least one of engineering design specifications, contractual constraints, and historical experience models into structured prior knowledge and mapping it to the target causal edges in the target causal network, and further combining rule conflict detection, penalty attenuation factor calculation, and nonlinear soft constraint modulation processing, the original causal influence strength can be corrected for business consistency. Thus, the edge weights in the target causal network not only reflect the strength of the data-level correlation but also the degree of constraint of engineering business rules on this correlation, thereby improving the rationality and interpretability of network edge weights in engineering audit scenarios.

[0175] Regarding the implementation details of calculating the audit causal score and causal completeness index of the engineering visa event in step S160, in some examples of the embodiments of this application, firstly, the target causal network is reconstructed using the causal influence strength after adjusting the weights, and a directed graph path search algorithm is called to extract the directed connected subgraph from the reconstructed target causal network that points from the underlying engineering business data source to the engineering visa event node, and the set of causal edges contained in the directed connected subgraph is used as the set of key causal edges.

[0176] In practical implementation, after the aforementioned domain knowledge constraint correction, each causal edge in the target causal network already possesses the adjusted causal influence strength. However, the network may still contain side structures or weakly related paths that have no direct supporting relationship with the current audited visa event. Therefore, the system takes the audited engineering visa event as the target node and performs path search processing on the reconstructed target causal network based on the topological relationship of the directed graph. For example, depth-first search or breadth-first search can be used to trace back from the visa event node along the directed edges to the underlying engineering business data source, thereby extracting all directed connected paths that can propagate to the visa event node, and using the set of causal edges covered by these paths as the key causal edge set.

[0177] For example, in a drainage fee verification scenario, the system can trace back from the verification fee node to drainage equipment start / stop records, on-site water accumulation monitoring records, rainfall monitoring records, and related construction log anomalies, thus forming a local directed connected subgraph around the verification event. Through this process, the originally global and complex causal network is condensed into a local support structure that is only related to the current verification event, which is beneficial for subsequent quantitative assessments of the verification matter itself and reduces interference from branches of the irrelevant graph.

[0178] Then, the adjusted weights of the causal influence strength of each causal edge in the key causal edge set are extracted, and the weighted sum is calculated with the preset engineering business importance weight coefficients assigned to the corresponding causal edge to calculate the audit causal score used to assess the logical support strength of the visa event.

[0179] It should be noted that different key causal edges do not all play the same supporting role in the formation of visa events. Some causal edges may directly correspond to critical changes in the status of the project, such as abnormal foundation conditions, equipment failure, structural displacement changes, and important design adjustments; while others reflect more general auxiliary influences. Therefore, in addition to using the causal influence strength after adjusting the weights of each edge, preset engineering business importance weight coefficients can be configured for different causal edges to reflect their relative supporting position in specific visa matters. Furthermore, the engineering business importance weight coefficients can be determined based on business rules, expert experience, or preset strategies.

[0180] For example, based on the set of key causal edges The audit causality score for assessing the logical support strength of visa incidents is calculated using the following model. : ;

[0181] In the formula, This indicates the node after weight adjustment. To the node The strength of causal influence This refers to the preset engineering business importance weight coefficients assigned to the corresponding causal edges. Through the above calculations, the local supporting relationships scattered in the set of key causal edges can be summarized into a single quantitative result. The higher the score, the stronger the prior factual support for the current visa event in the identified key supporting paths; the lower the score, the weaker the overall key chain supporting the visa event.

[0182] Subsequently, for the multiple engineering business data sources associated with the engineering visa event, they were each taken as the engineering business data sources to be tested. Based on the topological dependency relationship of the reconstructed target causal network, the set of valid parent nodes pointing to the engineering business data sources to be tested was extracted.

[0183] In practice, simply relying on causal scoring to determine whether a visa event has a supporting path is insufficient to reflect the consistency of the entire business chain. Therefore, the system further performs consistency checks on multiple engineering business data sources related to the visa event. Specifically, it iterates through the intermediate state nodes and underlying engineering business data source nodes associated with the current visa event, treating them sequentially as engineering business data sources to be tested. For any engineering business data source to be tested, the system extracts all direct predecessor nodes pointing to that data source based on the topological connections in the reconstructed target causal network, forming a set of valid parent nodes.

[0184] For example, in a certain visa matter, if the data source of the engineering business to be tested is the machine shift cost record, then its effective parent node set may include equipment start-up and shutdown records, construction progress abnormality records, or site status change records; if the data source of the engineering business to be tested is a certain material consumption record, then its effective parent node set may include design change records, construction procedure adjustment records, or component status change records, thereby constructing an upstream input set consistent with its topological location for each node to be tested.

[0185] Subsequently, the actual observed features of the effective parent node set are input into the conditional regression model corresponding to the business data source of the project under test for forward structural equation derivation to obtain the predicted state vector of the business data source of the project under test.

[0186] Here, the system can substitute the actual observed features corresponding to the effective parent node set of each data source under test into its corresponding conditional regression model, perform structural equation forward derivation, and thus calculate the predicted state vector that the data source under test should present under the current parent node state. This predicted state vector can be understood as the model derivation result of the theoretical state of the node under test under causal network constraints.

[0187] For example, in the scenario of a progress anomaly, the system can infer its predicted progress status based on the upstream weather conditions, equipment status, and personnel input status; in the scenario of a cost node, the system can infer its predicted cost status based on the upstream changes in workload, resource input, and process adjustment status. Thus, each data source of the project under test can obtain a corresponding theoretical state reference under the constraints of a unified causal network structure.

[0188] Furthermore, the norm deviation between the actual observed multidimensional vector and the predicted state vector of each engineering business data source is calculated. Then, the normalized ratio between the square of the norm deviation of each engineering business data source and the square of the norm of the actual observed multidimensional vector is calculated and averaged. The non-negative truncation result after subtracting the normalized ratio is used as the causal completeness index characterizing the consistency of global causal prediction.

[0189] For example, suppose the total number of engineering business data sources related to the current engineering visa event and participating in the consistency check is . , No. The actual observed multidimensional vector of a data source of a project under test, located in a dimensionless standardized state space, is denoted as […]. The predicted state vector derived from the conditional regression model is denoted as... Then the causal completeness index It can be calculated using the following formula: ;in, This represents the total number of engineering business data sources participating in the consistency verification. For traversal index, The squared L2 norm of a vector. To prevent the regularization of extremely small positive numbers with a denominator of zero.

[0190] In the above formula, the molecule The character represents the first The magnitude of the deviation between the actual state and the predicted state of the data source of the project under test, denominator This is used to normalize the amplitude differences between different business data sources. After calculating the relative deviation of all data sources participating in the test and taking the average, subtracting the average relative deviation from 1 yields an index characterizing overall consistency. The outermost non-negative truncation ensures that the obtained causal completeness index is not less than zero. If the actual observed state is generally close to the predicted state derived under the constraints of the target causal network, the causal completeness index is high; if the actual observed results at multiple nodes deviate significantly from the predicted results, the index decreases accordingly.

[0191] In this embodiment, by extracting the set of key causal edges from the reconstructed target causal network, calculating the weighted summation of the key causal edges, and further utilizing a conditional regression model to perform structural equation forward inference and consistency checks on the relevant engineering business data sources, an audit causal score characterizing the strength of prior support and a causal completeness index characterizing overall consistency can be obtained. Thus, the supporting path strength and internal network state consistency of the current engineering visa event are quantitatively expressed, thereby providing a calculable quantitative basis for audit judgments of visa matters.

[0192] Regarding the implementation details of identifying whether there is a risk of fraud in the engineering visa event in step S160, in some examples of the embodiments of this application, firstly, a first warning threshold for the audit causality score and a second warning threshold for the causal completeness index are set, and a two-dimensional risk assessment coordinate system is constructed using the first warning threshold and the second warning threshold as the judgment boundary.

[0193] In practical implementation, the system can combine historically approved visa samples, project types, engineering complexity, and fault tolerance requirements under the audit scenario to set corresponding early warning thresholds for the audit causality score and causality completeness index. For example, the first early warning threshold corresponding to the audit causality score can be denoted as... The second early warning threshold corresponding to the causal completeness index is denoted as... Based on this, a two-dimensional risk assessment coordinate system is constructed, using the audit causality score as one coordinate dimension and the causal completeness index as another, with the first and second early warning thresholds serving as dividing lines. This process allows for comparison of different engineering visa events within a unified quantitative space, thus avoiding potential biases that may arise when relying on a single indicator for judgment.

[0194] Then, the engineering visa events are mapped onto a two-dimensional risk assessment coordinate system, and risk type determination is performed based on the mapped location. On the one hand, when the audit causality score of an engineering visa event is lower than the first warning threshold and the causal completeness index is lower than the second warning threshold, the engineering visa event is identified as an engineering visa event with false risk and marked as an inflated anomaly.

[0195] In practice, the system maps the audit causality score and causality completeness index corresponding to a specific auditable engineering visa event onto a two-dimensional risk assessment coordinate system as a set of coordinate values. If the audit causality score of the engineering visa event is lower than the first warning threshold and the causality completeness index is lower than the second warning threshold, it indicates that the visa event is weak in both the dimensions of prior support strength and global consistency. In this case, the engineering visa event can be identified as an engineering visa event with a risk of falsification and marked as an inflated anomaly. This type of situation usually manifests as a lack of strong objective prior support for the visa matter, while the consistency within the relevant business data chain is also low.

[0196] On the other hand, when the audit causality score of an engineering visa event is lower than the first warning threshold and the causal completeness index is greater than or equal to the second warning threshold, the engineering visa event is identified as an engineering visa event with a risk of fraud and marked as a link forgery anomaly.

[0197] Here, if the audit causality score of the engineering visa event is lower than the first warning threshold, and the causal completeness index is greater than or equal to the second warning threshold, it indicates that although the visa event is weak in terms of the strength of its upstream support, its related data chain still maintains a high degree of consistency in form. In this case, the engineering visa event can be identified as an engineering visa event with a risk of fraud and marked as a link forgery anomaly. This type of situation usually manifests as a high degree of superficial consistency between data records related to the visa matter, but lacks a sufficiently strong upstream objective supporting relationship. If the engineering visa event does not meet the above high-risk judgment criteria, it can be judged as a non-target anomaly event or an event requiring further review. Through the above classification process, the system can perform hierarchical identification of visa matters with different risk characteristics, rather than treating all anomalies uniformly.

[0198] For engineering visa events identified as having a risk of fraud, from the key causal edge set Extract the causal influence strength after weight adjustment Target causal edges below a preset strength threshold are identified to determine abnormal causal chain breaks, and a visual audit interactive review report containing these abnormal causal chain breaks is automatically generated.

[0199] In practical implementation, after the system has identified a potential fraudulent risk in a particular engineering visa event at a macro level, it can further analyze and locate its supporting chain at a micro level. For example, the system traverses the set of key causal edges. It compares the adjusted weights of the causal influence of each causal edge with the preset strength threshold. The relationship is as follows. When the causal influence strength of a key causal edge after weight adjustment is lower than a preset strength threshold, the causal edge can be extracted as a target causal edge. Furthermore, the head node and tail node connected to the target causal edge can be identified as abnormal causal chain break nodes, thereby identifying relatively weak or discontinuous business positions in the current visa support chain.

[0200] When generating review results, the system can highlight the abnormal causal chain break nodes and their adjacent key causal edges in the graphical interface and organize them into a visual interactive audit review report. This report not only reflects the location of the engineering visa event in the two-dimensional risk assessment coordinate system, but also shows the local weak links in the set of key causal edges. This allows auditors to directly locate the business nodes that need to be focused on in the supporting chain when reviewing a visa risk, without having to check each node in the complete causal network one by one.

[0201] In the visual audit interactive review report, the fine-grained time alignment index mapping relationship established when performing time alignment processing on multi-source heterogeneous time series data is invoked to construct a multi-dimensional spatiotemporal tracing link for abnormal causal chain break nodes, so as to associate and trace back to locate the corresponding building information model component records and / or original waveforms of field sensors under a unified time axis.

[0202] In practical implementation, since the aforementioned time alignment steps have established a fine-grained time alignment index mapping relationship between a unified time axis and the underlying original sampling points, when an abnormal causal chain break node is identified, the system can further map the unified time position corresponding to the node back to the specific underlying business record based on this mapping relationship. For example, if an abnormal causal chain break node corresponds to a component status change record, the system can retrieve the building information model component change record or component snapshot at the corresponding time position; if an abnormal causal chain break node corresponds to a physical monitoring state, the system can retrieve the sensor's original waveform, monitoring curve, or related sampling record at the corresponding time position, forming a multi-dimensional spatiotemporal tracing link associated with the abnormal causal chain break node.

[0203] Through the above processing, when auditors view the visual audit interactive review report, they can not only see the classification results of engineering visa events in the risk coordinate system, but also further drill down to view the underlying business evidence corresponding to the abnormal causal chain break nodes. This allows them to correlate the visa risk identification results with specific business records, which helps improve the verifiability of risk assessment results and provides clearer location clues for manual review. Thus, the engineering visa risk identification results have both quantitative basis for judgment and can be linked to specific review objects and review clues.

[0204] Figure 4The diagram illustrates an example of a system operation mechanism principle of an engineering false visa auditing method based on multi-source temporal causal inference according to an embodiment of this application. The system can be divided into four main functional modules in its overall architecture: input data area, core processing area, scheduling and constraint area, and output and result area.

[0205] like Figure 4 As shown, firstly, the system comprehensively collects multi-source, heterogeneous underlying business facts from the engineering construction process through the input data area, including Building Information Modeling (BIM) logs, on-site sensor monitoring records, construction progress logs, and cost expenditure records, and simultaneously acquires auditable approval events as the core audit target. These multi-dimensional business data with varying sampling frequencies and complex structures collectively constitute the raw input materials for the system to conduct subsequent in-depth causal analysis.

[0206] Subsequently, the input data flows to the core processing area. Specifically, the heterogeneous data is first time-aligned using the Multi-Resolution Dynamic Time Normalization (MR-DTW) algorithm to generate a time series matrix corresponding to a unified time axis. Then, based on this time series matrix, the lag correlations between different engineering business data sources are calculated, and a time series graph embedding representation is extracted. Finally, the Multi-Source Time-Invariant Causal Discovery (MT-ICD) algorithm is used to calculate the causal influence strength between various business elements. During this process, the scheduling and constraint area uses prior domain knowledge such as engineering specifications and contract terms as engineering domain knowledge constraints. A nonlinear soft constraint mechanism is used to adjust the weights of corresponding causal edges in the target causal network, ensuring that the optimized target causal network balances data relationships and engineering business rules.

[0207] Finally, based on the knowledge-constrained optimized target causal network, the core processing area performs anomaly detection and evaluation, namely, calculating the audit causal score and causal completeness index. The system determines the risk type of the engineering visa event based on the mapping position of the evaluation indicators in the two-dimensional risk assessment coordinate system and outputs the evaluation results to the output and results areas. For engineering visa events identified as having false risks, the system generates an audit report containing abnormal causal chain breaks and underlying source information, thus forming a processing flow from multi-source engineering data processing to the output of false risk identification results.

[0208] To comprehensively verify the performance boundaries of the proposed method for auditing false engineering visas based on multi-source temporal causal inference (i.e., a technical solution integrating multi-resolution dynamic time normalization and multi-source temporal invariant causal discovery) in a controlled environment, and to objectively evaluate its effectiveness in real-world complex engineering scenarios, this study constructs a multi-source heterogeneous temporal simulation platform that follows the laws of engineering business for experimental verification. The experiment abandons traditional simple linear temporal simulation and instead strictly constructs a causal directed acyclic graph containing five key variable nodes according to the physical and management logic of engineering construction. The nodes in this graph specifically include: external environmental variables (such as rainfall and temperature) that affect construction progress as sources of interference; physical monitoring variables reflecting the true physical state of the structure or materials (corresponding to high-frequency sensor records on site); process status variables corresponding to the construction progress log; resource input variables corresponding to material procurement and labor cost expenditures (usually exhibiting discrete and non-uniform sampling characteristics); and final settlement and claim observation indicators driven by process status and resource input, i.e., visa event variables.

[0209] Based on the aforementioned normal causal evolution data flow, the experiment simulated three typical engineering fraud scenarios through anomaly injection to test the algorithm's risk identification capabilities in fine-grained features and causal structures. Specific scenarios included: the first was an artificial inflation anomaly simulation, where the amplitudes of visa event variables and resource input variables were artificially increased, but the transmission support between the front-end physical and state nodes was severed; the second was a link forgery anomaly simulation, where a superficial pseudo-correlation was established between process state variables and visa event variables by tampering with them, but the underlying physical monitoring variables did not show corresponding changes in physical state, forming a structural break with an effect but no cause; and the third was a time-series misalignment claim simulation, where a real engineering event occurred, but it was repeatedly or misaligned at an unreasonable time step.

[0210] To objectively evaluate the advantages of the method of this invention, two typical comparative baselines were set up in the experiment. The first was the traditional Vector Autoregressive Granger Causal Analysis (VAR-Granger), representing the conventional auditing methods that rely solely on linear temporal correlations and lack heterogeneous time alignment capabilities. The second was spatiotemporal deep learning anomaly detection models (such as ST-DeepND), representing "black box" prediction methods with strong nonlinear fitting capabilities but being purely data-driven and difficult to provide explicit causal explanation paths. Through comparative analysis with the above baseline models, the technical advantages of this invention in multi-source data alignment, resistance to environmental noise, and assessment of causal link integrity were further verified.

[0211] Figure 5This diagram illustrates a comparative simulation of different methods in identifying various engineering visa fraud scenarios. The simulation uses the F1-Score as the core quantitative evaluation indicator, with the horizontal axis covering three typical engineering visa fraud scenarios: isolated cost inflation, logical link forgery, and time-dislocation claims. The baseline methods used in the comparison include the traditional Vector Autoregressive Granger Causal Analysis (VAR-Granger, representing a conventional data verification method relying solely on linear time-series correlations) and the Spatiotemporal Deep Learning Anomaly Detection Model (ST-DeepND, representing a high-order, purely data-driven prediction method with strong nonlinear fitting capabilities but lacking explicit causal explanation).

[0212] Combination Figure 5 The simulation results, as shown in the histogram, reveal that in the relatively simple "isolated cost inflation" scenario (i.e., artificially inflating the final visa amount and resource input, but lacking front-end node support), both the ST-DeepND model and the method of this invention exhibit high recognition rates, with F1-Scores of 0.92 and 0.94 respectively, outperforming VAR-Granger's 0.80. However, in the more concealed "logical link forgery" scenario, VAR-Granger and ST-DeepND are easily deceived by this statistically highly correlated surface forged data by tampering with surface progress or management logs to conform to the surface statistical distribution, causing their F1-Scores to plummet to 0.65 and 0.66 respectively. In contrast, the method of this invention, benefiting from the multi-source time-invariant causal discovery mechanism (MT-ICD) and the cross-constraints of engineering domain knowledge, can identify the "causal chain break" phenomenon (i.e., structural break with an effect but no cause) between the high-frequency physical monitoring layer and the low-frequency management ledger layer, maintaining a high recognition rate even in this complex scenario.

[0213] Furthermore, in the scenario of "time misalignment claims" (i.e., actual engineering changes occurred, but were submitted at unreasonable time steps or in duplicate), the method of this invention also significantly outperformed VAR-Granger (0.70) and ST-DeepND (0.75) with an F1-Score of 0.88. This fully demonstrates that the multi-resolution dynamic time normalization algorithm (MR-DTW) adopted in this invention improves the accuracy of time alignment and the ability to reconstruct engineering processes when processing multi-source heterogeneous and non-uniformly sampled data. Thus, by deeply exploring the invariant causal mechanisms in the engineering environment, the method of this invention reduces the underlying limitations of traditional pure data-driven models that "emphasize correlation and neglect causality."

[0214] Figure 6This diagram illustrates a comparative simulation of the robustness of different methods in identifying fraudulent visas under varying environmental noise levels. The simulation uses AUROC (Area Under the Receiver Operating Characteristic) as a quantitative indicator to evaluate detection stability. The horizontal axis represents the gradually increasing environmental noise level (from 0 to 0.6), and the vertical axis represents the AUROC value for each model. The three broken lines in the diagram represent the performance degradation trends of the proposed method, the traditional VAR-Granger method, and conventional baseline algorithms (such as purely data-driven deep learning baseline methods) as the underlying data noise increases. In actual engineering field operations, physical sensors and ledger records often generate significant noise due to complex environmental interference. This experiment aims to verify the system's detection stability under harsh data environments.

[0215] Combination Figure 6 The trend lines show that all algorithms possess high recognition capabilities in zero or extremely low noise environments; however, as noise intensity increases, the performance of the baseline methods degrades, and the false alarm rate rises. Specifically, when the ambient noise level reaches 0.5, the AUROC value of the VAR-Granger method drops below 0.65, and the degradation of the conventional baseline algorithm is even more severe. Both methods show a significant decrease in evaluation metrics under strong noise conditions. In contrast, the AUROC curve of the method in this invention decays slowly. Even under extreme test conditions with an ambient noise level as high as 0.6, its AUROC value remains stable above 0.82, demonstrating high noise robustness.

[0216] The method of this invention maintains good performance under high noise conditions, mainly due to its processing mechanism: on the one hand, the multi-resolution dynamic time normalization algorithm (MR-DTW) compensates for time misalignment and flexibly aligns asynchronous, non-uniformly sampled multi-source heterogeneous data, helping to reduce interference caused by time mismatch; on the other hand, the multi-source time-invariant causality discovery mechanism (MT-ICD) helps filter out spurious statistical correlations caused by environmental noise fluctuations or system covariance. Simulation results show that, under the set noise conditions, the method of this invention can maintain good noise resistance stability.

[0217] Figure 7 This diagram illustrates a 3D response surface representation of an example of hyperparameter sensitivity analysis in an embodiment of this application. The simulation experiment aims to investigate the system's dependence on and robustness to core hyperparameters in practical engineering deployments. In the diagram, the X-axis is defined as the maximum time lag window length (denoted as...). The Y-axis represents the number of environment partitions (denoted as ). The Z-axis represents the Causal Completeness Index (CII) of the final calculated output of the system. The synergistic influence of the two key hyperparameters on the final visa audit evaluation indicators is visually demonstrated through a 3D color response surface.

[0218] Combination Figure 7 The 3D response surface morphology clearly shows that the Causal Completeness Index (CII) exhibits an extremely smooth and stable "plateau region" (i.e., the highlighted area at the top of the figure) over a large range of parameter combinations. Within this region, the mean CII score remains consistently above 0.85. This experimental result demonstrates that the proposed algorithm model is less prone to overfitting, is insensitive to hyperparameter changes, and possesses low parameter tuning difficulty and high generalization ability in practical engineering applications. Furthermore, this simulation surface provides a clear optimal guidance range for actual business deployment: it is recommended to set the maximum time lag window length to [value missing]. Within the scope, and set the number of environmental divisions. Within this range, the system can demonstrate good causal inference performance.

[0219] also, Figure 7 The edge region reflects the decreasing trend of the causal completeness index when the parameter settings deviate from the above range. On the one hand, when the time alignment window is too small, the temporal search range of the system is limited, which may make it difficult to fully capture the long time delays and hysteresis feedback relationships in engineering events; on the other hand, when the number of environment partitions is too small, the data base available for cross-environment comparison is reduced, which may affect the filtering effect on accidental co-occurrence relationships. The above simulation results can provide a reference for the parameter settings of the maximum temporal hysteresis window length and the number of environment partitions.

[0220] In this embodiment, a two-dimensional quantitative evaluation system based on Audit Causality Score (ACS) and Causality Completeness Index (CII) is constructed by combining "data-driven discovery" with "domain knowledge soft constraints". Multi-source simulation results based on engineering logic show that, under the set simulation conditions, this method maintains high recognition accuracy in logic link forgery scenarios and maintains a certain degree of stability under noise and parameter variations.

[0221] Figure 8 A structural block diagram of an example of an engineering fraud audit system based on multi-source temporal causal inference according to an embodiment of this application is shown.

[0222] like Figure 8 As shown, the engineering false visa audit system 800 based on multi-source time-series causal inference includes a multi-source data acquisition unit 810, a multi-resolution alignment unit 820, an interactive graphing and characterization unit 830, an invariant causal discovery unit 840, a causal network optimization unit 850, and a risk scoring and identification unit 860.

[0223] The multi-source data acquisition unit 810 is used to acquire multi-source heterogeneous time-series data of engineering projects and engineering visa events to be audited. The multi-source heterogeneous time-series data comes from multiple engineering business data sources, and the multiple engineering business data sources have different sampling frequencies.

[0224] The multi-resolution alignment unit 820 is used to perform time alignment processing on the multi-source heterogeneous time series data using a multi-resolution dynamic time normalization algorithm to generate a time series matrix corresponding to a unified time axis.

[0225] The interactive graph representation unit 830 is used to calculate the lag correlation between different engineering business data sources based on the time series matrix to construct a directed time series interactive graph, and to extract the time series graph embedding representation of each engineering business data source in the directed time series interactive graph.

[0226] The invariant causal discovery unit 840 is used to perform multi-source time-series invariant causal discovery in multiple pre-divided environments based on the time-series graph embedding representation, in order to determine the residual change characteristics and parameter consistency characteristics of the regression model in different environments, and to calculate the causal influence strength between different engineering business data sources based on the residual change characteristics and parameter consistency characteristics; each environment is used to indicate the corresponding construction status and / or data source attributes.

[0227] The causal network optimization unit 850 is used to construct a target causal network corresponding to the engineering visa event based on the causal influence intensity, and optimize the target causal network in combination with preset engineering domain knowledge constraints to adjust the weight of the corresponding causal edges in the target causal network; the engineering domain knowledge constraints include at least one of the following: engineering design specifications, contract terms constraints, and historical experience models.

[0228] The risk scoring and identification unit 860 is used to calculate the audit causality score and causal completeness index of the engineering visa event based on the target causal network after weight adjustment, and to identify whether there is a false risk in the engineering visa event based on the audit causality score and the causal completeness index.

[0229] In some embodiments, this application provides a non-volatile computer-readable storage medium storing one or more programs including execution instructions. These execution instructions can be read and executed by electronic devices (including but not limited to computers, servers, or network devices) to perform the steps of any of the above-described engineering false visa auditing methods based on multi-source temporal causal inference.

[0230] In some embodiments, this application also provides a computer program product, the computer program product including a computer program stored on a non-volatile computer-readable storage medium, the computer program including program instructions, which, when executed by a computer, cause the computer to perform the steps of any of the above-described engineering false visa audit methods based on multi-source temporal causal inference.

[0231] In some embodiments, this application also provides an electronic device comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of an engineering fraud auditing method based on multi-source temporal causal inference.

[0232] The above-described product can perform the methods provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects for performing the methods. Technical details not described in detail in this embodiment can be found in the methods provided in the embodiments of this application.

[0233] The electronic devices in this application can exist in various forms, including but not limited to: mobile communication devices, ultra-mobile personal computer devices, portable entertainment devices, or other airborne electronic devices with data interaction functions.

[0234] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0235] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device to execute the methods described in the various embodiments or some parts of the embodiments.

[0236] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for auditing fraudulent engineering visas based on multi-source temporal causal inference, characterized in that, The method includes: Acquire multi-source heterogeneous time-series data of engineering projects and engineering visa events to be audited. The multi-source heterogeneous time-series data comes from multiple engineering business data sources, and the multiple engineering business data sources have different sampling frequencies. A multi-resolution dynamic time normalization algorithm is used to perform time alignment processing on the multi-source heterogeneous time series data to generate a time series matrix corresponding to a unified time axis. Based on the time series matrix, the lag correlation between different engineering business data sources is calculated to construct a directed time series interaction graph, and the time series graph embedding representation of each engineering business data source in the directed time series interaction graph is extracted; Based on the time series graph embedding representation, multi-source time-invariant causal discovery is performed in multiple pre-divided environments to determine the residual variation characteristics and parameter consistency characteristics of the regression model under different environments, and to calculate the causal influence strength between different engineering business data sources based on the residual variation characteristics and parameter consistency characteristics; each environment is used to indicate the corresponding construction status and / or data source attributes; Based on the strength of the causal influence, a target causal network corresponding to the engineering visa event is constructed, and the target causal network is optimized in combination with preset engineering domain knowledge constraints to adjust the weights of the corresponding causal edges in the target causal network; the engineering domain knowledge constraints include at least one of the following: engineering design specifications, contract terms constraints, and historical experience models; The audit causality score and causal completeness index of the engineering visa event are calculated based on the weighted target causal network, and the presence of false risk of the engineering visa event is identified based on the audit causality score and the causal completeness index.

2. The method for auditing false engineering visas based on multi-source temporal causal inference according to claim 1, characterized in that, The step of employing a multi-resolution dynamic time normalization algorithm to perform time alignment processing on the multi-source heterogeneous time series data to generate a time series matrix corresponding to a unified time axis includes: Obtain the time series and corresponding timestamp sequences of different engineering business data sources in the multi-source heterogeneous time series data; A weighted distance measure function with a time difference penalty term is constructed. The weighted distance measure function is used to measure the difference between sequential sample points by multiplying the norm distance of the sample feature difference vector with the exponential weight function based on time penalty. The exponential weight function uses a preset penalty hyperparameter to control the sensitivity of path alignment to time difference. Downsampling is performed on the time series of the different engineering business data sources to obtain the corresponding coarse-grained time series, and the global cumulative distance between the coarse-grained time series is calculated based on the weighted distance metric function. The coarse-grained alignment path that minimizes the global cumulative distance is then extracted. The coarse-grained alignment path is mapped back to the original sampling resolution, and a local search constraint boundary with a preset time radius is set with the mapped path as the center. Within the local search constraint boundary, the weighted distance metric function is invoked to perform fine-resolution dynamic time-normalized matching iteration to obtain the corresponding fine-grained time-aligned index mapping relationship; Using the timestamp sequence of the preset benchmark engineering business data source as a unified time axis reference, and based on the fine-grained time alignment index mapping relationship, sampling point alignment processing is performed on each of the other engineering business data sources to generate a time series matrix corresponding to the unified time axis.

3. The method for auditing false engineering visas based on multi-source temporal causal inference according to claim 2, characterized in that, The step of calculating the lag correlation between different engineering business data sources based on the time series matrix to construct a directed time series interaction graph, and extracting the time series graph embedding representation of each engineering business data source in the directed time series interaction graph, includes: Zero-mean standardization is performed on the time series of each engineering business data source in the time series matrix to eliminate dimensional differences and baseline drift between different engineering business data sources; Set the maximum time-series lag window length and the lag decay coefficient that decreases over time; For any two engineering business data sources after standardization, the cumulative value of the cross sequence feature product is calculated within the lag window and the absolute value is taken, based on the historical observation value of the first engineering business data source and the current observation value of the second engineering business data source. Combined with the lag attenuation coefficient, the lag influence score in the corresponding direction is calculated. A preset significant noise truncation threshold is set, and asymmetric sparse filtering based on bidirectional comparison is performed on the lag effect score: when the positive lag effect score is strictly greater than the negative lag effect score and greater than the significant noise truncation threshold, the positive lag effect score is confirmed as an effective lag association and assigned as a directed edge weight; otherwise, the corresponding directed edge weight is reset to zero to construct an asymmetric adjacency matrix of the directed temporal interaction graph with sparse noise reduction characteristics. The symmetric fusion matrix of the asymmetric adjacency matrix is ​​extracted, and the degree distribution of each network node is calculated based on the symmetric fusion matrix to generate a degree matrix. Then, the degree matrix and the symmetric fusion matrix are combined to construct an undirected Laplacian matrix for graph analysis. The undirected Laplacian matrix is ​​subjected to eigenvalue decomposition to extract eigenvectors corresponding to the first few smallest non-zero eigenvalues, which are then concatenated to form a dimension-reduced feature matrix. The row vectors in the dimension-reduced feature matrix corresponding to each engineering business data source are used as the time-series graph embedding representation of each engineering business data source.

4. The method for auditing false engineering visas based on multi-source temporal causal inference according to claim 3, characterized in that, The step of performing multi-source time-invariant causal discovery based on the time-series graph embedding representation in pre-divided multiple environments to determine the residual variation characteristics and parameter consistency characteristics of the regression model under different environments, and calculating the causal influence strength between different engineering business data sources based on the residual variation characteristics and parameter consistency characteristics, includes: Based on the time-series graph embedding representation of each of the engineering business data sources, the embedding vector similarity between the target engineering business data source and the other engineering business data sources is calculated, and the engineering business data sources with embedding vector similarity greater than a preset similarity threshold are extracted as a set of candidate parent variables for constructing a regression model for the target engineering business data source, so as to achieve dimensionality reduction of the causal inference search space. For each pre-divided environment in the set of multiple pre-divided environments, for any test engineering business data source in the set of candidate parent variables, a baseline regression model that does not contain the test engineering business data source and a conditional regression model that contains the test engineering business data source are constructed respectively; wherein, the conditional regression model in each environment adopts a unified parameterized model structure. By fitting the baseline regression model and the conditional regression model under all environments, the baseline fit residuals and conditional fit residuals under all environments are aggregated to calculate the variance of the baseline residuals and the variance of the conditional residuals across environments, and the residual change characteristics are determined based on the relative reduction ratio of the baseline residual variance and the conditional residual variance. Extract the model parameter vectors of the conditional regression model under various environments, and determine the parameter consistency characteristics by calculating the vector space distance between the model parameter vectors under other environments and the model parameter vectors under a preset reference environment. The residual variation characteristics are multiplied by the average satisfaction of the parameter consistency characteristics across environments after removing the reference environment, in order to calculate and quantify the causal influence strength from the test engineering business data source to the target engineering business data source.

5. The method for auditing false engineering visas based on multi-source temporal causal inference according to claim 4, characterized in that, The optimization of the target causal network by incorporating preset engineering domain knowledge constraints to adjust the weights of corresponding causal edges in the target causal network includes: At least one of the engineering design specifications, contract terms and constraints and historical experience model is parsed and abstracted into entity relation triples to construct an engineering domain knowledge graph, and the engineering domain knowledge graph is transformed into a prior knowledge matrix with constraint strength labels. For any target causal edge in the target causal network, the data source business attributes of its head node and tail node are mapped to the corresponding coordinate positions of the prior knowledge matrix in order to extract the domain logic matching degree coefficient corresponding to the target causal edge. Based on the engineering domain knowledge constraints, the target causal edge is subjected to rule conflict detection: when the target causal edge indicates the occurrence of additional cost causal derivation, and the engineering sub-items corresponding to the head node and tail node are identified as non-additional cost items in the contract terms constraints, or when the evolution trend of the target causal edge violates the topological constraint relationship of the engineering design specification, a domain logic conflict is determined to be triggered. When the domain logic conflict is triggered, a penalty attenuation factor greater than zero is determined based on the domain logic matching degree coefficient, and the penalty attenuation factor decreases as the domain logic matching degree coefficient increases; when the domain logic conflict is not triggered, the penalty attenuation factor is set to zero. Based on the original causal influence strength of the target causal edge and the determined penalty attenuation factor, a nonlinear soft constraint modulation function based on prior knowledge is invoked to transform the penalty attenuation factor into a nonlinear attenuation effect on the original causal influence strength, so as to calculate the causal influence strength after weight adjustment and use it to reconstruct the target causal network.

6. The method for auditing false engineering visas based on multi-source temporal causal inference according to claim 5, characterized in that, The calculation of the audit causality score and causal completeness index of the engineering visa event based on the weighted target causal network includes: The target causal network is reconstructed using the adjusted weights of the causal influence strength, and a directed graph path search algorithm is called to extract a directed connected subgraph from the reconstructed target causal network that points from the underlying engineering business data source to the engineering visa event node, and the set of causal edges contained in the directed connected subgraph is taken as the key causal edge set. Extract the adjusted weighted causal influence strength of each causal edge in the key causal edge set, and sum it with the preset engineering business importance weight coefficient assigned to the corresponding causal edge to calculate the audit causal score used to assess the logical support strength of the visa event. For the multiple engineering business data sources associated with the engineering visa event, each is taken as the engineering business data source to be tested. Based on the topological dependency relationship of the reconstructed target causal network, the set of effective parent nodes pointing to the engineering business data source to be tested is extracted. The actual observed features of the effective parent node set are input into the conditional regression model corresponding to the business data source of the project under test for forward structural equation derivation to obtain the predicted state vector of the business data source of the project under test. The norm deviation between the actual observed multidimensional vector and the predicted state vector of each engineering business data source is calculated. Then, the normalized ratio between the square of the norm deviation of each engineering business data source and the square of the norm of the actual observed multidimensional vector is calculated and averaged. The non-negative truncation result after subtracting the normalized ratio is used as the causal completeness index characterizing the consistency of global causal prediction.

7. The method for auditing false engineering visas based on multi-source temporal causal inference according to claim 6, characterized in that, The process of identifying the risk of fraudulent activity in the engineering visa event based on the audit causality score and the causal completeness index includes: A first warning threshold is set for the audit causality score, and a second warning threshold is set for the causal completeness index. A two-dimensional risk assessment coordinate system is constructed using the first warning threshold and the second warning threshold as the judgment boundary. The engineering visa event is mapped onto the two-dimensional risk assessment coordinate system, and the risk type is determined based on the mapped location; When the audit causality score of the engineering visa event is lower than the first warning threshold and the causal completeness index is lower than the second warning threshold, the engineering visa event is identified as an engineering visa event with a risk of fraud and marked as an inflated anomaly. When the audit causality score of the engineering visa event is lower than the first warning threshold and the causal integrity index is greater than or equal to the second warning threshold, the engineering visa event is identified as an engineering visa event with a risk of fraud and marked as a link forgery anomaly. When the engineering visa event does not meet the above conditions, it is judged as a non-target abnormal event or an event requiring further review. For the identified engineering visa events with false risks, target causal edges with adjusted weights and causal influence strength below a preset strength threshold are extracted from the key causal edge set to identify abnormal causal chain nodes, and a visual audit interactive review report containing the abnormal causal chain nodes is automatically generated. Specifically, in the visual audit interactive review report, the fine-grained time alignment index mapping relationship established when performing time alignment processing on the multi-source heterogeneous time series data is invoked to construct a multi-dimensional spatiotemporal tracing link for the abnormal causal chain break node, so as to associate and trace back to locate the corresponding building information model component record and / or the original waveform of the field sensor under the unified time axis.

8. A system for auditing fraudulent engineering visas based on multi-source temporal causal inference, characterized in that, The system includes: A multi-source data acquisition unit is used to acquire multi-source heterogeneous time-series data of engineering projects and engineering visa events to be audited. The multi-source heterogeneous time-series data comes from multiple engineering business data sources, and the multiple engineering business data sources have different sampling frequencies. A multi-resolution alignment unit is used to perform time alignment processing on the multi-source heterogeneous time series data using a multi-resolution dynamic time normalization algorithm to generate a time series matrix corresponding to a unified time axis. An interactive graph representation unit is used to calculate the lag correlation between different engineering business data sources based on the time series matrix to construct a directed time series interactive graph, and to extract the time series graph embedding representation of each engineering business data source in the directed time series interactive graph; The invariant causal discovery unit is used to perform multi-source time-series invariant causal discovery in multiple pre-divided environments based on the time-series graph embedding representation, in order to determine the residual change characteristics and parameter consistency characteristics of the regression model in different environments, and to calculate the causal influence strength between different engineering business data sources based on the residual change characteristics and parameter consistency characteristics; each environment is used to indicate the corresponding construction status and / or data source attributes. The causal network optimization unit is used to construct a target causal network corresponding to the engineering visa event based on the causal influence intensity, and optimize the target causal network in combination with preset engineering domain knowledge constraints to adjust the weights of the corresponding causal edges in the target causal network; the engineering domain knowledge constraints include at least one of the following: engineering design specifications, contract terms constraints, and historical experience models; The risk scoring and identification unit is used to calculate the audit causality score and causal completeness index of the engineering visa event based on the target causal network after weight adjustment, and to identify whether there is a false risk in the engineering visa event based on the audit causality score and the causal completeness index.

9. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the method as described in any one of claims 1-7.

10. A storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-7.

Images